repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
TiKick
TiKick-main/setup.py
#!/usr/bin/env python # -*- coding: utf-8 -*- # Copyright 2021 The TARTRL Authors. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """""" import os from setuptools import setup, find_packages import setuptools def get_version() -> str: # https://packaging.python.org/guides/single-sourcing-package-version/ init = open(os.path.join("tmarl", "__init__.py"), "r").read().split() return init[init.index("__version__") + 2][1:-1] setup( name="tmarl", # Replace with your own username version=get_version(), description="marl algorithms", long_description=open("README.md", encoding="utf8").read(), long_description_content_type="text/markdown", author="tmarl", author_email="tmarl_contact@tartrl.cn", packages=setuptools.find_packages(), classifiers=[ "Development Status :: 3 - Alpha", "Intended Audience :: Science/Research", "Topic :: Scientific/Engineering :: Artificial Intelligence", "Topic :: Software Development :: Libraries :: Python Modules", "Programming Language :: Python :: 3", "License :: OSI Approved :: Apache License", "Operating System :: OS Independent", ], keywords="multi-agent reinforcement learning algorithms pytorch", python_requires='>=3.6', )
1,788
35.510204
74
py
TiKick
TiKick-main/tmarl/networks/policy_network.py
import torch import torch.nn as nn from tmarl.networks.utils.util import init, check from tmarl.networks.utils.mlp import MLPBase, MLPLayer from tmarl.networks.utils.rnn import RNNLayer from tmarl.networks.utils.act import ACTLayer from tmarl.networks.utils.popart import PopArt from tmarl.utils.util import get_shape_from_obs_space # networks are defined here class PolicyNetwork(nn.Module): def __init__(self, args, obs_space, action_space, device=torch.device("cpu")): super(PolicyNetwork, self).__init__() self.hidden_size = args.hidden_size self._gain = args.gain self._use_orthogonal = args.use_orthogonal self._activation_id = args.activation_id self._use_policy_active_masks = args.use_policy_active_masks self._use_naive_recurrent_policy = args.use_naive_recurrent_policy self._use_recurrent_policy = args.use_recurrent_policy self._use_influence_policy = args.use_influence_policy self._influence_layer_N = args.influence_layer_N self._use_policy_vhead = args.use_policy_vhead self._recurrent_N = args.recurrent_N self.tpdv = dict(dtype=torch.float32, device=device) obs_shape = get_shape_from_obs_space(obs_space) self._mixed_obs = False self.base = MLPBase(args, obs_shape, use_attn_internal=False, use_cat_self=True) input_size = self.base.output_size if self._use_naive_recurrent_policy or self._use_recurrent_policy: self.rnn = RNNLayer(input_size, self.hidden_size, self._recurrent_N, self._use_orthogonal) input_size = self.hidden_size if self._use_influence_policy: self.mlp = MLPLayer(obs_shape[0], self.hidden_size, self._influence_layer_N, self._use_orthogonal, self._activation_id) input_size += self.hidden_size self.act = ACTLayer(action_space, input_size, self._use_orthogonal, self._gain) if self._use_policy_vhead: init_method = [nn.init.xavier_uniform_, nn.init.orthogonal_][self._use_orthogonal] def init_(m): return init(m, init_method, lambda x: nn.init.constant_(x, 0)) if self._use_popart: self.v_out = init_(PopArt(input_size, 1, device=device)) else: self.v_out = init_(nn.Linear(input_size, 1)) self.to(device) def forward(self, obs, rnn_states, masks, available_actions=None, deterministic=False): if self._mixed_obs: for key in obs.keys(): obs[key] = check(obs[key]).to(**self.tpdv) else: obs = check(obs).to(**self.tpdv) rnn_states = check(rnn_states).to(**self.tpdv) masks = check(masks).to(**self.tpdv) if available_actions is not None: available_actions = check(available_actions).to(**self.tpdv) actor_features = self.base(obs) if self._use_naive_recurrent_policy or self._use_recurrent_policy: actor_features, rnn_states = self.rnn(actor_features, rnn_states, masks) if self._use_influence_policy: mlp_obs = self.mlp(obs) actor_features = torch.cat([actor_features, mlp_obs], dim=1) actions, action_log_probs = self.act(actor_features, available_actions, deterministic) return actions, action_log_probs, rnn_states def evaluate_actions(self, obs, rnn_states, action, masks, available_actions=None, active_masks=None): if self._mixed_obs: for key in obs.keys(): obs[key] = check(obs[key]).to(**self.tpdv) else: obs = check(obs).to(**self.tpdv) rnn_states = check(rnn_states).to(**self.tpdv) action = check(action).to(**self.tpdv) masks = check(masks).to(**self.tpdv) if available_actions is not None: available_actions = check(available_actions).to(**self.tpdv) if active_masks is not None: active_masks = check(active_masks).to(**self.tpdv) actor_features = self.base(obs) if self._use_naive_recurrent_policy or self._use_recurrent_policy: actor_features, rnn_states = self.rnn(actor_features, rnn_states, masks) if self._use_influence_policy: mlp_obs = self.mlp(obs) actor_features = torch.cat([actor_features, mlp_obs], dim=1) action_log_probs, dist_entropy = self.act.evaluate_actions(actor_features, action, available_actions, active_masks = active_masks if self._use_policy_active_masks else None) values = self.v_out(actor_features) if self._use_policy_vhead else None return action_log_probs, dist_entropy, values def get_policy_values(self, obs, rnn_states, masks): if self._mixed_obs: for key in obs.keys(): obs[key] = check(obs[key]).to(**self.tpdv) else: obs = check(obs).to(**self.tpdv) rnn_states = check(rnn_states).to(**self.tpdv) masks = check(masks).to(**self.tpdv) actor_features = self.base(obs) if self._use_naive_recurrent_policy or self._use_recurrent_policy: actor_features, rnn_states = self.rnn(actor_features, rnn_states, masks) if self._use_influence_policy: mlp_obs = self.mlp(obs) actor_features = torch.cat([actor_features, mlp_obs], dim=1) values = self.v_out(actor_features) return values
5,558
41.113636
181
py
TiKick
TiKick-main/tmarl/networks/utils/distributions.py
import torch import torch.nn as nn from .util import init """ Modify standard PyTorch distributions so they are compatible with this code. """ # # Standardize distribution interfaces # # Categorical class FixedCategorical(torch.distributions.Categorical): def sample(self): return super().sample().unsqueeze(-1) def log_probs(self, actions): return ( super() .log_prob(actions.squeeze(-1)) .view(actions.size(0), -1) .sum(-1) .unsqueeze(-1) ) def mode(self): return self.probs.argmax(dim=-1, keepdim=True) # Normal class FixedNormal(torch.distributions.Normal): def log_probs(self, actions): return super().log_prob(actions).sum(-1, keepdim=True) def entrop(self): return super.entropy().sum(-1) def mode(self): return self.mean # Bernoulli class FixedBernoulli(torch.distributions.Bernoulli): def log_probs(self, actions): return super.log_prob(actions).view(actions.size(0), -1).sum(-1).unsqueeze(-1) def entropy(self): return super().entropy().sum(-1) def mode(self): return torch.gt(self.probs, 0.5).float() class Categorical(nn.Module): def __init__(self, num_inputs, num_outputs, use_orthogonal=True, gain=0.01): super(Categorical, self).__init__() init_method = [nn.init.xavier_uniform_, nn.init.orthogonal_][use_orthogonal] def init_(m): return init(m, init_method, lambda x: nn.init.constant_(x, 0), gain) self.linear = init_(nn.Linear(num_inputs, num_outputs)) def forward(self, x, available_actions=None): x = self.linear(x) if available_actions is not None: x[available_actions == 0] = -1e10 return FixedCategorical(logits=x) class DiagGaussian(nn.Module): def __init__(self, num_inputs, num_outputs, use_orthogonal=True, gain=0.01): super(DiagGaussian, self).__init__() init_method = [nn.init.xavier_uniform_, nn.init.orthogonal_][use_orthogonal] def init_(m): return init(m, init_method, lambda x: nn.init.constant_(x, 0), gain) self.fc_mean = init_(nn.Linear(num_inputs, num_outputs)) self.logstd = AddBias(torch.zeros(num_outputs)) def forward(self, x): action_mean = self.fc_mean(x) # An ugly hack for my KFAC implementation. zeros = torch.zeros(action_mean.size()) if x.is_cuda: zeros = zeros.cuda() action_logstd = self.logstd(zeros) return FixedNormal(action_mean, action_logstd.exp()) class Bernoulli(nn.Module): def __init__(self, num_inputs, num_outputs, use_orthogonal=True, gain=0.01): super(Bernoulli, self).__init__() init_method = [nn.init.xavier_uniform_, nn.init.orthogonal_][use_orthogonal] def init_(m): return init(m, init_method, lambda x: nn.init.constant_(x, 0), gain) self.linear = init_(nn.Linear(num_inputs, num_outputs)) def forward(self, x): x = self.linear(x) return FixedBernoulli(logits=x) class AddBias(nn.Module): def __init__(self, bias): super(AddBias, self).__init__() self._bias = nn.Parameter(bias.unsqueeze(1)) def forward(self, x): if x.dim() == 2: bias = self._bias.t().view(1, -1) else: bias = self._bias.t().view(1, -1, 1, 1) return x + bias
3,466
27.891667
86
py
TiKick
TiKick-main/tmarl/networks/utils/mlp.py
import torch.nn as nn from .util import init, get_clones class MLPLayer(nn.Module): def __init__(self, input_dim, hidden_size, layer_N, use_orthogonal, activation_id): super(MLPLayer, self).__init__() self._layer_N = layer_N active_func = [nn.Tanh(), nn.ReLU(), nn.LeakyReLU(), nn.ELU()][activation_id] init_method = [nn.init.xavier_uniform_, nn.init.orthogonal_][use_orthogonal] gain = nn.init.calculate_gain(['tanh', 'relu', 'leaky_relu', 'leaky_relu'][activation_id]) def init_(m): return init(m, init_method, lambda x: nn.init.constant_(x, 0), gain=gain) self.fc1 = nn.Sequential( init_(nn.Linear(input_dim, hidden_size)), active_func, nn.LayerNorm(hidden_size)) self.fc_h = nn.Sequential(init_( nn.Linear(hidden_size, hidden_size)), active_func, nn.LayerNorm(hidden_size)) self.fc2 = get_clones(self.fc_h, self._layer_N) def forward(self, x): x = self.fc1(x) for i in range(self._layer_N): x = self.fc2[i](x) return x class MLPBase(nn.Module): def __init__(self, args, obs_shape, use_attn_internal=False, use_cat_self=True): super(MLPBase, self).__init__() self._use_feature_normalization = args.use_feature_normalization self._use_orthogonal = args.use_orthogonal self._activation_id = args.activation_id self._use_conv1d = args.use_conv1d self._stacked_frames = args.stacked_frames self._layer_N = args.layer_N self.hidden_size = args.hidden_size obs_dim = obs_shape[0] inputs_dim = obs_dim if self._use_feature_normalization: self.feature_norm = nn.LayerNorm(obs_dim) self.mlp = MLPLayer(inputs_dim, self.hidden_size, self._layer_N, self._use_orthogonal, self._activation_id) def forward(self, x): if self._use_feature_normalization: x = self.feature_norm(x) x = self.mlp(x) return x @property def output_size(self): return self.hidden_size
2,116
32.603175
98
py
TiKick
TiKick-main/tmarl/networks/utils/popart.py
import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F class PopArt(torch.nn.Module): def __init__(self, input_shape, output_shape, norm_axes=1, beta=0.99999, epsilon=1e-5, device=torch.device("cpu")): super(PopArt, self).__init__() self.beta = beta self.epsilon = epsilon self.norm_axes = norm_axes self.tpdv = dict(dtype=torch.float32, device=device) self.input_shape = input_shape self.output_shape = output_shape self.weight = nn.Parameter(torch.Tensor(output_shape, input_shape)).to(**self.tpdv) self.bias = nn.Parameter(torch.Tensor(output_shape)).to(**self.tpdv) self.stddev = nn.Parameter(torch.ones(output_shape), requires_grad=False).to(**self.tpdv) self.mean = nn.Parameter(torch.zeros(output_shape), requires_grad=False).to(**self.tpdv) self.mean_sq = nn.Parameter(torch.zeros(output_shape), requires_grad=False).to(**self.tpdv) self.debiasing_term = nn.Parameter(torch.tensor(0.0), requires_grad=False).to(**self.tpdv) self.reset_parameters() def reset_parameters(self): torch.nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5)) if self.bias is not None: fan_in, _ = torch.nn.init._calculate_fan_in_and_fan_out(self.weight) bound = 1 / math.sqrt(fan_in) torch.nn.init.uniform_(self.bias, -bound, bound) self.mean.zero_() self.mean_sq.zero_() self.debiasing_term.zero_() def forward(self, input_vector): if type(input_vector) == np.ndarray: input_vector = torch.from_numpy(input_vector) input_vector = input_vector.to(**self.tpdv) return F.linear(input_vector, self.weight, self.bias) @torch.no_grad() def update(self, input_vector): if type(input_vector) == np.ndarray: input_vector = torch.from_numpy(input_vector) input_vector = input_vector.to(**self.tpdv) old_mean, old_stddev = self.mean, self.stddev batch_mean = input_vector.mean(dim=tuple(range(self.norm_axes))) batch_sq_mean = (input_vector ** 2).mean(dim=tuple(range(self.norm_axes))) self.mean.mul_(self.beta).add_(batch_mean * (1.0 - self.beta)) self.mean_sq.mul_(self.beta).add_(batch_sq_mean * (1.0 - self.beta)) self.debiasing_term.mul_(self.beta).add_(1.0 * (1.0 - self.beta)) self.stddev = (self.mean_sq - self.mean ** 2).sqrt().clamp(min=1e-4) self.weight = self.weight * old_stddev / self.stddev self.bias = (old_stddev * self.bias + old_mean - self.mean) / self.stddev def debiased_mean_var(self): debiased_mean = self.mean / self.debiasing_term.clamp(min=self.epsilon) debiased_mean_sq = self.mean_sq / self.debiasing_term.clamp(min=self.epsilon) debiased_var = (debiased_mean_sq - debiased_mean ** 2).clamp(min=1e-2) return debiased_mean, debiased_var def normalize(self, input_vector): if type(input_vector) == np.ndarray: input_vector = torch.from_numpy(input_vector) input_vector = input_vector.to(**self.tpdv) mean, var = self.debiased_mean_var() out = (input_vector - mean[(None,) * self.norm_axes]) / torch.sqrt(var)[(None,) * self.norm_axes] return out def denormalize(self, input_vector): if type(input_vector) == np.ndarray: input_vector = torch.from_numpy(input_vector) input_vector = input_vector.to(**self.tpdv) mean, var = self.debiased_mean_var() out = input_vector * torch.sqrt(var)[(None,) * self.norm_axes] + mean[(None,) * self.norm_axes] out = out.cpu().numpy() return out
3,796
38.968421
119
py
TiKick
TiKick-main/tmarl/networks/utils/util.py
import copy import numpy as np import torch import torch.nn as nn def init(module, weight_init, bias_init, gain=1): weight_init(module.weight.data, gain=gain) bias_init(module.bias.data) return module def get_clones(module, N): return nn.ModuleList([copy.deepcopy(module) for i in range(N)]) def check(input): output = torch.from_numpy(input) if type(input) == np.ndarray else input return output
426
21.473684
76
py
TiKick
TiKick-main/tmarl/networks/utils/act.py
from .distributions import Bernoulli, Categorical, DiagGaussian import torch import torch.nn as nn class ACTLayer(nn.Module): def __init__(self, action_space, inputs_dim, use_orthogonal, gain): super(ACTLayer, self).__init__() self.multidiscrete_action = False self.continuous_action = False self.mixed_action = False if action_space.__class__.__name__ == "Discrete": action_dim = action_space.n self.action_out = Categorical(inputs_dim, action_dim, use_orthogonal, gain) elif action_space.__class__.__name__ == "Box": self.continuous_action = True action_dim = action_space.shape[0] self.action_out = DiagGaussian(inputs_dim, action_dim, use_orthogonal, gain) elif action_space.__class__.__name__ == "MultiBinary": action_dim = action_space.shape[0] self.action_out = Bernoulli(inputs_dim, action_dim, use_orthogonal, gain) elif action_space.__class__.__name__ == "MultiDiscrete": self.multidiscrete_action = True action_dims = action_space.high - action_space.low + 1 self.action_outs = [] for action_dim in action_dims: self.action_outs.append(Categorical(inputs_dim, action_dim, use_orthogonal, gain)) self.action_outs = nn.ModuleList(self.action_outs) else: # discrete + continous self.mixed_action = True continous_dim = action_space[0].shape[0] discrete_dim = action_space[1].n self.action_outs = nn.ModuleList([DiagGaussian(inputs_dim, continous_dim, use_orthogonal, gain), Categorical( inputs_dim, discrete_dim, use_orthogonal, gain)]) def forward(self, x, available_actions=None, deterministic=False): if self.mixed_action : actions = [] action_log_probs = [] for action_out in self.action_outs: action_logit = action_out(x) action = action_logit.mode() if deterministic else action_logit.sample() action_log_prob = action_logit.log_probs(action) actions.append(action.float()) action_log_probs.append(action_log_prob) actions = torch.cat(actions, -1) action_log_probs = torch.sum(torch.cat(action_log_probs, -1), -1, keepdim=True) elif self.multidiscrete_action: actions = [] action_log_probs = [] for action_out in self.action_outs: action_logit = action_out(x) action = action_logit.mode() if deterministic else action_logit.sample() action_log_prob = action_logit.log_probs(action) actions.append(action) action_log_probs.append(action_log_prob) actions = torch.cat(actions, -1) action_log_probs = torch.cat(action_log_probs, -1) elif self.continuous_action: action_logits = self.action_out(x) actions = action_logits.mode() if deterministic else action_logits.sample() action_log_probs = action_logits.log_probs(actions) else: action_logits = self.action_out(x, available_actions) actions = action_logits.mode() if deterministic else action_logits.sample() action_log_probs = action_logits.log_probs(actions) return actions, action_log_probs def get_probs(self, x, available_actions=None): if self.mixed_action or self.multidiscrete_action: action_probs = [] for action_out in self.action_outs: action_logit = action_out(x) action_prob = action_logit.probs action_probs.append(action_prob) action_probs = torch.cat(action_probs, -1) elif self.continuous_action: action_logits = self.action_out(x) action_probs = action_logits.probs else: action_logits = self.action_out(x, available_actions) action_probs = action_logits.probs return action_probs def get_log_1mp(self, x, action, available_actions=None, active_masks=None): action_logits = self.action_out(x, available_actions) action_prob = torch.gather(action_logits.probs, 1, action.long()) action_prob = torch.clamp(action_prob, 0, 1-1e-6) action_log_1mp = torch.log(1 - action_prob) return action_log_1mp def evaluate_actions(self, x, action, available_actions=None, active_masks=None): if self.mixed_action: a, b = action.split((2, 1), -1) b = b.long() action = [a, b] action_log_probs = [] dist_entropy = [] for action_out, act in zip(self.action_outs, action): action_logit = action_out(x) action_log_probs.append(action_logit.log_probs(act)) if active_masks is not None: if len(action_logit.entropy().shape) == len(active_masks.shape): dist_entropy.append((action_logit.entropy() * active_masks).sum()/active_masks.sum()) else: dist_entropy.append((action_logit.entropy() * active_masks.squeeze(-1)).sum()/active_masks.sum()) else: dist_entropy.append(action_logit.entropy().mean()) action_log_probs = torch.sum(torch.cat(action_log_probs, -1), -1, keepdim=True) dist_entropy = dist_entropy[0] * 0.0025 + dist_entropy[1] * 0.01 elif self.multidiscrete_action: action = torch.transpose(action, 0, 1) action_log_probs = [] dist_entropy = [] for action_out, act in zip(self.action_outs, action): action_logit = action_out(x) action_log_probs.append(action_logit.log_probs(act)) if active_masks is not None: dist_entropy.append((action_logit.entropy()*active_masks.squeeze(-1)).sum()/active_masks.sum()) else: dist_entropy.append(action_logit.entropy().mean()) action_log_probs = torch.cat(action_log_probs, -1) # ! could be wrong dist_entropy = torch.tensor(dist_entropy).mean() elif self.continuous_action: action_logits = self.action_out(x) action_log_probs = action_logits.log_probs(action) if active_masks is not None: dist_entropy = (action_logits.entropy()*active_masks).sum()/active_masks.sum() else: dist_entropy = action_logits.entropy().mean() else: action_logits = self.action_out(x, available_actions) action_log_probs = action_logits.log_probs(action) if active_masks is not None: dist_entropy = (action_logits.entropy()*active_masks.squeeze(-1)).sum()/active_masks.sum() else: dist_entropy = action_logits.entropy().mean() return action_log_probs, dist_entropy
7,195
46.342105
121
py
TiKick
TiKick-main/tmarl/networks/utils/rnn.py
import torch import torch.nn as nn class RNNLayer(nn.Module): def __init__(self, inputs_dim, outputs_dim, recurrent_N, use_orthogonal): super(RNNLayer, self).__init__() self._recurrent_N = recurrent_N self._use_orthogonal = use_orthogonal self.rnn = nn.GRU(inputs_dim, outputs_dim, num_layers=self._recurrent_N) for name, param in self.rnn.named_parameters(): if 'bias' in name: nn.init.constant_(param, 0) elif 'weight' in name: if self._use_orthogonal: nn.init.orthogonal_(param) else: nn.init.xavier_uniform_(param) self.norm = nn.LayerNorm(outputs_dim) def forward(self, x, hxs, masks): if x.size(0) == hxs.size(0): x, hxs = self.rnn(x.unsqueeze(0), (hxs * masks.repeat(1, self._recurrent_N).unsqueeze(-1)).transpose(0, 1).contiguous()) x = x.squeeze(0) hxs = hxs.transpose(0, 1) else: # x is a (T, N, -1) tensor that has been flatten to (T * N, -1) N = hxs.size(0) T = int(x.size(0) / N) # unflatten x = x.view(T, N, x.size(1)) # Same deal with masks masks = masks.view(T, N) # Let's figure out which steps in the sequence have a zero for any agent # We will always assume t=0 has a zero in it as that makes the logic cleaner has_zeros = ((masks[1:] == 0.0) .any(dim=-1) .nonzero() .squeeze() .cpu()) # +1 to correct the masks[1:] if has_zeros.dim() == 0: # Deal with scalar has_zeros = [has_zeros.item() + 1] else: has_zeros = (has_zeros + 1).numpy().tolist() # add t=0 and t=T to the list has_zeros = [0] + has_zeros + [T] hxs = hxs.transpose(0, 1) outputs = [] for i in range(len(has_zeros) - 1): # We can now process steps that don't have any zeros in masks together! # This is much faster start_idx = has_zeros[i] end_idx = has_zeros[i + 1] temp = (hxs * masks[start_idx].view(1, -1, 1).repeat(self._recurrent_N, 1, 1)).contiguous() rnn_scores, hxs = self.rnn(x[start_idx:end_idx], temp) outputs.append(rnn_scores) # assert len(outputs) == T # x is a (T, N, -1) tensor x = torch.cat(outputs, dim=0) # flatten x = x.reshape(T * N, -1) hxs = hxs.transpose(0, 1) x = self.norm(x) return x, hxs
2,816
34.2125
132
py
TiKick
TiKick-main/tmarl/drivers/shared_distributed/base_driver.py
import numpy as np import torch def _t2n(x): return x.detach().cpu().numpy() class Driver(object): def __init__(self, config, client=None): self.all_args = config['all_args'] self.envs = config['envs'] self.eval_envs = config['eval_envs'] self.device = config['device'] self.num_agents = config['num_agents'] if 'signal' in config: self.actor_id = config['signal'].actor_id self.weight_ids = config['signal'].weight_ids else: self.actor_id = 0 self.weight_ids = [0] # parameters self.env_name = self.all_args.env_name self.algorithm_name = self.all_args.algorithm_name self.experiment_name = self.all_args.experiment_name self.use_centralized_V = self.all_args.use_centralized_V self.use_obs_instead_of_state = self.all_args.use_obs_instead_of_state self.num_env_steps = self.all_args.num_env_steps if hasattr(self.all_args,'num_env_steps') else self.all_args.eval_num self.episode_length = self.all_args.episode_length self.n_rollout_threads = self.all_args.n_rollout_threads self.learner_n_rollout_threads = self.all_args.n_rollout_threads self.n_eval_rollout_threads = self.all_args.n_eval_rollout_threads self.hidden_size = self.all_args.hidden_size self.recurrent_N = self.all_args.recurrent_N # interval self.save_interval = self.all_args.save_interval self.use_eval = self.all_args.use_eval self.eval_interval = self.all_args.eval_interval self.log_interval = self.all_args.log_interval # dir self.model_dir = self.all_args.model_dir if self.algorithm_name == "rmappo": from tmarl.algorithms.r_mappo_distributed.mappo_algorithm import MAPPOAlgorithm as TrainAlgo from tmarl.algorithms.r_mappo_distributed.mappo_module import MAPPOModule as AlgoModule else: raise NotImplementedError if self.envs: share_observation_space = self.envs.share_observation_space[0] \ if self.use_centralized_V else self.envs.observation_space[0] # policy network self.algo_module = AlgoModule(self.all_args, self.envs.observation_space[0], share_observation_space, self.envs.action_space[0], device=self.device) else: share_observation_space = self.eval_envs.share_observation_space[0] \ if self.use_centralized_V else self.eval_envs.observation_space[0] # policy network self.algo_module = AlgoModule(self.all_args, self.eval_envs.observation_space[0], share_observation_space, self.eval_envs.action_space[0], device=self.device) if self.model_dir is not None: self.restore() # algorithm self.trainer = TrainAlgo(self.all_args, self.algo_module, device=self.device) # buffer from tmarl.replay_buffers.normal.shared_buffer import SharedReplayBuffer self.buffer = SharedReplayBuffer(self.all_args, self.num_agents, self.envs.observation_space[0] if self.envs else self.eval_envs.observation_space[0], share_observation_space, self.envs.action_space[0] if self.envs else self.eval_envs.action_space[0]) def run(self): raise NotImplementedError def warmup(self): raise NotImplementedError def collect(self, step): raise NotImplementedError def insert(self, data): raise NotImplementedError def restore(self): policy_actor_state_dict = torch.load(str(self.model_dir) + '/actor.pt', map_location=self.device) self.algo_module.actor.load_state_dict(policy_actor_state_dict)
4,244
39.04717
126
py
TiKick
TiKick-main/tmarl/algorithms/r_mappo_distributed/mappo_algorithm.py
import torch from tmarl.utils.valuenorm import ValueNorm # implement the loss of the MAPPO here class MAPPOAlgorithm(): def __init__(self, args, init_module, device=torch.device("cpu")): self.device = device self.tpdv = dict(dtype=torch.float32, device=device) self.algo_module = init_module self.clip_param = args.clip_param self.ppo_epoch = args.ppo_epoch self.num_mini_batch = args.num_mini_batch self.data_chunk_length = args.data_chunk_length self.policy_value_loss_coef = args.policy_value_loss_coef self.value_loss_coef = args.value_loss_coef self.entropy_coef = args.entropy_coef self.max_grad_norm = args.max_grad_norm self.huber_delta = args.huber_delta self._use_recurrent_policy = args.use_recurrent_policy self._use_naive_recurrent = args.use_naive_recurrent_policy self._use_max_grad_norm = args.use_max_grad_norm self._use_clipped_value_loss = args.use_clipped_value_loss self._use_huber_loss = args.use_huber_loss self._use_popart = args.use_popart self._use_valuenorm = args.use_valuenorm self._use_value_active_masks = args.use_value_active_masks self._use_policy_active_masks = args.use_policy_active_masks self._use_policy_vhead = args.use_policy_vhead assert (self._use_popart and self._use_valuenorm) == False, ("self._use_popart and self._use_valuenorm can not be set True simultaneously") if self._use_popart: self.value_normalizer = self.algo_module.critic.v_out if self._use_policy_vhead: self.policy_value_normalizer = self.algo_module.actor.v_out elif self._use_valuenorm: self.value_normalizer = ValueNorm(1, device = self.device) if self._use_policy_vhead: self.policy_value_normalizer = ValueNorm(1, device = self.device) else: self.value_normalizer = None if self._use_policy_vhead: self.policy_value_normalizer = None def prep_rollout(self): self.algo_module.actor.eval()
2,234
38.210526
147
py
TiKick
TiKick-main/tmarl/algorithms/r_mappo_distributed/mappo_module.py
import torch from tmarl.networks.policy_network import PolicyNetwork class MAPPOModule: def __init__(self, args, obs_space, share_obs_space, act_space, device=torch.device("cpu")): self.device = device self.lr = args.lr self.critic_lr = args.critic_lr self.opti_eps = args.opti_eps self.weight_decay = args.weight_decay self.obs_space = obs_space self.share_obs_space = share_obs_space self.act_space = act_space self.actor = PolicyNetwork(args, self.obs_space, self.act_space, self.device) self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=self.lr, eps=self.opti_eps, weight_decay=self.weight_decay) def get_actions(self, share_obs, obs, rnn_states_actor, rnn_states_critic, masks, available_actions=None, deterministic=False): actions, action_log_probs, rnn_states_actor = self.actor(obs, rnn_states_actor, masks, available_actions, deterministic) return None, actions, action_log_probs, rnn_states_actor, None
1,050
41.04
135
py
TiKick
TiKick-main/tmarl/replay_buffers/normal/shared_buffer.py
import torch import numpy as np from collections import defaultdict from tmarl.utils.util import check,get_shape_from_obs_space, get_shape_from_act_space def _flatten(T, N, x): return x.reshape(T * N, *x.shape[2:]) def _cast(x): return x.transpose(1, 2, 0, 3).reshape(-1, *x.shape[3:]) class SharedReplayBuffer(object): def __init__(self, args, num_agents, obs_space, share_obs_space, act_space): self.episode_length = args.episode_length self.n_rollout_threads = args.n_rollout_threads self.hidden_size = args.hidden_size self.recurrent_N = args.recurrent_N self.gamma = args.gamma self.gae_lambda = args.gae_lambda self._use_gae = args.use_gae self._use_popart = args.use_popart self._use_valuenorm = args.use_valuenorm self._use_proper_time_limits = args.use_proper_time_limits self._mixed_obs = False # for mixed observation obs_shape = get_shape_from_obs_space(obs_space) share_obs_shape = get_shape_from_obs_space(share_obs_space) # for mixed observation if 'Dict' in obs_shape.__class__.__name__: self._mixed_obs = True self.obs = {} self.share_obs = {} for key in obs_shape: self.obs[key] = np.zeros((self.episode_length + 1, self.n_rollout_threads, num_agents, *obs_shape[key].shape), dtype=np.float32) for key in share_obs_shape: self.share_obs[key] = np.zeros((self.episode_length + 1, self.n_rollout_threads, num_agents, *share_obs_shape[key].shape), dtype=np.float32) else: # deal with special attn format if type(obs_shape[-1]) == list: obs_shape = obs_shape[:1] if type(share_obs_shape[-1]) == list: share_obs_shape = share_obs_shape[:1] self.share_obs = np.zeros((self.episode_length + 1, self.n_rollout_threads, num_agents, *share_obs_shape), dtype=np.float32) self.obs = np.zeros((self.episode_length + 1, self.n_rollout_threads, num_agents, *obs_shape), dtype=np.float32) self.rnn_states = np.zeros((self.episode_length + 1, self.n_rollout_threads, num_agents, self.recurrent_N, self.hidden_size), dtype=np.float32) self.rnn_states_critic = np.zeros_like(self.rnn_states) self.value_preds = np.zeros( (self.episode_length + 1, self.n_rollout_threads, num_agents, 1), dtype=np.float32) self.returns = np.zeros_like(self.value_preds) if act_space.__class__.__name__ == 'Discrete': self.available_actions = np.ones((self.episode_length + 1, self.n_rollout_threads, num_agents, act_space.n), dtype=np.float32) else: self.available_actions = None act_shape = get_shape_from_act_space(act_space) self.actions = np.zeros( (self.episode_length, self.n_rollout_threads, num_agents, act_shape), dtype=np.float32) self.action_log_probs = np.zeros( (self.episode_length, self.n_rollout_threads, num_agents, act_shape), dtype=np.float32) self.rewards = np.zeros( (self.episode_length, self.n_rollout_threads, num_agents, 1), dtype=np.float32) self.masks = np.ones((self.episode_length + 1, self.n_rollout_threads, num_agents, 1), dtype=np.float32) self.bad_masks = np.ones_like(self.masks) self.active_masks = np.ones_like(self.masks) self.step = 0 def insert(self, share_obs, obs, rnn_states, rnn_states_critic, actions, action_log_probs, value_preds, rewards, masks, bad_masks=None, active_masks=None, available_actions=None): if self._mixed_obs: for key in self.share_obs.keys(): self.share_obs[key][self.step + 1] = share_obs[key].copy() for key in self.obs.keys(): self.obs[key][self.step + 1] = obs[key].copy() else: self.share_obs[self.step + 1] = share_obs.copy() self.obs[self.step + 1] = obs.copy() self.rnn_states[self.step + 1] = rnn_states.copy() self.rnn_states_critic[self.step + 1] = rnn_states_critic.copy() self.actions[self.step] = actions.copy() self.action_log_probs[self.step] = action_log_probs.copy() self.value_preds[self.step] = value_preds.copy() self.rewards[self.step] = rewards.copy() self.masks[self.step + 1] = masks.copy() if bad_masks is not None: self.bad_masks[self.step + 1] = bad_masks.copy() if active_masks is not None: self.active_masks[self.step + 1] = active_masks.copy() if available_actions is not None: self.available_actions[self.step + 1] = available_actions.copy() self.step = (self.step + 1) % self.episode_length def init_buffer(self,share_obs,obs): self.share_obs[0] = share_obs self.obs[0] = obs def chooseinsert(self, share_obs, obs, rnn_states, rnn_states_critic, actions, action_log_probs, value_preds, rewards, masks, bad_masks=None, active_masks=None, available_actions=None): self.share_obs[self.step] = share_obs.copy() self.obs[self.step] = obs.copy() self.rnn_states[self.step + 1] = rnn_states.copy() self.rnn_states_critic[self.step + 1] = rnn_states_critic.copy() self.actions[self.step] = actions.copy() self.action_log_probs[self.step] = action_log_probs.copy() self.value_preds[self.step] = value_preds.copy() self.rewards[self.step] = rewards.copy() self.masks[self.step + 1] = masks.copy() if bad_masks is not None: self.bad_masks[self.step + 1] = bad_masks.copy() if active_masks is not None: self.active_masks[self.step] = active_masks.copy() if available_actions is not None: self.available_actions[self.step] = available_actions.copy() self.step = (self.step + 1) % self.episode_length def after_update(self): if self._mixed_obs: for key in self.share_obs.keys(): self.share_obs[key][0] = self.share_obs[key][-1].copy() for key in self.obs.keys(): self.obs[key][0] = self.obs[key][-1].copy() else: self.share_obs[0] = self.share_obs[-1].copy() self.obs[0] = self.obs[-1].copy() self.rnn_states[0] = self.rnn_states[-1].copy() self.rnn_states_critic[0] = self.rnn_states_critic[-1].copy() self.masks[0] = self.masks[-1].copy() self.bad_masks[0] = self.bad_masks[-1].copy() self.active_masks[0] = self.active_masks[-1].copy() if self.available_actions is not None: self.available_actions[0] = self.available_actions[-1].copy() def chooseafter_update(self): self.rnn_states[0] = self.rnn_states[-1].copy() self.rnn_states_critic[0] = self.rnn_states_critic[-1].copy() self.masks[0] = self.masks[-1].copy() self.bad_masks[0] = self.bad_masks[-1].copy() def compute_returns(self, next_value, value_normalizer=None): if self._use_proper_time_limits: if self._use_gae: self.value_preds[-1] = next_value gae = 0 for step in reversed(range(self.rewards.shape[0])): if self._use_popart or self._use_valuenorm: # step + 1 delta = self.rewards[step] + self.gamma * value_normalizer.denormalize(self.value_preds[step + 1]) * self.masks[step + 1] \ - value_normalizer.denormalize(self.value_preds[step]) gae = delta + self.gamma * self.gae_lambda * gae * self.masks[step + 1] gae = gae * self.bad_masks[step + 1] self.returns[step] = gae + value_normalizer.denormalize(self.value_preds[step]) else: delta = self.rewards[step] + self.gamma * self.value_preds[step + 1] * self.masks[step + 1] - self.value_preds[step] gae = delta + self.gamma * self.gae_lambda * self.masks[step + 1] * gae gae = gae * self.bad_masks[step + 1] self.returns[step] = gae + self.value_preds[step] else: self.returns[-1] = next_value for step in reversed(range(self.rewards.shape[0])): if self._use_popart or self._use_valuenorm: self.returns[step] = (self.returns[step + 1] * self.gamma * self.masks[step + 1] + self.rewards[step]) * self.bad_masks[step + 1] \ + (1 - self.bad_masks[step + 1]) * value_normalizer.denormalize(self.value_preds[step]) else: self.returns[step] = (self.returns[step + 1] * self.gamma * self.masks[step + 1] + self.rewards[step]) * self.bad_masks[step + 1] \ + (1 - self.bad_masks[step + 1]) * self.value_preds[step] else: if self._use_gae: self.value_preds[-1] = next_value gae = 0 for step in reversed(range(self.rewards.shape[0])): if self._use_popart or self._use_valuenorm: delta = self.rewards[step] + self.gamma * value_normalizer.denormalize(self.value_preds[step + 1]) * self.masks[step + 1] \ - value_normalizer.denormalize(self.value_preds[step]) gae = delta + self.gamma * self.gae_lambda * self.masks[step + 1] * gae self.returns[step] = gae + value_normalizer.denormalize(self.value_preds[step]) else: delta = self.rewards[step] + self.gamma * self.value_preds[step + 1] * self.masks[step + 1] - self.value_preds[step] gae = delta + self.gamma * self.gae_lambda * self.masks[step + 1] * gae self.returns[step] = gae + self.value_preds[step] else: self.returns[-1] = next_value for step in reversed(range(self.rewards.shape[0])): self.returns[step] = self.returns[step + 1] * self.gamma * self.masks[step + 1] + self.rewards[step] def feed_forward_generator(self, advantages, num_mini_batch=None, mini_batch_size=None): episode_length, n_rollout_threads, num_agents = self.rewards.shape[0:3] batch_size = n_rollout_threads * episode_length * num_agents if mini_batch_size is None: assert batch_size >= num_mini_batch, ( "PPO requires the number of processes ({}) " "* number of steps ({}) * number of agents ({}) = {} " "to be greater than or equal to the number of PPO mini batches ({})." "".format(n_rollout_threads, episode_length, num_agents, n_rollout_threads * episode_length * num_agents, num_mini_batch)) mini_batch_size = batch_size // num_mini_batch rand = torch.randperm(batch_size).numpy() sampler = [rand[i*mini_batch_size:(i+1)*mini_batch_size] for i in range(num_mini_batch)] if self._mixed_obs: share_obs = {} obs = {} for key in self.share_obs.keys(): share_obs[key] = self.share_obs[key][:-1].reshape(-1, *self.share_obs[key].shape[3:]) for key in self.obs.keys(): obs[key] = self.obs[key][:-1].reshape(-1, *self.obs[key].shape[3:]) else: share_obs = self.share_obs[:-1].reshape(-1, *self.share_obs.shape[3:]) obs = self.obs[:-1].reshape(-1, *self.obs.shape[3:]) rnn_states = self.rnn_states[:-1].reshape(-1, *self.rnn_states.shape[3:]) rnn_states_critic = self.rnn_states_critic[:-1].reshape(-1, *self.rnn_states_critic.shape[3:]) actions = self.actions.reshape(-1, self.actions.shape[-1]) if self.available_actions is not None: available_actions = self.available_actions[:-1].reshape(-1, self.available_actions.shape[-1]) value_preds = self.value_preds[:-1].reshape(-1, 1) returns = self.returns[:-1].reshape(-1, 1) masks = self.masks[:-1].reshape(-1, 1) active_masks = self.active_masks[:-1].reshape(-1, 1) action_log_probs = self.action_log_probs.reshape(-1, self.action_log_probs.shape[-1]) advantages = advantages.reshape(-1, 1) for indices in sampler: # obs size [T+1 N M Dim]-->[T N M Dim]-->[T*N*M,Dim]-->[index,Dim] if self._mixed_obs: share_obs_batch = {} obs_batch = {} for key in share_obs.keys(): share_obs_batch[key] = share_obs[key][indices] for key in obs.keys(): obs_batch[key] = obs[key][indices] else: share_obs_batch = share_obs[indices] obs_batch = obs[indices] rnn_states_batch = rnn_states[indices] rnn_states_critic_batch = rnn_states_critic[indices] actions_batch = actions[indices] if self.available_actions is not None: available_actions_batch = available_actions[indices] else: available_actions_batch = None value_preds_batch = value_preds[indices] return_batch = returns[indices] masks_batch = masks[indices] active_masks_batch = active_masks[indices] old_action_log_probs_batch = action_log_probs[indices] if advantages is None: adv_targ = None else: adv_targ = advantages[indices] yield share_obs_batch, obs_batch, rnn_states_batch, rnn_states_critic_batch, actions_batch, value_preds_batch, return_batch, masks_batch, active_masks_batch, old_action_log_probs_batch, adv_targ, available_actions_batch def naive_recurrent_generator(self, advantages, num_mini_batch): episode_length, n_rollout_threads, num_agents = self.rewards.shape[0:3] batch_size = n_rollout_threads*num_agents assert n_rollout_threads*num_agents >= num_mini_batch, ( "PPO requires the number of processes ({})* number of agents ({}) " "to be greater than or equal to the number of " "PPO mini batches ({}).".format(n_rollout_threads, num_agents, num_mini_batch)) num_envs_per_batch = batch_size // num_mini_batch perm = torch.randperm(batch_size).numpy() if self._mixed_obs: share_obs = {} obs = {} for key in self.share_obs.keys(): share_obs[key] = self.share_obs[key].reshape(-1, batch_size, *self.share_obs[key].shape[3:]) for key in self.obs.keys(): obs[key] = self.obs[key].reshape(-1, batch_size, *self.obs[key].shape[3:]) else: share_obs = self.share_obs.reshape(-1, batch_size, *self.share_obs.shape[3:]) obs = self.obs.reshape(-1, batch_size, *self.obs.shape[3:]) rnn_states = self.rnn_states.reshape(-1, batch_size, *self.rnn_states.shape[3:]) rnn_states_critic = self.rnn_states_critic.reshape(-1, batch_size, *self.rnn_states_critic.shape[3:]) actions = self.actions.reshape(-1, batch_size, self.actions.shape[-1]) if self.available_actions is not None: available_actions = self.available_actions.reshape(-1, batch_size, self.available_actions.shape[-1]) value_preds = self.value_preds.reshape(-1, batch_size, 1) returns = self.returns.reshape(-1, batch_size, 1) masks = self.masks.reshape(-1, batch_size, 1) active_masks = self.active_masks.reshape(-1, batch_size, 1) action_log_probs = self.action_log_probs.reshape(-1, batch_size, self.action_log_probs.shape[-1]) advantages = advantages.reshape(-1, batch_size, 1) for start_ind in range(0, batch_size, num_envs_per_batch): if self._mixed_obs: share_obs_batch = defaultdict(list) obs_batch = defaultdict(list) else: share_obs_batch = [] obs_batch = [] rnn_states_batch = [] rnn_states_critic_batch = [] actions_batch = [] available_actions_batch = [] value_preds_batch = [] return_batch = [] masks_batch = [] active_masks_batch = [] old_action_log_probs_batch = [] adv_targ = [] for offset in range(num_envs_per_batch): ind = perm[start_ind + offset] if self._mixed_obs: for key in share_obs.keys(): share_obs_batch[key].append(share_obs[key][:-1, ind]) for key in obs.keys(): obs_batch[key].append(obs[key][:-1, ind]) else: share_obs_batch.append(share_obs[:-1, ind]) obs_batch.append(obs[:-1, ind]) rnn_states_batch.append(rnn_states[0:1, ind]) rnn_states_critic_batch.append(rnn_states_critic[0:1, ind]) actions_batch.append(actions[:, ind]) if self.available_actions is not None: available_actions_batch.append(available_actions[:-1, ind]) value_preds_batch.append(value_preds[:-1, ind]) return_batch.append(returns[:-1, ind]) masks_batch.append(masks[:-1, ind]) active_masks_batch.append(active_masks[:-1, ind]) old_action_log_probs_batch.append(action_log_probs[:, ind]) adv_targ.append(advantages[:, ind]) # [N[T, dim]] T, N = self.episode_length, num_envs_per_batch # These are all from_numpys of size (T, N, -1) if self._mixed_obs: for key in share_obs_batch.keys(): share_obs_batch[key] = np.stack(share_obs_batch[key], 1) for key in obs_batch.keys(): obs_batch[key] = np.stack(obs_batch[key], 1) else: share_obs_batch = np.stack(share_obs_batch, 1) obs_batch = np.stack(obs_batch, 1) actions_batch = np.stack(actions_batch, 1) if self.available_actions is not None: available_actions_batch = np.stack(available_actions_batch, 1) value_preds_batch = np.stack(value_preds_batch, 1) return_batch = np.stack(return_batch, 1) masks_batch = np.stack(masks_batch, 1) active_masks_batch = np.stack(active_masks_batch, 1) old_action_log_probs_batch = np.stack(old_action_log_probs_batch, 1) adv_targ = np.stack(adv_targ, 1) # States is just a (N, dim) from_numpy [N[1,dim]] rnn_states_batch = np.stack(rnn_states_batch).reshape(N, *self.rnn_states.shape[3:]) rnn_states_critic_batch = np.stack(rnn_states_critic_batch).reshape(N, *self.rnn_states_critic.shape[3:]) # Flatten the (T, N, ...) from_numpys to (T * N, ...) if self._mixed_obs: for key in share_obs_batch.keys(): share_obs_batch[key] = _flatten(T, N, share_obs_batch[key]) for key in obs_batch.keys(): obs_batch[key] = _flatten(T, N, obs_batch[key]) else: share_obs_batch = _flatten(T, N, share_obs_batch) obs_batch = _flatten(T, N, obs_batch) actions_batch = _flatten(T, N, actions_batch) if self.available_actions is not None: available_actions_batch = _flatten(T, N, available_actions_batch) else: available_actions_batch = None value_preds_batch = _flatten(T, N, value_preds_batch) return_batch = _flatten(T, N, return_batch) masks_batch = _flatten(T, N, masks_batch) active_masks_batch = _flatten(T, N, active_masks_batch) old_action_log_probs_batch = _flatten(T, N, old_action_log_probs_batch) adv_targ = _flatten(T, N, adv_targ) yield share_obs_batch, obs_batch, rnn_states_batch, rnn_states_critic_batch, actions_batch, value_preds_batch, return_batch, masks_batch, active_masks_batch, old_action_log_probs_batch, adv_targ, available_actions_batch def recurrent_generator(self, advantages, num_mini_batch, data_chunk_length): episode_length, n_rollout_threads, num_agents = self.rewards.shape[0:3] batch_size = n_rollout_threads * episode_length * num_agents data_chunks = batch_size // data_chunk_length # [C=r*T*M/L] mini_batch_size = data_chunks // num_mini_batch assert n_rollout_threads * episode_length * num_agents >= data_chunk_length, ( "PPO requires the number of processes ({})* number of agents ({}) * episode length ({}) " "to be greater than or equal to the number of " "data chunk length ({}).".format(n_rollout_threads, num_agents, episode_length ,data_chunk_length)) rand = torch.randperm(data_chunks).numpy() sampler = [rand[i*mini_batch_size:(i+1)*mini_batch_size] for i in range(num_mini_batch)] if self._mixed_obs: share_obs = {} obs = {} for key in self.share_obs.keys(): if len(self.share_obs[key].shape) == 6: share_obs[key] = self.share_obs[key][:-1].transpose(1, 2, 0, 3, 4, 5).reshape(-1, *self.share_obs[key].shape[3:]) elif len(self.share_obs[key].shape) == 5: share_obs[key] = self.share_obs[key][:-1].transpose(1, 2, 0, 3, 4).reshape(-1, *self.share_obs[key].shape[3:]) else: share_obs[key] = _cast(self.share_obs[key][:-1]) for key in self.obs.keys(): if len(self.obs[key].shape) == 6: obs[key] = self.obs[key][:-1].transpose(1, 2, 0, 3, 4, 5).reshape(-1, *self.obs[key].shape[3:]) elif len(self.obs[key].shape) == 5: obs[key] = self.obs[key][:-1].transpose(1, 2, 0, 3, 4).reshape(-1, *self.obs[key].shape[3:]) else: obs[key] = _cast(self.obs[key][:-1]) else: if len(self.share_obs.shape) > 4: share_obs = self.share_obs[:-1].transpose(1, 2, 0, 3, 4, 5).reshape(-1, *self.share_obs.shape[3:]) obs = self.obs[:-1].transpose(1, 2, 0, 3, 4, 5).reshape(-1, *self.obs.shape[3:]) else: share_obs = _cast(self.share_obs[:-1]) obs = _cast(self.obs[:-1]) actions = _cast(self.actions) action_log_probs = _cast(self.action_log_probs) advantages = _cast(advantages) value_preds = _cast(self.value_preds[:-1]) returns = _cast(self.returns[:-1]) masks = _cast(self.masks[:-1]) active_masks = _cast(self.active_masks[:-1]) # rnn_states = _cast(self.rnn_states[:-1]) # rnn_states_critic = _cast(self.rnn_states_critic[:-1]) rnn_states = self.rnn_states[:-1].transpose(1, 2, 0, 3, 4).reshape(-1, *self.rnn_states.shape[3:]) rnn_states_critic = self.rnn_states_critic[:-1].transpose(1, 2, 0, 3, 4).reshape(-1, *self.rnn_states_critic.shape[3:]) if self.available_actions is not None: available_actions = _cast(self.available_actions[:-1]) for indices in sampler: if self._mixed_obs: share_obs_batch = defaultdict(list) obs_batch = defaultdict(list) else: share_obs_batch = [] obs_batch = [] rnn_states_batch = [] rnn_states_critic_batch = [] actions_batch = [] available_actions_batch = [] value_preds_batch = [] return_batch = [] masks_batch = [] active_masks_batch = [] old_action_log_probs_batch = [] adv_targ = [] for index in indices: ind = index * data_chunk_length # size [T+1 N M Dim]-->[T N M Dim]-->[N,M,T,Dim]-->[N*M*T,Dim]-->[L,Dim] if self._mixed_obs: for key in share_obs.keys(): share_obs_batch[key].append(share_obs[key][ind:ind+data_chunk_length]) for key in obs.keys(): obs_batch[key].append(obs[key][ind:ind+data_chunk_length]) else: share_obs_batch.append(share_obs[ind:ind+data_chunk_length]) obs_batch.append(obs[ind:ind+data_chunk_length]) actions_batch.append(actions[ind:ind+data_chunk_length]) if self.available_actions is not None: available_actions_batch.append(available_actions[ind:ind+data_chunk_length]) value_preds_batch.append(value_preds[ind:ind+data_chunk_length]) return_batch.append(returns[ind:ind+data_chunk_length]) masks_batch.append(masks[ind:ind+data_chunk_length]) active_masks_batch.append(active_masks[ind:ind+data_chunk_length]) old_action_log_probs_batch.append(action_log_probs[ind:ind+data_chunk_length]) adv_targ.append(advantages[ind:ind+data_chunk_length]) # size [T+1 N M Dim]-->[T N M Dim]-->[N M T Dim]-->[N*M*T,Dim]-->[1,Dim] rnn_states_batch.append(rnn_states[ind]) rnn_states_critic_batch.append(rnn_states_critic[ind]) L, N = data_chunk_length, mini_batch_size # These are all from_numpys of size (L, N, Dim) if self._mixed_obs: for key in share_obs_batch.keys(): share_obs_batch[key] = np.stack(share_obs_batch[key], axis=1) for key in obs_batch.keys(): obs_batch[key] = np.stack(obs_batch[key], axis=1) else: share_obs_batch = np.stack(share_obs_batch, axis=1) obs_batch = np.stack(obs_batch, axis=1) actions_batch = np.stack(actions_batch, axis=1) if self.available_actions is not None: available_actions_batch = np.stack(available_actions_batch, axis=1) value_preds_batch = np.stack(value_preds_batch, axis=1) return_batch = np.stack(return_batch, axis=1) masks_batch = np.stack(masks_batch, axis=1) active_masks_batch = np.stack(active_masks_batch, axis=1) old_action_log_probs_batch = np.stack(old_action_log_probs_batch, axis=1) adv_targ = np.stack(adv_targ, axis=1) # States is just a (N, -1) from_numpy rnn_states_batch = np.stack(rnn_states_batch).reshape(N, *self.rnn_states.shape[3:]) rnn_states_critic_batch = np.stack(rnn_states_critic_batch).reshape(N, *self.rnn_states_critic.shape[3:]) # Flatten the (L, N, ...) from_numpys to (L * N, ...) if self._mixed_obs: for key in share_obs_batch.keys(): share_obs_batch[key] = _flatten(L, N, share_obs_batch[key]) for key in obs_batch.keys(): obs_batch[key] = _flatten(L, N, obs_batch[key]) else: share_obs_batch = _flatten(L, N, share_obs_batch) obs_batch = _flatten(L, N, obs_batch) actions_batch = _flatten(L, N, actions_batch) if self.available_actions is not None: available_actions_batch = _flatten(L, N, available_actions_batch) else: available_actions_batch = None value_preds_batch = _flatten(L, N, value_preds_batch) return_batch = _flatten(L, N, return_batch) masks_batch = _flatten(L, N, masks_batch) active_masks_batch = _flatten(L, N, active_masks_batch) old_action_log_probs_batch = _flatten(L, N, old_action_log_probs_batch) adv_targ = _flatten(L, N, adv_targ) yield share_obs_batch, obs_batch, rnn_states_batch, rnn_states_critic_batch, actions_batch, value_preds_batch, return_batch, masks_batch, active_masks_batch, old_action_log_probs_batch, adv_targ, available_actions_batch
28,769
52.081181
231
py
TiKick
TiKick-main/tmarl/configs/config.py
#!/usr/bin/env python # -*- coding: utf-8 -*- # Copyright 2021 The TARTRL Authors. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """""" import argparse def get_config(): parser = argparse.ArgumentParser( description='TiKick', formatter_class=argparse.RawDescriptionHelpFormatter) # prepare parameters parser.add_argument("--algorithm_name", type=str, default='rmappo', choices=["rmappo"]) parser.add_argument("--experiment_name", type=str, default="check", help="an identifier to distinguish different experiment.") parser.add_argument("--seed", type=int, default=1, help="Random seed for numpy/torch") parser.add_argument("--disable_cuda", action='store_true', default=False, help="by default False, will use GPU to train; or else will use CPU;") parser.add_argument("--cuda_deterministic", action='store_false', default=True, help="by default, make sure random seed effective. if set, bypass such function.") parser.add_argument("--n_rollout_threads", type=int, default=2, help="Number of parallel envs for training rollout") parser.add_argument("--n_eval_rollout_threads", type=int, default=1, help="Number of parallel envs for evaluating rollout") parser.add_argument("--n_render_rollout_threads", type=int, default=1, help="Number of parallel envs for rendering rollout") parser.add_argument("--eval_num", type=int, default=1, help='Number of environment steps to evaluate (default: 1)') # env parameters parser.add_argument("--env_name", type=str, default='StarCraft2', help="specify the name of environment") parser.add_argument("--use_obs_instead_of_state", action='store_true', default=False, help="Whether to use global state or concatenated obs") # replay buffer parameters parser.add_argument("--episode_length", type=int, default=200, help="Max length for any episode") # network parameters parser.add_argument("--separate_policy", action='store_true', default=False, help='Whether agent seperate the policy') parser.add_argument("--use_centralized_V", action='store_false', default=True, help="Whether to use centralized V function") parser.add_argument("--use_conv1d", action='store_true', default=False, help="Whether to use conv1d") parser.add_argument("--stacked_frames", type=int, default=1, help="Dimension of hidden layers for actor/critic networks") parser.add_argument("--use_stacked_frames", action='store_true', default=False, help="Whether to use stacked_frames") parser.add_argument("--hidden_size", type=int, default=256, help="Dimension of hidden layers for actor/critic networks") # TODO @zoeyuchao. The same comment might in need of change. parser.add_argument("--layer_N", type=int, default=3, help="Number of layers for actor/critic networks") parser.add_argument("--activation_id", type=int, default=1, help="choose 0 to use tanh, 1 to use relu, 2 to use leaky relu, 3 to use elu") parser.add_argument("--use_popart", action='store_true', default=False, help="by default False, use PopArt to normalize rewards.") parser.add_argument("--use_valuenorm", action='store_false', default=True, help="by default True, use running mean and std to normalize rewards.") parser.add_argument("--use_feature_normalization", action='store_false', default=True, help="Whether to apply layernorm to the inputs") parser.add_argument("--use_orthogonal", action='store_false', default=True, help="Whether to use Orthogonal initialization for weights and 0 initialization for biases") parser.add_argument("--gain", type=float, default=0.01, help="The gain # of last action layer") parser.add_argument("--cnn_layers_params", type=str, default=None, help="The parameters of cnn layer") parser.add_argument("--use_maxpool2d", action='store_true', default=False, help="Whether to apply layernorm to the inputs") # recurrent parameters parser.add_argument("--use_naive_recurrent_policy", action='store_true', default=False, help='Whether to use a naive recurrent policy') parser.add_argument("--use_recurrent_policy", action='store_false', default=True, help='use a recurrent policy') parser.add_argument("--recurrent_N", type=int, default=1, help="The number of recurrent layers.") parser.add_argument("--data_chunk_length", type=int, default=25, help="Time length of chunks used to train a recurrent_policy") parser.add_argument("--use_influence_policy", action='store_true', default=False, help='use a recurrent policy') parser.add_argument("--influence_layer_N", type=int, default=1, help="Number of layers for actor/critic networks") # optimizer parameters parser.add_argument("--lr", type=float, default=5e-4, help='learning rate (default: 5e-4)') parser.add_argument("--tau", type=float, default=0.995, help='soft update polyak (default: 0.995)') parser.add_argument("--critic_lr", type=float, default=5e-4, help='critic learning rate (default: 5e-4)') parser.add_argument("--opti_eps", type=float, default=1e-5, help='RMSprop optimizer epsilon (default: 1e-5)') parser.add_argument("--weight_decay", type=float, default=0) # ppo parameters parser.add_argument("--ppo_epoch", type=int, default=15, help='number of ppo epochs (default: 15)') parser.add_argument("--use_policy_vhead", action='store_true', default=False, help="by default, do not use policy vhead. if set, use policy vhead.") parser.add_argument("--use_clipped_value_loss", action='store_false', default=True, help="by default, clip loss value. If set, do not clip loss value.") parser.add_argument("--clip_param", type=float, default=0.2, help='ppo clip parameter (default: 0.2)') parser.add_argument("--num_mini_batch", type=int, default=1, help='number of batches for ppo (default: 1)') parser.add_argument("--policy_value_loss_coef", type=float, default=1, help='policy value loss coefficient (default: 0.5)') parser.add_argument("--entropy_coef", type=float, default=0.01, help='entropy term coefficient (default: 0.01)') parser.add_argument("--value_loss_coef", type=float, default=1, help='value loss coefficient (default: 0.5)') parser.add_argument("--use_max_grad_norm", action='store_false', default=True, help="by default, use max norm of gradients. If set, do not use.") parser.add_argument("--max_grad_norm", type=float, default=10.0, help='max norm of gradients (default: 0.5)') parser.add_argument("--use_gae", action='store_false', default=True, help='use generalized advantage estimation') parser.add_argument("--gamma", type=float, default=0.99, help='discount factor for rewards (default: 0.99)') parser.add_argument("--gae_lambda", type=float, default=0.95, help='gae lambda parameter (default: 0.95)') parser.add_argument("--use_proper_time_limits", action='store_true', default=False, help='compute returns taking into account time limits') parser.add_argument("--use_huber_loss", action='store_false', default=True, help="by default, use huber loss. If set, do not use huber loss.") parser.add_argument("--use_value_active_masks", action='store_false', default=True, help="by default True, whether to mask useless data in value loss.") parser.add_argument("--use_policy_active_masks", action='store_false', default=True, help="by default True, whether to mask useless data in policy loss.") parser.add_argument("--huber_delta", type=float, default=10.0, help=" coefficience of huber loss.") # save parameters parser.add_argument("--save_interval", type=int, default=1, help="time duration between contiunous twice models saving.") # log parameters parser.add_argument("--log_interval", type=int, default=5, help="time duration between contiunous twice log printing.") # eval parameters parser.add_argument("--use_eval", action='store_true', default=False, help="by default, do not start evaluation. If set`, start evaluation alongside with training.") parser.add_argument("--eval_interval", type=int, default=25, help="time duration between contiunous twice evaluation progress.") parser.add_argument("--eval_episodes", type=int, default=64, help="number of episodes of a single evaluation.") # pretrained parameters parser.add_argument("--model_dir", type=str, default=None, help="by default None. set the path to pretrained model.") parser.add_argument("--replay_save_dir", type=str, default=None, help="replay file save dir") # replay buffer parameters return parser
10,665
55.734043
146
py
TiKick
TiKick-main/tmarl/runners/base_evaluator.py
#!/usr/bin/env python # -*- coding: utf-8 -*- # Copyright 2021 The TARTRL Authors. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """""" import random import numpy as np import torch from tmarl.configs.config import get_config from tmarl.runners.base_runner import Runner def set_seed(seed): random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) class Evaluator(Runner): def __init__(self, argv,program_type=None, client=None): super().__init__(argv) parser = get_config() all_args = self.extra_args_func(argv, parser) all_args.cuda = not all_args.disable_cuda self.algorithm_name = all_args.algorithm_name # cuda if not all_args.disable_cuda and torch.cuda.is_available(): device = torch.device("cuda:0") if all_args.cuda_deterministic: torch.backends.cudnn.benchmark = False torch.backends.cudnn.deterministic = True else: print("choose to use cpu...") device = torch.device("cpu") # run dir run_dir = self.setup_run_dir(all_args) # env init Env_Class, SubprocVecEnv, DummyVecEnv = self.get_env() eval_envs = self.env_init( all_args, Env_Class, SubprocVecEnv, DummyVecEnv) num_agents = all_args.num_agents config = { "all_args": all_args, "envs": None, "eval_envs": eval_envs, "num_agents": num_agents, "device": device, "run_dir": run_dir, } self.all_args, self.envs, self.eval_envs, self.config \ = all_args, None, eval_envs, config self.driver = self.init_driver() def run(self): # run experiments self.driver.run() self.stop() def stop(self): pass def extra_args_func(self, argv, parser): raise NotImplementedError def get_env(self): raise NotImplementedError def init_driver(self): raise NotImplementedError def make_eval_env(self, all_args, Env_Class, SubprocVecEnv, DummyVecEnv): def get_env_fn(rank): def init_env(): env = Env_Class(all_args) env.seed(all_args.seed * 50000 + rank * 10000) return env return init_env if all_args.n_eval_rollout_threads == 1: return DummyVecEnv([get_env_fn(0)]) else: return SubprocVecEnv([get_env_fn(i) for i in range(all_args.n_eval_rollout_threads)]) def env_init(self, all_args, Env_Class, SubprocVecEnv, DummyVecEnv): eval_envs = self.make_eval_env( all_args, Env_Class, SubprocVecEnv, DummyVecEnv) if all_args.use_eval else None return eval_envs def setup_run_dir(self, all_args): return None
3,402
28.08547
97
py
TiKick
TiKick-main/tmarl/runners/base_runner.py
#!/usr/bin/env python # -*- coding: utf-8 -*- # Copyright 2021 The TARTRL Authors. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """""" import os import random import socket import setproctitle import numpy as np from pathlib import Path import torch from tmarl.configs.config import get_config def set_seed(seed): random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) class Runner: def __init__(self, argv): self.argv = argv def run(self): # main run raise NotImplementedError
1,079
22.478261
74
py
TiKick
TiKick-main/tmarl/utils/valuenorm.py
import numpy as np import torch import torch.nn as nn class ValueNorm(nn.Module): """ Normalize a vector of observations - across the first norm_axes dimensions""" def __init__(self, input_shape, norm_axes=1, beta=0.99999, per_element_update=False, epsilon=1e-5, device=torch.device("cpu")): super(ValueNorm, self).__init__() self.input_shape = input_shape self.norm_axes = norm_axes self.epsilon = epsilon self.beta = beta self.per_element_update = per_element_update self.tpdv = dict(dtype=torch.float32, device=device) self.running_mean = nn.Parameter(torch.zeros(input_shape), requires_grad=False).to(**self.tpdv) self.running_mean_sq = nn.Parameter(torch.zeros(input_shape), requires_grad=False).to(**self.tpdv) self.debiasing_term = nn.Parameter(torch.tensor(0.0), requires_grad=False).to(**self.tpdv) self.reset_parameters() def reset_parameters(self): self.running_mean.zero_() self.running_mean_sq.zero_() self.debiasing_term.zero_() def running_mean_var(self): debiased_mean = self.running_mean / self.debiasing_term.clamp(min=self.epsilon) debiased_mean_sq = self.running_mean_sq / self.debiasing_term.clamp(min=self.epsilon) debiased_var = (debiased_mean_sq - debiased_mean ** 2).clamp(min=1e-2) return debiased_mean, debiased_var @torch.no_grad() def update(self, input_vector): if type(input_vector) == np.ndarray: input_vector = torch.from_numpy(input_vector) input_vector = input_vector.to(**self.tpdv) batch_mean = input_vector.mean(dim=tuple(range(self.norm_axes))) batch_sq_mean = (input_vector ** 2).mean(dim=tuple(range(self.norm_axes))) if self.per_element_update: batch_size = np.prod(input_vector.size()[:self.norm_axes]) weight = self.beta ** batch_size else: weight = self.beta self.running_mean.mul_(weight).add_(batch_mean * (1.0 - weight)) self.running_mean_sq.mul_(weight).add_(batch_sq_mean * (1.0 - weight)) self.debiasing_term.mul_(weight).add_(1.0 * (1.0 - weight)) def normalize(self, input_vector): # Make sure input is float32 if type(input_vector) == np.ndarray: input_vector = torch.from_numpy(input_vector) input_vector = input_vector.to(**self.tpdv) mean, var = self.running_mean_var() out = (input_vector - mean[(None,) * self.norm_axes]) / torch.sqrt(var)[(None,) * self.norm_axes] return out def denormalize(self, input_vector): """ Transform normalized data back into original distribution """ if type(input_vector) == np.ndarray: input_vector = torch.from_numpy(input_vector) input_vector = input_vector.to(**self.tpdv) mean, var = self.running_mean_var() out = input_vector * torch.sqrt(var)[(None,) * self.norm_axes] + mean[(None,) * self.norm_axes] out = out.cpu().numpy() return out
3,110
37.8875
131
py
TiKick
TiKick-main/tmarl/utils/util.py
import copy import numpy as np import math import gym import torch import torch.nn as nn import torch.nn.functional as F import torch.distributed as dist from torch.autograd import Variable from gym.spaces import Box, Discrete, Tuple def check(input): if type(input) == np.ndarray: return torch.from_numpy(input) def get_gard_norm(it): sum_grad = 0 for x in it: if x.grad is None: continue sum_grad += x.grad.norm() ** 2 return math.sqrt(sum_grad) def update_linear_schedule(optimizer, epoch, total_num_epochs, initial_lr): """Decreases the learning rate linearly""" lr = initial_lr - (initial_lr * (epoch / float(total_num_epochs))) for param_group in optimizer.param_groups: param_group['lr'] = lr def huber_loss(e, d): a = (abs(e) <= d).float() b = (e > d).float() return a*e**2/2 + b*d*(abs(e)-d/2) def mse_loss(e): return e**2/2 def get_shape_from_obs_space(obs_space): if obs_space.__class__.__name__ == 'Box': obs_shape = obs_space.shape elif obs_space.__class__.__name__ == 'list': obs_shape = obs_space elif obs_space.__class__.__name__ == 'Dict': obs_shape = obs_space.spaces else: raise NotImplementedError return obs_shape def get_shape_from_act_space(act_space): if act_space.__class__.__name__ == 'Discrete': act_shape = 1 elif act_space.__class__.__name__ == "MultiDiscrete": act_shape = act_space.shape elif act_space.__class__.__name__ == "Box": act_shape = act_space.shape[0] elif act_space.__class__.__name__ == "MultiBinary": act_shape = act_space.shape[0] else: # agar act_shape = act_space[0].shape[0] + 1 return act_shape def tile_images(img_nhwc): """ Tile N images into one big PxQ image (P,Q) are chosen to be as close as possible, and if N is square, then P=Q. input: img_nhwc, list or array of images, ndim=4 once turned into array n = batch index, h = height, w = width, c = channel returns: bigim_HWc, ndarray with ndim=3 """ img_nhwc = np.asarray(img_nhwc) N, h, w, c = img_nhwc.shape H = int(np.ceil(np.sqrt(N))) W = int(np.ceil(float(N)/H)) img_nhwc = np.array( list(img_nhwc) + [img_nhwc[0]*0 for _ in range(N, H*W)]) img_HWhwc = img_nhwc.reshape(H, W, h, w, c) img_HhWwc = img_HWhwc.transpose(0, 2, 1, 3, 4) img_Hh_Ww_c = img_HhWwc.reshape(H*h, W*w, c) return img_Hh_Ww_c def to_torch(input): return torch.from_numpy(input) if type(input) == np.ndarray else input def to_numpy(x): return x.detach().cpu().numpy() class FixedCategorical(torch.distributions.Categorical): def sample(self): return super().sample() def log_probs(self, actions): return ( super() .log_prob(actions.squeeze(-1)) .view(actions.size(0), -1) .sum(-1) .unsqueeze(-1) ) def mode(self): return self.probs.argmax(dim=-1, keepdim=True) class MultiDiscrete(gym.Space): """ - The multi-discrete action space consists of a series of discrete action spaces with different parameters - It can be adapted to both a Discrete action space or a continuous (Box) action space - It is useful to represent game controllers or keyboards where each key can be represented as a discrete action space - It is parametrized by passing an array of arrays containing [min, max] for each discrete action space where the discrete action space can take any integers from `min` to `max` (both inclusive) Note: A value of 0 always need to represent the NOOP action. e.g. Nintendo Game Controller - Can be conceptualized as 3 discrete action spaces: 1) Arrow Keys: Discrete 5 - NOOP[0], UP[1], RIGHT[2], DOWN[3], LEFT[4] - params: min: 0, max: 4 2) Button A: Discrete 2 - NOOP[0], Pressed[1] - params: min: 0, max: 1 3) Button B: Discrete 2 - NOOP[0], Pressed[1] - params: min: 0, max: 1 - Can be initialized as MultiDiscrete([ [0,4], [0,1], [0,1] ]) """ def __init__(self, array_of_param_array): self.low = np.array([x[0] for x in array_of_param_array]) self.high = np.array([x[1] for x in array_of_param_array]) self.num_discrete_space = self.low.shape[0] self.n = np.sum(self.high) + 2 def sample(self): """ Returns a array with one sample from each discrete action space """ # For each row: round(random .* (max - min) + min, 0) random_array = np.random.rand(self.num_discrete_space) return [int(x) for x in np.floor(np.multiply((self.high - self.low + 1.), random_array) + self.low)] def contains(self, x): return len(x) == self.num_discrete_space and (np.array(x) >= self.low).all() and (np.array(x) <= self.high).all() @property def shape(self): return self.num_discrete_space def __repr__(self): return "MultiDiscrete" + str(self.num_discrete_space) def __eq__(self, other): return np.array_equal(self.low, other.low) and np.array_equal(self.high, other.high) class DecayThenFlatSchedule(): def __init__(self, start, finish, time_length, decay="exp"): self.start = start self.finish = finish self.time_length = time_length self.delta = (self.start - self.finish) / self.time_length self.decay = decay if self.decay in ["exp"]: self.exp_scaling = (-1) * self.time_length / \ np.log(self.finish) if self.finish > 0 else 1 def eval(self, T): if self.decay in ["linear"]: return max(self.finish, self.start - self.delta * T) elif self.decay in ["exp"]: return min(self.start, max(self.finish, np.exp(- T / self.exp_scaling))) pass def huber_loss(e, d): a = (abs(e) <= d).float() b = (e > d).float() return a*e**2/2 + b*d*(abs(e)-d/2) def mse_loss(e): return e**2 def init(module, weight_init, bias_init, gain=1): weight_init(module.weight.data, gain=gain) bias_init(module.bias.data) return module def get_clones(module, N): return nn.ModuleList([copy.deepcopy(module) for i in range(N)]) # https://github.com/ikostrikov/pytorch-ddpg-naf/blob/master/ddpg.py#L11 def soft_update(target, source, tau): """ Perform DDPG soft update (move target params toward source based on weight factor tau) Inputs: target (torch.nn.Module): Net to copy parameters to source (torch.nn.Module): Net whose parameters to copy tau (float, 0 < x < 1): Weight factor for update """ for target_param, param in zip(target.parameters(), source.parameters()): target_param.data.copy_( target_param.data * (1.0 - tau) + param.data * tau) # https://github.com/ikostrikov/pytorch-ddpg-naf/blob/master/ddpg.py#L15 def hard_update(target, source): """ Copy network parameters from source to target Inputs: target (torch.nn.Module): Net to copy parameters to source (torch.nn.Module): Net whose parameters to copy """ for target_param, param in zip(target.parameters(), source.parameters()): target_param.data.copy_(param.data) # https://github.com/seba-1511/dist_tuto.pth/blob/gh-pages/train_dist.py def average_gradients(model): """ Gradient averaging. """ size = float(dist.get_world_size()) for param in model.parameters(): dist.all_reduce(param.grad.data, op=dist.reduce_op.SUM, group=0) param.grad.data /= size def onehot_from_logits(logits, avail_logits=None, eps=0.0): """ Given batch of logits, return one-hot sample using epsilon greedy strategy (based on given epsilon) """ # get best (according to current policy) actions in one-hot form logits = to_torch(logits) dim = len(logits.shape) - 1 if avail_logits is not None: avail_logits = to_torch(avail_logits) logits[avail_logits == 0] = -1e10 argmax_acs = (logits == logits.max(dim, keepdim=True)[0]).float() if eps == 0.0: return argmax_acs # get random actions in one-hot form rand_acs = Variable(torch.eye(logits.shape[1])[[np.random.choice( range(logits.shape[1]), size=logits.shape[0])]], requires_grad=False) # chooses between best and random actions using epsilon greedy return torch.stack([argmax_acs[i] if r > eps else rand_acs[i] for i, r in enumerate(torch.rand(logits.shape[0]))]) # modified for PyTorch from https://github.com/ericjang/gumbel-softmax/blob/master/Categorical%20VAE.ipynb def sample_gumbel(shape, eps=1e-20, tens_type=torch.FloatTensor): """Sample from Gumbel(0, 1)""" U = Variable(tens_type(*shape).uniform_(), requires_grad=False) return -torch.log(-torch.log(U + eps) + eps) # modified for PyTorch from https://github.com/ericjang/gumbel-softmax/blob/master/Categorical%20VAE.ipynb def gumbel_softmax_sample(logits, avail_logits, temperature, device=torch.device('cpu')): """ Draw a sample from the Gumbel-Softmax distribution""" if str(device) == 'cpu': y = logits + sample_gumbel(logits.shape, tens_type=type(logits.data)) else: y = (logits.cpu() + sample_gumbel(logits.shape, tens_type=type(logits.data))).cuda() dim = len(logits.shape) - 1 if avail_logits is not None: avail_logits = to_torch(avail_logits).to(device) y[avail_logits == 0] = -1e10 return F.softmax(y / temperature, dim=dim) # modified for PyTorch from https://github.com/ericjang/gumbel-softmax/blob/master/Categorical%20VAE.ipynb def gumbel_softmax(logits, avail_logits=None, temperature=1.0, hard=False, device=torch.device('cpu')): """Sample from the Gumbel-Softmax distribution and optionally discretize. Args: logits: [batch_size, n_class] unnormalized log-probs temperature: non-negative scalar hard: if True, take argmax, but differentiate w.r.t. soft sample y Returns: [batch_size, n_class] sample from the Gumbel-Softmax distribution. If hard=True, then the returned sample will be one-hot, otherwise it will be a probabilitiy distribution that sums to 1 across classes """ y = gumbel_softmax_sample(logits, avail_logits, temperature, device) if hard: y_hard = onehot_from_logits(y) y = (y_hard - y).detach() + y return y def gaussian_noise(shape, std): return torch.empty(shape).normal_(mean=0, std=std) def get_obs_shape(obs_space): if obs_space.__class__.__name__ == "Box": obs_shape = obs_space.shape elif obs_space.__class__.__name__ == "list": obs_shape = obs_space else: raise NotImplementedError return obs_shape def get_dim_from_space(space): if isinstance(space, Box): dim = space.shape[0] elif isinstance(space, Discrete): dim = space.n elif isinstance(space, Tuple): dim = sum([get_dim_from_space(sp) for sp in space]) elif "MultiDiscrete" in space.__class__.__name__: return (space.high - space.low) + 1 elif isinstance(space, list): dim = space[0] else: raise Exception("Unrecognized space: ", type(space)) return dim def get_state_dim(observation_dict, action_dict): combined_obs_dim = sum([get_dim_from_space(space) for space in observation_dict.values()]) combined_act_dim = 0 for space in action_dict.values(): dim = get_dim_from_space(space) if isinstance(dim, np.ndarray): combined_act_dim += int(sum(dim)) else: combined_act_dim += dim return combined_obs_dim, combined_act_dim, combined_obs_dim+combined_act_dim def get_cent_act_dim(action_space): cent_act_dim = 0 for space in action_space: dim = get_dim_from_space(space) if isinstance(dim, np.ndarray): cent_act_dim += int(sum(dim)) else: cent_act_dim += dim return cent_act_dim def is_discrete(space): if isinstance(space, Discrete) or "MultiDiscrete" in space.__class__.__name__: return True else: return False def is_multidiscrete(space): if "MultiDiscrete" in space.__class__.__name__: return True else: return False def make_onehot(int_action, action_dim, seq_len=None): if type(int_action) == torch.Tensor: int_action = int_action.cpu().numpy() if not seq_len: return np.eye(action_dim)[int_action] if seq_len: onehot_actions = [] for i in range(seq_len): onehot_action = np.eye(action_dim)[int_action[i]] onehot_actions.append(onehot_action) return np.stack(onehot_actions) def avail_choose(x, avail_x=None): x = to_torch(x) if avail_x is not None: avail_x = to_torch(avail_x) x[avail_x == 0] = -1e10 return x # FixedCategorical(logits=x) def tile_images(img_nhwc): """ Tile N images into one big PxQ image (P,Q) are chosen to be as close as possible, and if N is square, then P=Q. input: img_nhwc, list or array of images, ndim=4 once turned into array n = batch index, h = height, w = width, c = channel returns: bigim_HWc, ndarray with ndim=3 """ img_nhwc = np.asarray(img_nhwc) N, h, w, c = img_nhwc.shape H = int(np.ceil(np.sqrt(N))) W = int(np.ceil(float(N)/H)) img_nhwc = np.array( list(img_nhwc) + [img_nhwc[0]*0 for _ in range(N, H*W)]) img_HWhwc = img_nhwc.reshape(H, W, h, w, c) img_HhWwc = img_HWhwc.transpose(0, 2, 1, 3, 4) img_Hh_Ww_c = img_HhWwc.reshape(H*h, W*w, c) return img_Hh_Ww_c
13,893
31.846336
122
py
TiKick
TiKick-main/tmarl/utils/gpu_mem_track.py
# code from https://github.com/Oldpan/Pytorch-Memory-Utils import gc import datetime import inspect import torch import numpy as np dtype_memory_size_dict = { torch.float64: 64/8, torch.double: 64/8, torch.float32: 32/8, torch.float: 32/8, torch.float16: 16/8, torch.half: 16/8, torch.int64: 64/8, torch.long: 64/8, torch.int32: 32/8, torch.int: 32/8, torch.int16: 16/8, torch.short: 16/6, torch.uint8: 8/8, torch.int8: 8/8, } # compatibility of torch1.0 if getattr(torch, "bfloat16", None) is not None: dtype_memory_size_dict[torch.bfloat16] = 16/8 if getattr(torch, "bool", None) is not None: dtype_memory_size_dict[torch.bool] = 8/8 # pytorch use 1 byte for a bool, see https://github.com/pytorch/pytorch/issues/41571 def get_mem_space(x): try: ret = dtype_memory_size_dict[x] except KeyError: print(f"dtype {x} is not supported!") return ret class MemTracker(object): """ Class used to track pytorch memory usage Arguments: detail(bool, default True): whether the function shows the detail gpu memory usage path(str): where to save log file verbose(bool, default False): whether show the trivial exception device(int): GPU number, default is 0 """ def __init__(self, detail=True, path='', verbose=False, device=0): self.print_detail = detail self.last_tensor_sizes = set() self.gpu_profile_fn = path + f'{datetime.datetime.now():%d-%b-%y-%H:%M:%S}-gpu_mem_track.txt' self.verbose = verbose self.begin = True self.device = device def get_tensors(self): for obj in gc.get_objects(): try: if torch.is_tensor(obj) or (hasattr(obj, 'data') and torch.is_tensor(obj.data)): tensor = obj else: continue if tensor.is_cuda: yield tensor except Exception as e: if self.verbose: print('A trivial exception occured: {}'.format(e)) def get_tensor_usage(self): sizes = [np.prod(np.array(tensor.size())) * get_mem_space(tensor.dtype) for tensor in self.get_tensors()] return np.sum(sizes) / 1024**2 def get_allocate_usage(self): return torch.cuda.memory_allocated() / 1024**2 def clear_cache(self): gc.collect() torch.cuda.empty_cache() def print_all_gpu_tensor(self, file=None): for x in self.get_tensors(): print(x.size(), x.dtype, np.prod(np.array(x.size()))*get_mem_space(x.dtype)/1024**2, file=file) def track(self): """ Track the GPU memory usage """ frameinfo = inspect.stack()[1] where_str = frameinfo.filename + ' line ' + str(frameinfo.lineno) + ': ' + frameinfo.function with open(self.gpu_profile_fn, 'a+') as f: if self.begin: f.write(f"GPU Memory Track | {datetime.datetime.now():%d-%b-%y-%H:%M:%S} |" f" Total Tensor Used Memory:{self.get_tensor_usage():<7.1f}Mb" f" Total Allocated Memory:{self.get_allocate_usage():<7.1f}Mb\n\n") self.begin = False if self.print_detail is True: ts_list = [(tensor.size(), tensor.dtype) for tensor in self.get_tensors()] new_tensor_sizes = {(type(x), tuple(x.size()), ts_list.count((x.size(), x.dtype)), np.prod(np.array(x.size()))*get_mem_space(x.dtype)/1024**2, x.dtype) for x in self.get_tensors()} for t, s, n, m, data_type in new_tensor_sizes - self.last_tensor_sizes: f.write(f'+ | {str(n)} * Size:{str(s):<20} | Memory: {str(m*n)[:6]} M | {str(t):<20} | {data_type}\n') for t, s, n, m, data_type in self.last_tensor_sizes - new_tensor_sizes: f.write(f'- | {str(n)} * Size:{str(s):<20} | Memory: {str(m*n)[:6]} M | {str(t):<20} | {data_type}\n') self.last_tensor_sizes = new_tensor_sizes f.write(f"\nAt {where_str:<50}" f" Total Tensor Used Memory:{self.get_tensor_usage():<7.1f}Mb" f" Total Allocated Memory:{self.get_allocate_usage():<7.1f}Mb\n\n")
4,432
36.888889
129
py
TiKick
TiKick-main/tmarl/utils/modelsize_estimate.py
# code from https://github.com/Oldpan/Pytorch-Memory-Utils import torch.nn as nn import numpy as np def modelsize(model, input, type_size=4): para = sum([np.prod(list(p.size())) for p in model.parameters()]) # print('Model {} : Number of params: {}'.format(model._get_name(), para)) print('Model {} : params: {:4f}M'.format(model._get_name(), para * type_size / 1000 / 1000)) input_ = input.clone() input_.requires_grad_(requires_grad=False) mods = list(model.modules()) out_sizes = [] for i in range(1, len(mods)): m = mods[i] if isinstance(m, nn.ReLU): if m.inplace: continue out = m(input_) out_sizes.append(np.array(out.size())) input_ = out total_nums = 0 for i in range(len(out_sizes)): s = out_sizes[i] nums = np.prod(np.array(s)) total_nums += nums # print('Model {} : Number of intermedite variables without backward: {}'.format(model._get_name(), total_nums)) # print('Model {} : Number of intermedite variables with backward: {}'.format(model._get_name(), total_nums*2)) print('Model {} : intermedite variables: {:3f} M (without backward)' .format(model._get_name(), total_nums * type_size / 1000 / 1000)) print('Model {} : intermedite variables: {:3f} M (with backward)' .format(model._get_name(), total_nums * type_size*2 / 1000 / 1000))
1,428
34.725
116
py
RobDanns
RobDanns-main/deep_learning/tools/corruptions-inference-tinyimagenet.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Train a classification model.""" from __future__ import print_function import argparse import numpy as np import os import sys import torch import multiprocessing as mp import math import pdb import torch.utils.data import torchvision.datasets as datasets import torchvision.transforms as transforms from pycls.config import assert_cfg from pycls.config import cfg from pycls.config import dump_cfg from pycls.datasets import loader from pycls.models import model_builder from pycls.utils.meters import TestMeter from pycls.utils.meters import TrainMeter from PIL import Image import pycls.models.losses as losses import pycls.models.optimizer as optim import pycls.utils.checkpoint as cu import pycls.utils.distributed as du import pycls.utils.logging as lu import pycls.utils.metrics as mu import pycls.utils.multiprocessing as mpu import pycls.utils.net as nu import pycls.datasets.paths as dp import time from datetime import datetime from tensorboardX import SummaryWriter from torchvision.utils import save_image from skimage.util import random_noise print("Let's use GPU :", torch.cuda.current_device()) logger = lu.get_logger(__name__) def parse_args(): """Parses the arguments.""" parser = argparse.ArgumentParser( description='Train a classification model' ) parser.add_argument( '--cfg', dest='cfg_file', help='Config file', required=True, type=str ) parser.add_argument( 'opts', help='See pycls/core/config.py for all options', default=None, nargs=argparse.REMAINDER ) if len(sys.argv) == 1: parser.print_help() sys.exit(1) return parser.parse_args() # TEST(VAL) DATA_LOADER FOR TINY_IMAGENET200 def parseClasses(file): classes = [] filenames = [] with open(file) as f: lines = f.readlines() lines = [x.strip() for x in lines] for x in range(0, len(lines)): tokens = lines[x].split() classes.append(tokens[1]) filenames.append(tokens[0]) return filenames, classes def load_allimages(dir): images = [] if not os.path.isdir(dir): sys.exit(-1) for root, _, fnames in sorted(os.walk(dir)): for fname in sorted(fnames): #if datasets.folder.is_image_file(fname): if datasets.folder.has_file_allowed_extension(fname,['.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif']): path = os.path.join(root, fname) item = path images.append(item) return images class TinyImageNet(torch.utils.data.Dataset): """ TinyImageNet200 validation dataloader.""" def __init__(self, img_path, gt_path, class_to_idx=None, transform=None): self.img_path = img_path self.transform = transform self.gt_path = gt_path self.class_to_idx = class_to_idx self.classidx = [] self.imgs, self.classnames = parseClasses(gt_path) for classname in self.classnames: self.classidx.append(self.class_to_idx[classname]) def __getitem__(self, index): """inputs: Index, retrns: tuple(im, label)""" img = None with open(os.path.join(self.img_path, self.imgs[index]), 'rb') as f: img = Image.open(f) img = img.convert('RGB') if self.transform is not None: img = self.transform(img) label = self.classidx[index] return img, label def __len__(self): return len(self.imgs) def is_eval_epoch(cur_epoch): """Determines if the model should be evaluated at the current epoch.""" return ( (cur_epoch + 1) % cfg.TRAIN.EVAL_PERIOD == 0 or (cur_epoch + 1) == cfg.OPTIM.MAX_EPOCH ) def log_model_info(model, writer_eval=None): """Logs model info""" logger.info('Model:\n{}'.format(model)) params = mu.params_count(model) flops = mu.flops_count(model) logger.info('Params: {:,}'.format(params)) logger.info('Flops: {:,}'.format(flops)) logger.info('Number of node: {:,}'.format(cfg.RGRAPH.GROUP_NUM)) # logger.info('{}, {}'.format(params,flops)) if writer_eval is not None: writer_eval.add_scalar('Params', params, 1) writer_eval.add_scalar('Flops', flops, 1) return params, flops @torch.no_grad() def eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval=None, params=0, flops=0, is_master=False): """Evaluates the model on the test set.""" # Enable eval mode model.eval() test_meter.iter_tic() for cur_iter, (inputs, labels) in enumerate(test_loader): # Transfer the data to the current GPU device inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) # Compute the predictions preds = model(inputs) # Compute the errors top1_err, top5_err = mu.topk_errors(preds, labels, [1, 5]) # Combine the errors across the GPUs if cfg.NUM_GPUS > 1: top1_err, top5_err = du.scaled_all_reduce([top1_err, top5_err]) # Copy the errors from GPU to CPU (sync point) top1_err, top5_err = top1_err.item(), top5_err.item() test_meter.iter_toc() # Update and log stats test_meter.update_stats( top1_err, top5_err, inputs.size(0) * cfg.NUM_GPUS ) test_meter.log_iter_stats(cur_epoch, cur_iter) test_meter.iter_tic() # Log epoch stats test_meter.log_epoch_stats(cur_epoch, writer_eval, params, flops, model, is_master=is_master) eval_stats = test_meter.get_epoch_stats(cur_epoch) test_meter.reset() if cfg.RGRAPH.SAVE_GRAPH: adj_dict = nu.model2adj(model) adj_dict = {**adj_dict, 'top1_err': eval_stats['top1_err']} os.makedirs('{}/graphs/{}'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN), exist_ok=True) np.savez('{}/graphs/{}/{}.npz'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN, cur_epoch), **adj_dict) # return eval_stats def save_noisy_image(img, name): if img.size(2) == 32: img = img.view(img.size(0), 3, 32, 32) save_image(img, name) if img.size(2) == 64: img = img.view(img.size(0), 3, 64, 64) save_image(img, name) else: img = img.view(img.size(0), 3, 224, 224) save_image(img, name) ## Functions to save noisy images. # def gaussian_noise(test_loader): # print("Adding gaussian_noise") # for data in test_loader: # img, _ = data[0], data[1] # gaussian_img_05 = torch.tensor(random_noise(img, mode='gaussian', mean=0, var=0.05, clip=True)) # gaussian_img_2 = torch.tensor(random_noise(img, mode='gaussian', mean=0, var=0.2, clip=True)) # gaussian_img_4 = torch.tensor(random_noise(img, mode='gaussian', mean=0, var=0.4, clip=True)) # gaussian_img_6 = torch.tensor(random_noise(img, mode='gaussian', mean=0, var=0.6, clip=True)) # save_noisy_image(gaussian_img_05, r"noisy-images/gaussian_05.png") # save_noisy_image(gaussian_img_2, r"noisy-images/gaussian_2.png") # save_noisy_image(gaussian_img_4, r"noisy-images/gaussian_4.png") # save_noisy_image(gaussian_img_6, r"noisy-images/gaussian_6.png") # break # def salt_pepper_noise(test_loader): # print("Adding salt_pepper_noise") # for data in test_loader: # img, _ = data[0], data[1] # s_vs_p_5 = torch.tensor(random_noise(img, mode='s&p', salt_vs_pepper=0.5, clip=True)) # s_vs_p_6 = torch.tensor(random_noise(img, mode='s&p', salt_vs_pepper=0.6, clip=True)) # s_vs_p_7 = torch.tensor(random_noise(img, mode='s&p', salt_vs_pepper=0.7, clip=True)) # save_noisy_image(s_vs_p_5, r"noisy-images/s&p_5.png") # break # def speckle_noise(test_loader): # print("Adding speckle_noise") # for data in test_loader: # img, _ = data[0], data[1] # speckle_img_05 = torch.tensor(random_noise(img, mode='speckle', mean=0, var=0.05, clip=True)) # speckle_img_2 = torch.tensor(random_noise(img, mode='speckle', mean=0, var=0.2, clip=True)) # speckle_img_4 = torch.tensor(random_noise(img, mode='speckle', mean=0, var=0.4, clip=True)) # speckle_img_6 = torch.tensor(random_noise(img, mode='speckle', mean=0, var=0.6, clip=True)) # save_noisy_image(speckle_img_05, r"noisy-images/speckle_05.png") # save_noisy_image(speckle_img_2, r"noisy-images/speckle_2.png") # save_noisy_image(speckle_img_4, r"noisy-images/speckle_4.png") # save_noisy_image(speckle_img_6, r"noisy-images/speckle_6.png") # break def train_model(writer_train=None, writer_eval=None, is_master=False): """Trains the model.""" # Fit flops/params if cfg.TRAIN.AUTO_MATCH and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: mode = 'flops' # flops or params if cfg.TRAIN.DATASET == 'cifar10': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': # ResNet20 stats_baseline = 40813184 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 64: stats_baseline = 48957952 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'cifar100': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': # ResNet20 stats_baseline = 40813184 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'tinyimagenet200': pre_repeat = 9 if cfg.MODEL.TYPE == 'resnet': if 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 18: # ResNet18 stats_baseline = 1820000000 elif 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 34: # ResNet34 stats_baseline = 3663761408 elif 'sep' in cfg.RESNET.TRANS_FUN: # ResNet34-sep stats_baseline = 553614592 elif 'bottleneck' in cfg.RESNET.TRANS_FUN: # ResNet50 stats_baseline = 4089184256 elif cfg.MODEL.TYPE == 'efficientnet': # EfficientNet stats_baseline = 385824092 elif cfg.MODEL.TYPE == 'cnn': # CNN if cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 166438912 cfg.defrost() stats = model_builder.build_model_stats(mode) if stats != stats_baseline: # 1st round: set first stage dim for i in range(pre_repeat): scale = round(math.sqrt(stats_baseline / stats), 2) first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first = int(round(first * scale)) cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 step = 1 while True: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first += flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if stats == stats_baseline: break if flag != flag_init: if cfg.RGRAPH.UPPER == False: # make sure the stats is SMALLER than baseline if flag < 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break else: if flag > 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break # 2nd round: set other stage dim first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [int(round(dim / first)) for dim in cfg.RGRAPH.DIM_LIST] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 if 'share' not in cfg.RESNET.TRANS_FUN: for i in range(1, len(cfg.RGRAPH.DIM_LIST)): for j in range(ratio_list[i]): cfg.RGRAPH.DIM_LIST[i] += flag_init stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if flag_init != flag: cfg.RGRAPH.DIM_LIST[i] -= flag_init break stats = model_builder.build_model_stats(mode) print('FINAL', cfg.RGRAPH.GROUP_NUM, cfg.RGRAPH.DIM_LIST, stats, stats_baseline, stats < stats_baseline) # Build the model (before the loaders to ease debugging) model = model_builder.build_model() params, flops = log_model_info(model, writer_eval) # Define the loss function loss_fun = losses.get_loss_fun() # Construct the optimizer optimizer = optim.construct_optimizer(model) # Load a checkpoint if applicable start_epoch = 0 if cu.had_checkpoint(): print("Checking for a checkpoint") last_checkpoint = cu.get_checkpoint_last() print("Last Checkpoint : ", last_checkpoint) checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model, optimizer) logger.info('Loaded checkpoint from: {}'.format(last_checkpoint)) if checkpoint_epoch == cfg.OPTIM.MAX_EPOCH: exit() start_epoch = checkpoint_epoch else: start_epoch = checkpoint_epoch + 1 print("Epoch = ", start_epoch) # Create data loaders data_path = dp.get_data_path(cfg.TRAIN.DATASET) # Retrieve the data path for the dataset traindir = os.path.join(data_path, cfg.TRAIN.SPLIT) valdir = os.path.join(data_path, cfg.TEST.SPLIT, 'images') valgtfile = os.path.join(data_path, cfg.TEST.SPLIT, 'val_annotations.txt') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) # create training dataset and loader train_loader = torch.utils.data.DataLoader( datasets.ImageFolder(traindir, transforms.Compose([ transforms.Resize(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])), batch_size=int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS), shuffle=True, num_workers=cfg.DATA_LOADER.NUM_WORKERS, pin_memory=True) # create validation dataset test_dataset = TinyImageNet( valdir, valgtfile, class_to_idx=train_loader.dataset.class_to_idx.copy(), transform=transforms.Compose([ transforms.Resize(224), transforms.ToTensor(), normalize])) # create validation loader test_loader = torch.utils.data.DataLoader( test_dataset, batch_size=int(cfg.TEST.BATCH_SIZE / cfg.NUM_GPUS), shuffle=False, num_workers=cfg.DATA_LOADER.NUM_WORKERS, pin_memory=cfg.DATA_LOADER.PIN_MEMORY, drop_last=False) # Create meters test_meter = TestMeter(len(test_loader)) if cfg.ONLINE_FLOPS: model_dummy = model_builder.build_model() IMAGE_SIZE = 224 n_flops, n_params = mu.measure_model(model_dummy, IMAGE_SIZE, IMAGE_SIZE) logger.info('FLOPs: %.2fM, Params: %.2fM' % (n_flops / 1e6, n_params / 1e6)) del (model_dummy) # Perform the training loop logger.info('Start epoch: {}'.format(start_epoch + 1)) if start_epoch == cfg.OPTIM.MAX_EPOCH: cur_epoch = start_epoch - 1 eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval, params, flops, is_master=is_master) noise_mode = ['gaussian', 'speckle', 's&p'] noise_std = [0.001, 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6] # change the variance values as desired. model.eval() accuracies_gaussian = [] accuracies_saltpepper = [] accuracies_speckle = [] for mode in noise_mode: for level in noise_std: print("Adding noise={} at level={} to images".format(mode, level)) ctr = 0 correct = 0 total = 0 for cur_iter, (inputs, labels) in enumerate(test_loader): if not 's&p' in mode: noisy_img = torch.tensor(random_noise(inputs, mode=mode, mean=0, var=level, clip=True)) else: noisy_img = torch.tensor(random_noise(inputs, mode=mode, salt_vs_pepper=0.5, clip=True)) noisy_img, labels = noisy_img.cuda(), labels.cuda(non_blocking=True) outputs = model(noisy_img.float()) _, predicted = torch.max(outputs.data, 1) ctr += 1 total += labels.size(0) correct += (predicted == labels).sum() if total > X: # replace X with the number of images to be generated for adversarial attacks. break acc = 100 * float(correct) / total print("acc =", round(acc, 2), "correct =", float(correct), "total =", total) if 'gaussian' in mode: print('Robust Accuracy = {:.3f} with level = {:.2f}'.format(acc, level)) accuracies_gaussian.append(round(acc, 2)) print("Guassian Accuracies after append :", accuracies_gaussian) elif 'speckle' in mode: print('Robust Accuracy = {:.3f} with level = {:.2f}'.format(acc, level)) accuracies_speckle.append(round(acc, 2)) print("Speckle Accuracies after append :", accuracies_speckle) elif 's&p' in mode: print('Robust Accuracy = {:.3f} for S&P noise'.format(acc)) accuracies_saltpepper.append(round(acc, 2)) print("Salt&Pepper Accuracies after append :", accuracies_saltpepper) break else: print("noise mode not supported") # gaussian_noise(test_loader) # salt_pepper_noise(test_loader) # speckle_noise(test_loader) # Change the number of variable as desired number of outputs. gaus_001, gaus_01, gaus_05, gaus_1, gaus_2, gaus_3, gaus_4, gaus_5, gaus_6 = (items for items in accuracies_gaussian) speck_001, speck_01, speck_05, speck_1, speck_2, speck_3, speck_4, speck_5, speck_6 = (items for items in accuracies_speckle) saltpepper = accuracies_saltpepper[0] # load the top1 error and top5 error from the evaluation results f = open("{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH), "r") c_ids = [] for i in f.readlines(): sub_id = list(map(float, i.split(","))) c_ids.append(sub_id[3:5]) topK_errors = [sum(i) / len(c_ids) for i in zip(*c_ids)] top1_error, top5_error = topK_errors[0], topK_errors[1] result_gaussian = ', '.join( [str(cfg.RGRAPH.GROUP_NUM), str(cfg.RGRAPH.P), str(cfg.RGRAPH.SPARSITY), '{:.3f}'.format(top1_error), '{:.3f}'.format(top5_error), str(gaus_001), str(gaus_01), str(gaus_05), str(gaus_1), str(gaus_2), str(gaus_3), str(gaus_4), str(gaus_5), str(gaus_6)]) result_speck = ', '.join( [str(cfg.RGRAPH.GROUP_NUM), str(cfg.RGRAPH.P), str(cfg.RGRAPH.SPARSITY), '{:.3f}'.format(top1_error), '{:.3f}'.format(top5_error), str(speck_001), str(speck_01), str(speck_05), str(speck_1), str(speck_2), str(speck_3), str(speck_4), str(speck_5), str(speck_6)]) result_sp = ', '.join( [str(cfg.RGRAPH.GROUP_NUM), str(cfg.RGRAPH.P), str(cfg.RGRAPH.SPARSITY), '{:.3f}'.format(top1_error), '{:.3f}'.format(top5_error), str(saltpepper)]) with open("{}/gaus_noise_stats.txt".format(cfg.OUT_DIR), "a") as text_file: print(" Writing Text File with accuracies Gaussian:{} ".format(accuracies_gaussian)) text_file.write(result_gaussian + '\n') with open("{}/saltpepper_noise_stats.txt".format(cfg.OUT_DIR), "a") as text_file: print(" Writing Text File with accuracies Salt & Pepper:{} ".format(accuracies_saltpepper)) text_file.write(result_sp + '\n') with open("{}/speckle_noise_stats.txt".format(cfg.OUT_DIR), "a") as text_file: print(" Writing Text File with accuracies Speckle:{} ".format(accuracies_speckle)) text_file.write(result_speck + '\n') def single_proc_train(): """Performs single process training.""" # Setup logging lu.setup_logging() # Show the config logger.info('Config:\n{}'.format(cfg)) # Setup tensorboard if provided writer_train = None writer_eval = None ## If use tensorboard if cfg.TENSORBOARD and du.is_master_proc() and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: comment = '' current_time = datetime.now().strftime('%b%d_%H-%M-%S') logdir_train = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_train') logdir_eval = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_eval') if not os.path.exists(logdir_train): os.makedirs(logdir_train) if not os.path.exists(logdir_eval): os.makedirs(logdir_eval) writer_train = SummaryWriter(logdir_train) writer_eval = SummaryWriter(logdir_eval) # Fix the RNG seeds (see RNG comment in core/config.py for discussion) np.random.seed(cfg.RGRAPH.SEED_TRAIN) torch.manual_seed(cfg.RGRAPH.SEED_TRAIN) # Configure the CUDNN backend torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK # Launch inference + adversarial run train_model(writer_train, writer_eval, is_master=du.is_master_proc()) if writer_train is not None and writer_eval is not None: writer_train.close() writer_eval.close() def check_seed_exists(i): fname = "{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH) if os.path.isfile(fname): with open(fname, 'r') as f: lines = f.readlines() if len(lines) > i: return True return False def main(): # Parse cmd line args args = parse_args() # Load config options cfg.merge_from_file(args.cfg_file) cfg.merge_from_list(args.opts) assert_cfg() # cfg.freeze() # Ensure that the output dir exists os.makedirs(cfg.OUT_DIR, exist_ok=True) # Save the config dump_cfg() for i, cfg.RGRAPH.SEED_TRAIN in enumerate(range(cfg.RGRAPH.SEED_TRAIN_START, cfg.RGRAPH.SEED_TRAIN_END)): # check if a seed has been run if not check_seed_exists(i): print("Launching inference for seed {}".format(i)) single_proc_train() else: print('Inference seed {} already exists, stopping inference'.format(cfg.RGRAPH.SEED_TRAIN)) if __name__ == '__main__': main()
25,928
41.092532
139
py
RobDanns
RobDanns-main/deep_learning/tools/train_resnet18_on_tinyimagenet200.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Train a classification model.""" from __future__ import print_function import argparse import numpy as np import os import sys import torch import multiprocessing as mp import math import pdb import torch.utils.data import torchvision.datasets as datasets import torchvision.transforms as transforms from pycls.config import assert_cfg from pycls.config import cfg from pycls.config import dump_cfg from pycls.datasets import loader from pycls.models import model_builder from pycls.utils.meters import TestMeter from pycls.utils.meters import TrainMeter from PIL import Image import pycls.models.losses as losses import pycls.models.optimizer as optim import pycls.utils.checkpoint as cu import pycls.utils.distributed as du import pycls.utils.logging as lu import pycls.utils.metrics as mu import pycls.utils.multiprocessing as mpu import pycls.utils.net as nu import pycls.datasets.paths as dp import time from datetime import datetime from tensorboardX import SummaryWriter logger = lu.get_logger(__name__) print("Let's use GPU :", torch.cuda.current_device()) def parse_args(): """Parses the arguments.""" parser = argparse.ArgumentParser( description='Train a classification model' ) parser.add_argument( '--cfg', dest='cfg_file', help='Config file', required=True, type=str ) parser.add_argument( 'opts', help='See pycls/core/config.py for all options', default=None, nargs=argparse.REMAINDER ) if len(sys.argv) == 1: parser.print_help() sys.exit(1) return parser.parse_args() # TEST/VAL DATA_LOADER FOR TINY_IMAGENET200 def parseClasses(file): classes = [] filenames = [] with open(file) as f: lines = f.readlines() lines = [x.strip() for x in lines] for x in range(0, len(lines)): tokens = lines[x].split() classes.append(tokens[1]) filenames.append(tokens[0]) return filenames, classes def load_allimages(dir): images = [] if not os.path.isdir(dir): sys.exit(-1) for root, _, fnames in sorted(os.walk(dir)): for fname in sorted(fnames): #if datasets.folder.is_image_file(fname): if datasets.folder.has_file_allowed_extension(fname,['.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif']): path = os.path.join(root, fname) item = path images.append(item) return images class TinyImageNet(torch.utils.data.Dataset): """ TinyImageNet200 validation dataloader.""" def __init__(self, img_path, gt_path, class_to_idx=None, transform=None): self.img_path = img_path self.transform = transform self.gt_path = gt_path self.class_to_idx = class_to_idx self.classidx = [] self.imgs, self.classnames = parseClasses(gt_path) # logger.info('Number of images: {}'.format(len(self.imgs))) # logger.info('Number of classes: {}'.format(len(self.classnames))) for classname in self.classnames: self.classidx.append(self.class_to_idx[classname]) def __getitem__(self, index): """inputs: Index, retrns: tuple(im, label)""" img = None with open(os.path.join(self.img_path, self.imgs[index]), 'rb') as f: img = Image.open(f) img = img.convert('RGB') if self.transform is not None: img = self.transform(img) label = self.classidx[index] return img, label def __len__(self): return len(self.imgs) def is_eval_epoch(cur_epoch): """Determines if the model should be evaluated at the current epoch.""" return ( (cur_epoch + 1) % cfg.TRAIN.EVAL_PERIOD == 0 or (cur_epoch + 1) == cfg.OPTIM.MAX_EPOCH ) def log_model_info(model, writer_eval=None): """Logs model info""" logger.info('Model:\n{}'.format(model)) params = mu.params_count(model) flops = mu.flops_count(model) logger.info('Params: {:,}'.format(params)) logger.info('Flops: {:,}'.format(flops)) logger.info('Number of node: {:,}'.format(cfg.RGRAPH.GROUP_NUM)) # logger.info('{}, {}'.format(params,flops)) if writer_eval is not None: writer_eval.add_scalar('Params', params, 1) writer_eval.add_scalar('Flops', flops, 1) return params, flops def train_epoch( train_loader, model, loss_fun, optimizer, train_meter, cur_epoch, writer_train=None, params=0, flops=0, is_master=False): """Performs one epoch of training.""" # Shuffle the data loader.shuffle(train_loader, cur_epoch) # Update the learning rate lr = optim.get_epoch_lr(cur_epoch) optim.set_lr(optimizer, lr) # Enable training mode model.train() train_meter.iter_tic() for cur_iter, (inputs, labels) in enumerate(train_loader): # Transfer the data to the current GPU device inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) # Perform the forward pass preds = model(inputs) # Compute the loss loss = loss_fun(preds, labels) # Perform the backward pass optimizer.zero_grad() loss.backward() # Update the parameters optimizer.step() # Compute the errors top1_err, top5_err = mu.topk_errors(preds, labels, [1, 5]) # Combine the stats across the GPUs if cfg.NUM_GPUS > 1: loss, top1_err, top5_err = du.scaled_all_reduce( [loss, top1_err, top5_err] ) # Copy the stats from GPU to CPU (sync point) loss, top1_err, top5_err = loss.item(), top1_err.item(), top5_err.item() train_meter.iter_toc() # Update and log stats train_meter.update_stats( top1_err, top5_err, loss, lr, inputs.size(0) * cfg.NUM_GPUS ) train_meter.log_iter_stats(cur_epoch, cur_iter) train_meter.iter_tic() # Log epoch stats train_meter.log_epoch_stats(cur_epoch, writer_train, params, flops, is_master=is_master) trg_stats = train_meter.get_epoch_stats(cur_epoch) train_meter.reset() return trg_stats @torch.no_grad() def eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval=None, params=0, flops=0, is_master=False): """Evaluates the model on the test set.""" # Enable eval mode model.eval() test_meter.iter_tic() for cur_iter, (inputs, labels) in enumerate(test_loader): # Transfer the data to the current GPU device inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) # Compute the predictions preds = model(inputs) # Compute the errors top1_err, top5_err = mu.topk_errors(preds, labels, [1, 5]) # Combine the errors across the GPUs if cfg.NUM_GPUS > 1: top1_err, top5_err = du.scaled_all_reduce([top1_err, top5_err]) # Copy the errors from GPU to CPU (sync point) top1_err, top5_err = top1_err.item(), top5_err.item() test_meter.iter_toc() # Update and log stats test_meter.update_stats( top1_err, top5_err, inputs.size(0) * cfg.NUM_GPUS ) test_meter.log_iter_stats(cur_epoch, cur_iter) test_meter.iter_tic() # Log epoch stats # test_meter.log_epoch_stats(cur_epoch,writer_eval,params,flops) test_meter.log_epoch_stats(cur_epoch, writer_eval, params, flops, model, is_master=is_master) eval_stats = test_meter.get_epoch_stats(cur_epoch) test_meter.reset() if cfg.RGRAPH.SAVE_GRAPH: adj_dict = nu.model2adj(model) adj_dict = {**adj_dict, 'top1_err': eval_stats['top1_err']} os.makedirs('{}/graphs/{}'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN), exist_ok=True) np.savez('{}/graphs/{}/{}.npz'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN, cur_epoch), **adj_dict) return eval_stats def train_model(writer_train=None, writer_eval=None, is_master=False): """Trains the model.""" # Fit flops/params if cfg.TRAIN.AUTO_MATCH and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: mode = 'flops' # flops or params if cfg.TRAIN.DATASET == 'cifar10': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': stats_baseline = 40813184 elif cfg.MODEL.TYPE == 'mlpnet': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 64: stats_baseline = 48957952 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'cifar100': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': stats_baseline = 40813184 elif cfg.MODEL.TYPE == 'mlpnet': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'tinyimagenet200': pre_repeat = 9 if cfg.MODEL.TYPE == 'resnet': if 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 18: # ResNet18 stats_baseline = 1820000000 elif 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 34: # ResNet34 stats_baseline = 3663761408 elif 'sep' in cfg.RESNET.TRANS_FUN: # ResNet34-sep stats_baseline = 553614592 elif 'bottleneck' in cfg.RESNET.TRANS_FUN: # ResNet50 stats_baseline = 4089184256 elif cfg.MODEL.TYPE == 'efficientnet': # EfficientNet stats_baseline = 385824092 elif cfg.MODEL.TYPE == 'cnn': # CNN if cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 166438912 cfg.defrost() stats = model_builder.build_model_stats(mode) if stats != stats_baseline: # 1st round: set first stage dim for i in range(pre_repeat): scale = round(math.sqrt(stats_baseline / stats), 2) first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first = int(round(first * scale)) cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 step = 1 while True: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first += flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if stats == stats_baseline: break if flag != flag_init: if cfg.RGRAPH.UPPER == False: # make sure the stats is SMALLER than baseline if flag < 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break else: if flag > 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break # 2nd round: set other stage dim first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [int(round(dim / first)) for dim in cfg.RGRAPH.DIM_LIST] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 if 'share' not in cfg.RESNET.TRANS_FUN: for i in range(1, len(cfg.RGRAPH.DIM_LIST)): for j in range(ratio_list[i]): cfg.RGRAPH.DIM_LIST[i] += flag_init stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if flag_init != flag: cfg.RGRAPH.DIM_LIST[i] -= flag_init break stats = model_builder.build_model_stats(mode) print('FINAL', cfg.RGRAPH.GROUP_NUM, cfg.RGRAPH.DIM_LIST, stats, stats_baseline, stats < stats_baseline) # Build the model (before the loaders to ease debugging) model = model_builder.build_model() params, flops = log_model_info(model, writer_eval) # Define the loss function loss_fun = losses.get_loss_fun() # Construct the optimizer optimizer = optim.construct_optimizer(model) # Load a checkpoint if applicable start_epoch = 0 if cfg.TRAIN.AUTO_RESUME and cu.has_checkpoint(): last_checkpoint = cu.get_checkpoint_last() checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model, optimizer) logger.info('Loaded checkpoint from: {}'.format(last_checkpoint)) if checkpoint_epoch == cfg.OPTIM.MAX_EPOCH: exit() start_epoch = checkpoint_epoch else: start_epoch = checkpoint_epoch + 1 # Create data loaders # Retrieve the data path for the dataset data_path = dp.get_data_path(cfg.TRAIN.DATASET) traindir = os.path.join(data_path, cfg.TRAIN.SPLIT) valdir = os.path.join(data_path, cfg.TEST.SPLIT, 'images') valgtfile = os.path.join(data_path, cfg.TEST.SPLIT, 'val_annotations.txt') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) # create training dataset and loader train_loader = torch.utils.data.DataLoader( datasets.ImageFolder(traindir, transforms.Compose([ transforms.Resize(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])), batch_size=int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS), shuffle=True, num_workers=cfg.DATA_LOADER.NUM_WORKERS, pin_memory=True) # create validation dataset test_dataset = TinyImageNet( valdir, valgtfile, class_to_idx=train_loader.dataset.class_to_idx.copy(), transform=transforms.Compose([ transforms.Resize(224), transforms.ToTensor(), normalize])) # create validation loader test_loader = torch.utils.data.DataLoader( test_dataset, batch_size=int(cfg.TEST.BATCH_SIZE / cfg.NUM_GPUS), shuffle=False, num_workers=cfg.DATA_LOADER.NUM_WORKERS, pin_memory=cfg.DATA_LOADER.PIN_MEMORY, drop_last=False) # Create meters train_meter = TrainMeter(len(train_loader)) test_meter = TestMeter(len(test_loader)) # Create meters for fgsm test_meter_fgsm = TestMeter(len(test_loader_adv)) if cfg.ONLINE_FLOPS: model_dummy = model_builder.build_model() IMAGE_SIZE = 224 n_flops, n_params = mu.measure_model(model_dummy, IMAGE_SIZE, IMAGE_SIZE) logger.info('FLOPs: %.2fM, Params: %.2fM' % (n_flops / 1e6, n_params / 1e6)) del (model_dummy) # Perform the training loop logger.info('Start epoch: {}'.format(start_epoch + 1)) # do eval at initialization initial_eval_stats = eval_epoch(test_loader, model, test_meter, -1, writer_eval, params, flops, is_master=is_master) if start_epoch == cfg.OPTIM.MAX_EPOCH: cur_epoch = start_epoch - 1 last_epoch_eval_stats = eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval, params, flops, is_master=is_master) else: for cur_epoch in range(start_epoch, cfg.OPTIM.MAX_EPOCH): print('Epoch {} Started'.format(cur_epoch)) # Train for one epoch trg_stats = train_epoch( train_loader, model, loss_fun, optimizer, train_meter, cur_epoch, writer_train, is_master=is_master ) # Compute precise BN stats if cfg.BN.USE_PRECISE_STATS: nu.compute_precise_bn_stats(model, train_loader) # Save a checkpoint if cu.is_checkpoint_epoch(cur_epoch): checkpoint_file = cu.save_checkpoint(model, optimizer, cur_epoch) logger.info('Wrote checkpoint to: {}'.format(checkpoint_file)) # Evaluate the model if is_eval_epoch(cur_epoch): eval_stats = eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval, params, flops, is_master=is_master) def single_proc_train(): """Performs single process training.""" # Setup logging lu.setup_logging() # Show the config logger.info('Config:\n{}'.format(cfg)) # Setup tensorboard if provided writer_train = None writer_eval = None ## If use tensorboard if cfg.TENSORBOARD and du.is_master_proc() and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: comment = '' current_time = datetime.now().strftime('%b%d_%H-%M-%S') logdir_train = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_train') logdir_eval = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_eval') if not os.path.exists(logdir_train): os.makedirs(logdir_train) if not os.path.exists(logdir_eval): os.makedirs(logdir_eval) writer_train = SummaryWriter(logdir_train) writer_eval = SummaryWriter(logdir_eval) # Fix the RNG seeds (see RNG comment in core/config.py for discussion) np.random.seed(cfg.RGRAPH.SEED_TRAIN) torch.manual_seed(cfg.RGRAPH.SEED_TRAIN) # Configure the CUDNN backend torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK # Train the model train_model(writer_train, writer_eval, is_master=du.is_master_proc()) if writer_train is not None and writer_eval is not None: writer_train.close() writer_eval.close() def check_seed_exists(i): fname = "{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH) if os.path.isfile(fname): with open(fname, 'r') as f: lines = f.readlines() if len(lines) > i: return True return False def main(): # Parse cmd line args args = parse_args() # Load config options cfg.merge_from_file(args.cfg_file) cfg.merge_from_list(args.opts) assert_cfg() # cfg.freeze() # Ensure that the output dir exists os.makedirs(cfg.OUT_DIR, exist_ok=True) # Save the config dump_cfg() for i, cfg.RGRAPH.SEED_TRAIN in enumerate(range(cfg.RGRAPH.SEED_TRAIN_START, cfg.RGRAPH.SEED_TRAIN_END)): # check if a seed has been run if not check_seed_exists(i): if cfg.NUM_GPUS > 1: mpu.multi_proc_run(num_proc=cfg.NUM_GPUS, fun=single_proc_train) else: single_proc_train() else: print('Seed {} exists, skip!'.format(cfg.RGRAPH.SEED_TRAIN)) if __name__ == '__main__': main()
21,617
37.741935
129
py
RobDanns
RobDanns-main/deep_learning/tools/adversarial-inference-tinyimagenet200.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Train a classification model.""" from __future__ import print_function import argparse import numpy as np import os import sys import torch import multiprocessing as mp import math import pdb import torch.utils.data import torchvision.datasets as datasets import torchvision.transforms as transforms from pycls.config import assert_cfg from pycls.config import cfg from pycls.config import dump_cfg from pycls.datasets import loader from pycls.models import model_builder from pycls.utils.meters import TestMeter from pycls.utils.meters import TrainMeter from PIL import Image import pycls.models.losses as losses import pycls.models.optimizer as optim import pycls.utils.checkpoint as cu import pycls.utils.distributed as du import pycls.utils.logging as lu import pycls.utils.metrics as mu import pycls.utils.multiprocessing as mpu import pycls.utils.net as nu import pycls.datasets.paths as dp import time from datetime import datetime from tensorboardX import SummaryWriter print("Let's use GPU :", torch.cuda.current_device()) logger = lu.get_logger(__name__) def parse_args(): """Parses the arguments.""" parser = argparse.ArgumentParser( description='Train a classification model' ) parser.add_argument( '--cfg', dest='cfg_file', help='Config file', required=True, type=str ) parser.add_argument( 'opts', help='See pycls/core/config.py for all options', default=None, nargs=argparse.REMAINDER ) if len(sys.argv) == 1: parser.print_help() sys.exit(1) return parser.parse_args() # TEST/VAL DATA_LOADER FOR TINY_IMAGENET200 def parseClasses(file): classes = [] filenames = [] with open(file) as f: lines = f.readlines() lines = [x.strip() for x in lines] for x in range(0, len(lines)): tokens = lines[x].split() classes.append(tokens[1]) filenames.append(tokens[0]) return filenames, classes def load_allimages(dir): images = [] if not os.path.isdir(dir): sys.exit(-1) for root, _, fnames in sorted(os.walk(dir)): for fname in sorted(fnames): # if datasets.folder.is_image_file(fname): if datasets.folder.has_file_allowed_extension(fname,['.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif']): path = os.path.join(root, fname) item = path images.append(item) return images class TinyImageNet(torch.utils.data.Dataset): """ TinyImageNet200 validation dataloader.""" def __init__(self, img_path, gt_path, class_to_idx=None, transform=None): self.img_path = img_path self.transform = transform self.gt_path = gt_path self.class_to_idx = class_to_idx self.classidx = [] self.imgs, self.classnames = parseClasses(gt_path) for classname in self.classnames: self.classidx.append(self.class_to_idx[classname]) def __getitem__(self, index): """inputs: Index, retrns: tuple(im, label)""" img = None with open(os.path.join(self.img_path, self.imgs[index]), 'rb') as f: img = Image.open(f) img = img.convert('RGB') if self.transform is not None: img = self.transform(img) label = self.classidx[index] return img, label def __len__(self): return len(self.imgs) def is_eval_epoch(cur_epoch): """Determines if the model should be evaluated at the current epoch.""" return ( (cur_epoch + 1) % cfg.TRAIN.EVAL_PERIOD == 0 or (cur_epoch + 1) == cfg.OPTIM.MAX_EPOCH ) def log_model_info(model, writer_eval=None): """Logs model info""" logger.info('Model:\n{}'.format(model)) params = mu.params_count(model) flops = mu.flops_count(model) logger.info('Params: {:,}'.format(params)) logger.info('Flops: {:,}'.format(flops)) logger.info('Number of node: {:,}'.format(cfg.RGRAPH.GROUP_NUM)) # logger.info('{}, {}'.format(params,flops)) if writer_eval is not None: writer_eval.add_scalar('Params', params, 1) writer_eval.add_scalar('Flops', flops, 1) return params, flops @torch.no_grad() def eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval=None, params=0, flops=0, is_master=False): """Evaluates the model on the test set.""" # Enable eval mode model.eval() test_meter.iter_tic() for cur_iter, (inputs, labels) in enumerate(test_loader): # Transfer the data to the current GPU device inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) # Compute the predictions preds = model(inputs) # Compute the errors top1_err, top5_err = mu.topk_errors(preds, labels, [1, 5]) # Combine the errors across the GPUs if cfg.NUM_GPUS > 1: top1_err, top5_err = du.scaled_all_reduce([top1_err, top5_err]) # Copy the errors from GPU to CPU (sync point) top1_err, top5_err = top1_err.item(), top5_err.item() test_meter.iter_toc() # Update and log stats test_meter.update_stats( top1_err, top5_err, inputs.size(0) * cfg.NUM_GPUS ) test_meter.log_iter_stats(cur_epoch, cur_iter) test_meter.iter_tic() # Log epoch stats # test_meter.log_epoch_stats(cur_epoch,writer_eval,params,flops) test_meter.log_epoch_stats(cur_epoch, writer_eval, params, flops, model, is_master=is_master) eval_stats = test_meter.get_epoch_stats(cur_epoch) test_meter.reset() if cfg.RGRAPH.SAVE_GRAPH: adj_dict = nu.model2adj(model) adj_dict = {**adj_dict, 'top1_err': eval_stats['top1_err']} os.makedirs('{}/graphs/{}'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN), exist_ok=True) np.savez('{}/graphs/{}/{}.npz'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN, cur_epoch), **adj_dict) # return eval_stats class Normalize(torch.nn.Module): def __init__(self, mean, std): super(Normalize, self).__init__() self.register_buffer('mean', torch.Tensor(mean)) self.register_buffer('std', torch.Tensor(std)) def forward(self, input): # Broadcasting mean = self.mean.reshape(1,3,1,1) std = self.std.reshape(1,3,1,1) norm_img = (input - mean) / std return norm_img # Helper class for printing model layers class PrintLayer(torch.nn.Module): def __init__(self): super(PrintLayer, self).__init__() def forward(self, x): # Do your print / debug stuff here print(x) return x def train_model(writer_train=None, writer_eval=None, is_master=False): """Trains the model.""" # Fit flops/params if cfg.TRAIN.AUTO_MATCH and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: mode = 'flops' # flops or params if cfg.TRAIN.DATASET == 'cifar10': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': # ResNet20 stats_baseline = 40813184 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 64: stats_baseline = 48957952 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'cifar100': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': # ResNet20 stats_baseline = 40813184 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'tinyimagenet200': pre_repeat = 9 if cfg.MODEL.TYPE == 'resnet': if 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 18: # ResNet18 stats_baseline = 1820000000 elif 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 34: # ResNet34 stats_baseline = 3663761408 elif 'sep' in cfg.RESNET.TRANS_FUN: # ResNet34-sep stats_baseline = 553614592 elif 'bottleneck' in cfg.RESNET.TRANS_FUN: # ResNet50 stats_baseline = 4089184256 elif cfg.MODEL.TYPE == 'efficientnet': # EfficientNet stats_baseline = 385824092 elif cfg.MODEL.TYPE == 'cnn': # CNN if cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 166438912 cfg.defrost() stats = model_builder.build_model_stats(mode) if stats != stats_baseline: # 1st round: set first stage dim for i in range(pre_repeat): scale = round(math.sqrt(stats_baseline / stats), 2) first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first = int(round(first * scale)) cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 step = 1 while True: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first += flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if stats == stats_baseline: break if flag != flag_init: if cfg.RGRAPH.UPPER == False: # make sure the stats is SMALLER than baseline if flag < 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break else: if flag > 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break # 2nd round: set other stage dim first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [int(round(dim / first)) for dim in cfg.RGRAPH.DIM_LIST] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 if 'share' not in cfg.RESNET.TRANS_FUN: for i in range(1, len(cfg.RGRAPH.DIM_LIST)): for j in range(ratio_list[i]): cfg.RGRAPH.DIM_LIST[i] += flag_init stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if flag_init != flag: cfg.RGRAPH.DIM_LIST[i] -= flag_init break stats = model_builder.build_model_stats(mode) print('FINAL', cfg.RGRAPH.GROUP_NUM, cfg.RGRAPH.DIM_LIST, stats, stats_baseline, stats < stats_baseline) # Build the model (before the loaders to ease debugging) model = model_builder.build_model() params, flops = log_model_info(model, writer_eval) # for name, param in model.named_parameters(): # print(name, param.shape) # Define the loss function loss_fun = losses.get_loss_fun() # Construct the optimizer optimizer = optim.construct_optimizer(model) # Load a checkpoint if applicable start_epoch = 0 if cu.had_checkpoint(): print("Checking for a checkpoint") last_checkpoint = cu.get_checkpoint_last() print("Last Checkpoint : ", last_checkpoint) checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model, optimizer) logger.info('Loaded checkpoint from: {}'.format(last_checkpoint)) if checkpoint_epoch == cfg.OPTIM.MAX_EPOCH: exit() start_epoch = checkpoint_epoch else: start_epoch = checkpoint_epoch + 1 print("Epoch = ", start_epoch) # Create data loaders data_path = dp.get_data_path(cfg.TRAIN.DATASET) # Retrieve the data path for the dataset traindir = os.path.join(data_path, cfg.TRAIN.SPLIT) valdir = os.path.join(data_path, cfg.TEST.SPLIT, 'images') valgtfile = os.path.join(data_path, cfg.TEST.SPLIT, 'val_annotations.txt') # normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) # create training dataset and loader train_loader = torch.utils.data.DataLoader( datasets.ImageFolder(traindir, transforms.Compose([ transforms.Resize(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])), batch_size=int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS), shuffle=True, num_workers=cfg.DATA_LOADER.NUM_WORKERS, pin_memory=True) # create validation dataset test_dataset = TinyImageNet( valdir, valgtfile, class_to_idx=train_loader.dataset.class_to_idx.copy(), transform=transforms.Compose([ transforms.Resize(224), transforms.ToTensor(), normalize])) # create validation loader test_loader = torch.utils.data.DataLoader( test_dataset, batch_size=int(cfg.TEST.BATCH_SIZE / cfg.NUM_GPUS), shuffle=False, num_workers=cfg.DATA_LOADER.NUM_WORKERS, pin_memory=cfg.DATA_LOADER.PIN_MEMORY, drop_last=False) # create adversarial dataset adv_dataset = TinyImageNet( valdir, valgtfile, class_to_idx=train_loader.dataset.class_to_idx.copy(), transform=transforms.Compose([ transforms.Resize(224), transforms.ToTensor()])) # create adversarial loader test_loader_adv = torch.utils.data.DataLoader( adv_dataset, batch_size=1, shuffle=True, num_workers=cfg.DATA_LOADER.NUM_WORKERS, pin_memory=cfg.DATA_LOADER.PIN_MEMORY, drop_last=False) # Create meters test_meter = TestMeter(len(test_loader)) test_meter_adv = TestMeter(len(test_loader_adv)) if cfg.ONLINE_FLOPS: model_dummy = model_builder.build_model() IMAGE_SIZE = 224 n_flops, n_params = mu.measure_model(model_dummy, IMAGE_SIZE, IMAGE_SIZE) logger.info('FLOPs: %.2fM, Params: %.2fM' % (n_flops / 1e6, n_params / 1e6)) del (model_dummy) # Perform the training loop logger.info('Start epoch: {}'.format(start_epoch + 1)) if start_epoch == cfg.OPTIM.MAX_EPOCH: cur_epoch = start_epoch - 1 eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval, params, flops, is_master=is_master) # when epsilon=0 --> PGD, epsilon=1 --> CW, otherwise FGSM-->replace eps1, eps2, ... with required epsilon of attack versions epsilons = [0, eps1, eps2, ... epsN, 1] # Per-channel mean and SD values in BGR order for TinyImageNet dataset tinyimagenet_MEAN = [0.485, 0.456, 0.406] tinyimagenet_SD = [0.229, 0.224, 0.225] accuracies = [] # add normalization layer to the model norm_layer = Normalize(mean=tinyimagenet_MEAN, std=tinyimagenet_SD) net = torch.nn.Sequential(norm_layer, model).cuda() net = net.eval() for epsilon in epsilons: if epsilon == 0: print("Running PGD Attack") atk = torchattacks.PGD(net, eps=1/510, alpha=2/225, steps=7) # for relevant dataset, use parameters from torchattacks official notebook elif epsilon == 1: print("Running CW Attack") atk = torchattacks.CW(net, c=0.1, kappa=0, steps=100, lr=0.01) # choose suitable values for c, kappa, steps, and lr. else: print("Running FGSM Attacks on epsilon :", epsilon) atk = torchattacks.FGSM(net, eps=epsilon) ctr = 0 correct = 0 total = 0 for cur_iter, (inputs, labels) in enumerate(test_loader_adv): inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) adv_images = atk(inputs, labels) outputs = net(adv_images) _, predicted = torch.max(outputs.data, 1) ctr += 1 total += 1 correct += (predicted == labels).sum() if ctr > X: # replace X with the number of images to be generated for adversarial attacks. print(ctr, " images done for epsilon:", epsilon) break acc = 100 * float(correct) / total print("acc =", round(acc, 2), "correct =", float(correct), "total =", total) accuracies.append(round(acc, 2)) print('Attack Accuracy = {:.3f} with epsilon = {:.4f}'.format(acc, epsilon)) print("accuracies after apend :", accuracies) # save items inside accuracies list to separate float objects, update the # of variables according to requirement. accPGD, accFGSM1, accFGSM2, accFGSM3, accFGSM4, accFGSM5, accFGSM6, accFGSM7, accCW = (items for items in accuracies) # load the top1 error and top5 error from the evaluation results f = open("{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH), "r") c_ids = [] for i in f.readlines(): sub_id = list(map(float, i.split(","))) c_ids.append(sub_id[3:5]) topK_errors = [sum(i) / len(c_ids) for i in zip(*c_ids)] top1_error, top5_error = topK_errors[0], topK_errors[1] result_info = ', '.join( [str(cfg.RGRAPH.GROUP_NUM), str(cfg.RGRAPH.P), str(cfg.RGRAPH.SPARSITY), '{:.3f}'.format(top1_error), '{:.3f}'.format(top5_error), str(accPGD), str(accFGSM1), str(accFGSM2), str(accFGSM3), str(accFGSM4), str(accFGSM5), str(accFGSM6), str(accFGSM7), str(accCW)]) with open("{}/stats.txt".format(cfg.OUT_DIR), "a") as text_file: print(" Writing Text File with accuracies {} ".format(accuracies)) text_file.write(result_info + '\n') def single_proc_train(): """Performs single process training.""" # Setup logging lu.setup_logging() # Show the config logger.info('Config:\n{}'.format(cfg)) # Setup tensorboard if provided writer_train = None writer_eval = None ## If use tensorboard if cfg.TENSORBOARD and du.is_master_proc() and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: comment = '' current_time = datetime.now().strftime('%b%d_%H-%M-%S') logdir_train = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_train') logdir_eval = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_eval') if not os.path.exists(logdir_train): os.makedirs(logdir_train) if not os.path.exists(logdir_eval): os.makedirs(logdir_eval) writer_train = SummaryWriter(logdir_train) writer_eval = SummaryWriter(logdir_eval) # Fix the RNG seeds (see RNG comment in core/config.py for discussion) np.random.seed(cfg.RGRAPH.SEED_TRAIN) torch.manual_seed(cfg.RGRAPH.SEED_TRAIN) # Configure the CUDNN backend torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK # Launch inference + adversarial run train_model(writer_train, writer_eval, is_master=du.is_master_proc()) if writer_train is not None and writer_eval is not None: writer_train.close() writer_eval.close() def check_seed_exists(i): fname = "{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH) if os.path.isfile(fname): with open(fname, 'r') as f: lines = f.readlines() if len(lines) > i: return True return False def main(): # Parse cmd line args args = parse_args() # Load config options cfg.merge_from_file(args.cfg_file) cfg.merge_from_list(args.opts) assert_cfg() # cfg.freeze() # Ensure that the output dir exists os.makedirs(cfg.OUT_DIR, exist_ok=True) # Save the config dump_cfg() for i, cfg.RGRAPH.SEED_TRAIN in enumerate(range(cfg.RGRAPH.SEED_TRAIN_START, cfg.RGRAPH.SEED_TRAIN_END)): # check if a seed has been run if not check_seed_exists(i): print("Launching inference for seed {}".format(i)) single_proc_train() else: print('Inference seed {} already exists, stopping inference'.format(cfg.RGRAPH.SEED_TRAIN)) if __name__ == '__main__': main()
23,184
38.768439
147
py
RobDanns
RobDanns-main/deep_learning/tools/adversarial-inference.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Train a classification model.""" import argparse import pickle import numpy as np import os import sys import torch import math import torchvision import torchattacks from pycls.config import assert_cfg from pycls.config import cfg from pycls.config import dump_cfg from pycls.datasets import loader from pycls.models import model_builder from pycls.utils.meters import TestMeter import pycls.models.losses as losses import pycls.models.optimizer as optim import pycls.utils.checkpoint as cu import pycls.utils.distributed as du import pycls.utils.logging as lu import pycls.utils.metrics as mu import pycls.utils.multiprocessing as mpu import pycls.utils.net as nu import pycls.datasets.transforms as transforms from datetime import datetime from tensorboardX import SummaryWriter import foolbox as fb import art import art.attacks.evasion as evasion from art.estimators.classification import PyTorchClassifier print("Using GPU :", torch.cuda.current_device()) logger = lu.get_logger(__name__) def parse_args(): """Parses the arguments.""" parser = argparse.ArgumentParser( description='Train a classification model' ) parser.add_argument( '--cfg', dest='cfg_file', help='Config file', required=True, type=str ) parser.add_argument( 'opts', help='See pycls/core/config.py for all options', default=None, nargs=argparse.REMAINDER ) if len(sys.argv) == 1: parser.print_help() sys.exit(1) return parser.parse_args() def is_eval_epoch(cur_epoch): """Determines if the model should be evaluated at the current epoch.""" return ( (cur_epoch + 1) % cfg.TRAIN.EVAL_PERIOD == 0 or (cur_epoch + 1) == cfg.OPTIM.MAX_EPOCH ) def log_model_info(model, writer_eval=None): """Logs model info""" logger.info('Model:\n{}'.format(model)) params = mu.params_count(model) flops = mu.flops_count(model) logger.info('Params: {:,}'.format(params)) logger.info('Flops: {:,}'.format(flops)) logger.info('Number of node: {:,}'.format(cfg.RGRAPH.GROUP_NUM)) # logger.info('{}, {}'.format(params,flops)) if writer_eval is not None: writer_eval.add_scalar('Params', params, 1) writer_eval.add_scalar('Flops', flops, 1) return params, flops @torch.no_grad() def eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval=None, params=0, flops=0, is_master=False): """Evaluates the model on the test set.""" # Enable eval mode model.eval() test_meter.iter_tic() # val_input_imgs, for cur_iter, (inputs, labels) in enumerate(test_loader): # Transfer the data to the current GPU device inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) # Compute the predictions preds = model(inputs) # Compute the errors top1_err, top5_err = mu.topk_errors(preds, labels, [1, 5]) # Combine the errors across the GPUs if cfg.NUM_GPUS > 1: top1_err, top5_err = du.scaled_all_reduce([top1_err, top5_err]) # Copy the errors from GPU to CPU (sync point) top1_err, top5_err = top1_err.item(), top5_err.item() test_meter.iter_toc() # Update and log stats test_meter.update_stats( top1_err, top5_err, inputs.size(0) * cfg.NUM_GPUS ) test_meter.log_iter_stats(cur_epoch, cur_iter) test_meter.iter_tic() # Log epoch stats # test_meter.log_epoch_stats(cur_epoch,writer_eval,params,flops) test_meter.log_epoch_stats(cur_epoch, writer_eval, params, flops, model, is_master=is_master) stats = test_meter.get_epoch_stats(cur_epoch) test_meter.reset() if cfg.RGRAPH.SAVE_GRAPH: adj_dict = nu.model2adj(model) adj_dict = {**adj_dict, 'top1_err': stats['top1_err']} os.makedirs('{}/graphs/{}'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN), exist_ok=True) np.savez('{}/graphs/{}/{}.npz'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN, cur_epoch), **adj_dict) class Normalize(torch.nn.Module): def __init__(self, mean, std): super(Normalize, self).__init__() self.register_buffer('mean', torch.Tensor(mean)) self.register_buffer('std', torch.Tensor(std)) def forward(self, input): # Broadcasting mean = self.mean.reshape(1,3,1,1) std = self.std.reshape(1,3,1,1) norm_img = (input - mean) / std return norm_img # Helper class for printing model layers class PrintLayer(torch.nn.Module): def __init__(self): super(PrintLayer, self).__init__() def forward(self, x): # Do your print / debug stuff here print(x) return x def train_model(writer_eval=None, is_master=False): """Trains the model.""" # Fit flops/params if cfg.TRAIN.AUTO_MATCH and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: mode = 'flops' # flops or params if cfg.TRAIN.DATASET == 'cifar10': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': # ResNet20 stats_baseline = 40813184 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 64: stats_baseline = 48957952 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'cifar100': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': if cfg.MODEL.DEPTH == 20: stats_baseline = 40813184 # ResNet20 elif cfg.MODEL.DEPTH == 26: stats_baseline = 56140000 # ResNet26 elif cfg.MODEL.DEPTH == 34: stats_baseline = 71480000 # ResNet34 elif cfg.MODEL.DEPTH == 38: stats_baseline = 86819000 # ResNet38 elif cfg.MODEL.DEPTH == 50: stats_baseline = 130000000 # ResNet50 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'tinyimagenet200': pre_repeat = 9 if cfg.MODEL.TYPE == 'resnet': if 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 18: # ResNet18 stats_baseline = 1820000000 elif 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 34: # ResNet34 stats_baseline = 3663761408 elif 'sep' in cfg.RESNET.TRANS_FUN: # ResNet34-sep stats_baseline = 553614592 elif 'bottleneck' in cfg.RESNET.TRANS_FUN: # ResNet50 stats_baseline = 4089184256 elif cfg.MODEL.TYPE == 'efficientnet': # EfficientNet stats_baseline = 385824092 elif cfg.MODEL.TYPE == 'cnn': # CNN if cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 166438912 elif cfg.TRAIN.DATASET == 'imagenet': pre_repeat = 9 if cfg.MODEL.TYPE == 'resnet': if 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 18: # ResNet18 stats_baseline = 1820000000 elif 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 34: # ResNet34 stats_baseline = 3663761408 elif 'sep' in cfg.RESNET.TRANS_FUN: # ResNet34-sep stats_baseline = 553614592 elif 'bottleneck' in cfg.RESNET.TRANS_FUN: # ResNet50 stats_baseline = 4089184256 elif cfg.MODEL.TYPE == 'efficientnet': # EfficientNet stats_baseline = 385824092 elif cfg.MODEL.TYPE == 'cnn': # CNN if cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 166438912 cfg.defrost() stats = model_builder.build_model_stats(mode) if stats != stats_baseline: # 1st round: set first stage dim for i in range(pre_repeat): scale = round(math.sqrt(stats_baseline / stats), 2) first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first = int(round(first * scale)) cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 step = 1 while True: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first += flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if stats == stats_baseline: break if flag != flag_init: if cfg.RGRAPH.UPPER == False: # make sure the stats is SMALLER than baseline if flag < 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break else: if flag > 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break # 2nd round: set other stage dim first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [int(round(dim / first)) for dim in cfg.RGRAPH.DIM_LIST] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 if 'share' not in cfg.RESNET.TRANS_FUN: for i in range(1, len(cfg.RGRAPH.DIM_LIST)): for j in range(ratio_list[i]): cfg.RGRAPH.DIM_LIST[i] += flag_init stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if flag_init != flag: cfg.RGRAPH.DIM_LIST[i] -= flag_init break stats = model_builder.build_model_stats(mode) print('FINAL', cfg.RGRAPH.GROUP_NUM, cfg.RGRAPH.DIM_LIST, stats, stats_baseline, stats < stats_baseline) # Build the model (before the loaders to ease debugging) model = model_builder.build_model() params, flops = log_model_info(model, writer_eval) if cfg.IS_INFERENCE and cfg.IS_DDP: model = torch.nn.parallel.DataParallel(model) # for name, param in model.named_parameters(): # print(name, param.shape) # Define the loss function loss_fun = losses.get_loss_fun() # Construct the optimizer optimizer = optim.construct_optimizer(model) # Load a checkpoint if applicable start_epoch = 0 if cu.had_checkpoint(): print("Checking for a checkpoint") last_checkpoint = cu.get_checkpoint_last() print("Last Checkpoint : ", last_checkpoint) checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model, optimizer) logger.info('Loaded checkpoint from: {}'.format(last_checkpoint)) if checkpoint_epoch == cfg.OPTIM.MAX_EPOCH: exit() start_epoch = checkpoint_epoch else: start_epoch = checkpoint_epoch + 1 print("Epoch = ", start_epoch) # Create data loaders test_loader = loader.construct_test_loader() test_loader_adv = loader.construct_test_loader_adv() # Create meters test_meter = TestMeter(len(test_loader)) test_meter_adv = TestMeter(len(test_loader_adv)) if cfg.ONLINE_FLOPS: model_dummy = model_builder.build_model() IMAGE_SIZE = 224 n_flops, n_params = mu.measure_model(model_dummy, IMAGE_SIZE, IMAGE_SIZE) logger.info('FLOPs: %.2fM, Params: %.2fM' % (n_flops / 1e6, n_params / 1e6)) del (model_dummy) # Perform the training loop logger.info('Start epoch: {}'.format(start_epoch + 1)) if start_epoch == cfg.OPTIM.MAX_EPOCH: cur_epoch = start_epoch - 1 eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval, params, flops, is_master=is_master) # when epsilon=0, 1 --> PGD, epsilon=2, 3 --> CW, otherwise FGSM-->replace eps1, eps2, ... with required epsilon of attack versions epsilons = [0, 1, eps1, eps2, ... epsN, 2, 3] # Per-channel mean and SD values in BGR order for ImageNet dataset cifar10_MEAN = [0.491, 0.482, 0.4465] cifar10_SD = [0.247, 0.243, 0.262] cifar100_MEAN = [0.507, 0.487, 0.441] cifar100_SD = [0.267, 0.256, 0.276] imagenet_MEAN = [0.406, 0.456, 0.485] imagenet_SD = [0.225, 0.224, 0.229] accuracies = [] # replace the MEAN and SD variable in the following line for the relevant dataset. norm_layer = Normalize(mean=cifar10_MEAN, std=cifar10_SD) net = torch.nn.Sequential(norm_layer, model).cuda() # net = torch.nn.Sequential(norm_layer, PrintLayer(), model).cuda() net = net.eval() print("Adversarial Loader Batch Size =", test_loader_adv.batch_size) for epsilon in epsilons: if epsilon == 0: print("Running PGD Attack") atk_ta = torchattacks.PGD(net, eps=6/255, alpha=2/255, steps=7) # for relevant dataset, use parameters from torchattacks official notebook elif epsilon == 1: print("Running PGD Attack") atk_ta = torchattacks.PGD(net, eps=9/255, alpha=2/255, steps=7) # for relevant dataset, use parameters from torchattacks official notebook elif epsilon == 2: print("Running Torchattacks.CW") atk_ta = torchattacks.CW(net, c=0.15, kappa=0, steps=100, lr=0.01) # replace the values of c and steps according to hyperparameters reported in the paper. elif epsilon == 3: print("Running Torchattacks.CW") atk_ta = torchattacks.CW(net, c=0.25, kappa=0, steps=100, lr=0.01) # replace the values of c and steps according to hyperparameters reported in the paper. # For Foolbox or ART attacks, uncomment the following lines. # print("-> FoolBox.CW") # fmodel = fb.PyTorchModel(net, bounds=(0, 1)) # atk_fb = fb.attacks.L2CarliniWagnerAttack(binary_search_steps=1, initial_const=0.05, # confidence=0, steps=100, stepsize=0.01) # print("-> Adversarial Robustness Toolbox.CW") # classifier = PyTorchClassifier(model=net, clip_values=(0, 1), # loss=loss_fun, # optimizer=optimizer, # input_shape=(3, 32, 32), nb_classes=10) # atk_art = evasion.CarliniL2Method(batch_size=1, classifier=classifier, # binary_search_steps=1, initial_const=0.05, # confidence=0, max_iter=100, # learning_rate=0.01) else: print("Running FGSM Attacks on epsilon :", epsilon) atk_ta = torchattacks.FGSM(net, eps=epsilon) ctr = 0 correct_ta = 0 # correct_fb = 0 # correct_art = 0 total = 0 for cur_iter, (inputs, labels) in enumerate(test_loader_adv): inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) inputs = inputs.float().div(255) adv_images_ta = atk_ta(inputs, labels) # _, adv_images_fb, _ = atk_fb(fmodel, inputs, labels, epsilons=1) # adv_images_art = torch.tensor(atk_art.generate(inputsnp, labelsnp)).cuda() adv_inputs_ta = adv_images_ta.float() # adv_inputs_fb = adv_images_fb.float() # adv_inputs_art = adv_images_art.float() outputs_ta = net(adv_inputs_ta) # outputs_fb = net(adv_inputs_fb) # outputs_art = net(adv_inputs_art) _, predicted_ta = torch.max(outputs_ta.data, 1) # _, predicted_fb = torch.max(outputs_fb.data, 1) # _, predicted_art = torch.max(outputs_art.data, 1) ctr += 1 total += 1 correct_ta += (predicted_ta == labels).sum() # correct_fb += (predicted_fb == labels).sum() # correct_art += (predicted_art == labels).sum() if ctr > X: # replace X with the number of images to be generated for adversarial attacks. print(ctr, " images done for epsilon:", epsilon) break acc_ta = 100 * float(correct_ta) / total # acc_fb = 100 * float(correct_fb) / total # acc_art = 100 * float(correct_art) / total print("ta acc =", round(acc_ta, 2), ", ta correct =", float(correct_ta), ", total =", total) # print("fb acc =", round(acc_fb, 2), ", fb correct =", float(correct_fb), ", total =", total) # print("art acc =", round(acc_art, 2), ", art correct =", float(correct_art), ", total =", total) accuracies.append(round(acc_ta, 2)) print('Attack Accuracy = {:.3f} with epsilon = {:.2f}'.format(acc_ta, epsilon)) print("accuracies after apend :", accuracies) # save items inside accuracies list to separate float objects, update the # of variables according to requirement. accPGD_6by255, accPGD_9by255, accFGSM1, accFGSM2, accFGSM3, accFGSM4, accFGSM5, accCW_15, accCW_25 = (items for items in accuracies) # load the top1 error and top5 error from the evaluation results f = open("{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH), "r") c_ids = [] for i in f.readlines(): sub_id = list(map(float, i.split(","))) c_ids.append(sub_id[3:5]) topK_errors = [sum(i) / len(c_ids) for i in zip(*c_ids)] top1_error, top5_error = topK_errors[0], topK_errors[1] result_info = ', '.join( [str(cfg.RGRAPH.GROUP_NUM), str(cfg.RGRAPH.P), str(cfg.RGRAPH.SPARSITY), '{:.3f}'.format(top1_error), '{:.3f}'.format(top5_error), str(accPGD_6by255), str(accPGD_9by255), str(accFGSM1), str(accFGSM2), str(accFGSM3), str(accFGSM4), str(accFGSM5), str(accCW_15), str(accCW_25)]) # with open("{}/stats.txt".format(cfg.OUT_DIR), "a") as text_file: print(" Writing Text File with accuracies {} ".format(accuracies)) text_file.write(result_info + '\n') def single_proc_train(): """Performs single process training.""" # Setup logging lu.setup_logging() # Show the config logger.info('Config:\n{}'.format(cfg)) # Setup tensorboard if provided writer_train = None writer_eval = None # If use tensorboard if cfg.TENSORBOARD and du.is_master_proc() and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: comment = '' current_time = datetime.now().strftime('%b%d_%H-%M-%S') logdir_train = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_train') logdir_eval = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_eval') if not os.path.exists(logdir_train): os.makedirs(logdir_train) if not os.path.exists(logdir_eval): os.makedirs(logdir_eval) writer_train = SummaryWriter(logdir_train) writer_eval = SummaryWriter(logdir_eval) # Fix the RNG seeds (see RNG comment in core/config.py for discussion) np.random.seed(cfg.RGRAPH.SEED_TRAIN) torch.manual_seed(cfg.RGRAPH.SEED_TRAIN) # Configure the CUDNN backend torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK # Launch inference + adversarial run train_model(writer_eval, is_master=du.is_master_proc()) if writer_eval is not None: # writer_train.close() writer_eval.close() def check_seed_exists(i): fname = "{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH) if os.path.isfile(fname): with open(fname, 'r') as f: lines = f.readlines() if len(lines) > i: return True return False def main(): # Parse cmd line args args = parse_args() # Load config options cfg.merge_from_file(args.cfg_file) cfg.merge_from_list(args.opts) assert_cfg() # cfg.freeze() # Ensure that the output dir exists os.makedirs(cfg.OUT_DIR, exist_ok=True) # Save the config dump_cfg() for i, cfg.RGRAPH.SEED_TRAIN in enumerate(range(cfg.RGRAPH.SEED_TRAIN_START, cfg.RGRAPH.SEED_TRAIN_END)): # check if a seed has been run if not check_seed_exists(i): print("Launching inference for seed {}".format(i)) single_proc_train() else: print('Trained seed {} already exists, stopping inference'.format(cfg.RGRAPH.SEED_TRAIN)) if __name__ == '__main__': main()
23,798
41.72711
166
py
RobDanns
RobDanns-main/deep_learning/tools/corruptions-inference.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Train a classification model.""" import argparse import pickle import numpy as np import os import sys import torch import math import torchvision import torchattacks from pycls.config import assert_cfg from pycls.config import cfg from pycls.config import dump_cfg from pycls.datasets import loader from pycls.models import model_builder from pycls.utils.meters import TestMeter import pycls.models.losses as losses import pycls.models.optimizer as optim import pycls.utils.checkpoint as cu import pycls.utils.distributed as du import pycls.utils.logging as lu import pycls.utils.metrics as mu import pycls.utils.multiprocessing as mpu import pycls.utils.net as nu import pycls.datasets.transforms as transforms from datetime import datetime from tensorboardX import SummaryWriter from torchvision.utils import save_image from skimage.util import random_noise print("Using GPU :", torch.cuda.current_device()) logger = lu.get_logger(__name__) def parse_args(): """Parses the arguments.""" parser = argparse.ArgumentParser( description='Train a classification model' ) parser.add_argument( '--cfg', dest='cfg_file', help='Config file', required=True, type=str ) parser.add_argument( 'opts', help='See pycls/core/config.py for all options', default=None, nargs=argparse.REMAINDER ) if len(sys.argv) == 1: parser.print_help() sys.exit(1) return parser.parse_args() def is_eval_epoch(cur_epoch): """Determines if the model should be evaluated at the current epoch.""" return ( (cur_epoch + 1) % cfg.TRAIN.EVAL_PERIOD == 0 or (cur_epoch + 1) == cfg.OPTIM.MAX_EPOCH ) def log_model_info(model, writer_eval=None): """Logs model info""" logger.info('Model:\n{}'.format(model)) params = mu.params_count(model) flops = mu.flops_count(model) logger.info('Params: {:,}'.format(params)) logger.info('Flops: {:,}'.format(flops)) logger.info('Number of node: {:,}'.format(cfg.RGRAPH.GROUP_NUM)) # logger.info('{}, {}'.format(params,flops)) if writer_eval is not None: writer_eval.add_scalar('Params', params, 1) writer_eval.add_scalar('Flops', flops, 1) return params, flops @torch.no_grad() def eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval=None, params=0, flops=0, is_master=False): """Evaluates the model on the test set.""" # Enable eval mode model.eval() test_meter.iter_tic() # val_input_imgs, for cur_iter, (inputs, labels) in enumerate(test_loader): # Transfer the data to the current GPU device inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) # Compute the predictions preds = model(inputs) # Compute the errors top1_err, top5_err = mu.topk_errors(preds, labels, [1, 5]) # Combine the errors across the GPUs if cfg.NUM_GPUS > 1: top1_err, top5_err = du.scaled_all_reduce([top1_err, top5_err]) # Copy the errors from GPU to CPU (sync point) top1_err, top5_err = top1_err.item(), top5_err.item() test_meter.iter_toc() # Update and log stats test_meter.update_stats( top1_err, top5_err, inputs.size(0) * cfg.NUM_GPUS ) test_meter.log_iter_stats(cur_epoch, cur_iter) test_meter.iter_tic() # Log epoch stats test_meter.log_epoch_stats(cur_epoch, writer_eval, params, flops, model, is_master=is_master) stats = test_meter.get_epoch_stats(cur_epoch) test_meter.reset() if cfg.RGRAPH.SAVE_GRAPH: adj_dict = nu.model2adj(model) adj_dict = {**adj_dict, 'top1_err': stats['top1_err']} os.makedirs('{}/graphs/{}'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN), exist_ok=True) np.savez('{}/graphs/{}/{}.npz'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN, cur_epoch), **adj_dict) def save_noisy_image(img, name): if img.size(2) == 32: img = img.view(img.size(0), 3, 32, 32) save_image(img, name) else: img = img.view(img.size(0), 3, 224, 224) save_image(img, name) ## Functions to save noisy images. # def gaussian_noise(test_loader): # print("Adding gaussian_noise") # for data in test_loader: # img, _ = data[0], data[1] # gaussian_img_05 = torch.tensor(random_noise(img, mode='gaussian', mean=0, var=0.05, clip=True)) # gaussian_img_2 = torch.tensor(random_noise(img, mode='gaussian', mean=0, var=0.2, clip=True)) # gaussian_img_4 = torch.tensor(random_noise(img, mode='gaussian', mean=0, var=0.4, clip=True)) # gaussian_img_6 = torch.tensor(random_noise(img, mode='gaussian', mean=0, var=0.6, clip=True)) # save_noisy_image(gaussian_img_05, r"noisy-images/gaussian_05.png") # save_noisy_image(gaussian_img_2, r"noisy-images/gaussian_2.png") # save_noisy_image(gaussian_img_4, r"noisy-images/gaussian_4.png") # save_noisy_image(gaussian_img_6, r"noisy-images/gaussian_6.png") # break # def salt_pepper_noise(test_loader): # print("Adding salt_pepper_noise") # for data in test_loader: # img, _ = data[0], data[1] # s_vs_p_5 = torch.tensor(random_noise(img, mode='s&p', salt_vs_pepper=0.5, clip=True)) # s_vs_p_6 = torch.tensor(random_noise(img, mode='s&p', salt_vs_pepper=0.6, clip=True)) # s_vs_p_7 = torch.tensor(random_noise(img, mode='s&p', salt_vs_pepper=0.7, clip=True)) # save_noisy_image(s_vs_p_5, r"noisy-images/s&p_5.png") # save_noisy_image(s_vs_p_6, r"noisy-images/s&p_6.png") # save_noisy_image(s_vs_p_7, r"noisy-images/s&p_7.png") # break # def speckle_noise(test_loader): # print("Adding speckle_noise") # for data in test_loader: # img, _ = data[0], data[1] # speckle_img_05 = torch.tensor(random_noise(img, mode='speckle', mean=0, var=0.05, clip=True)) # speckle_img_2 = torch.tensor(random_noise(img, mode='speckle', mean=0, var=0.2, clip=True)) # speckle_img_4 = torch.tensor(random_noise(img, mode='speckle', mean=0, var=0.4, clip=True)) # speckle_img_6 = torch.tensor(random_noise(img, mode='speckle', mean=0, var=0.6, clip=True)) # save_noisy_image(speckle_img_05, r"noisy-images/speckle_05.png") # save_noisy_image(speckle_img_2, r"noisy-images/speckle_2.png") # save_noisy_image(speckle_img_4, r"noisy-images/speckle_4.png") # save_noisy_image(speckle_img_6, r"noisy-images/speckle_6.png") # break def train_model(writer_eval=None, is_master=False): """Trains the model.""" # Fit flops/params if cfg.TRAIN.AUTO_MATCH and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: mode = 'flops' # flops or params if cfg.TRAIN.DATASET == 'cifar10': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': # ResNet20 stats_baseline = 40813184 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 64: stats_baseline = 48957952 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'cifar100': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': # ResNet20 if cfg.MODEL.DEPTH == 20: stats_baseline = 40813184 # ResNet20 elif cfg.MODEL.DEPTH == 26: stats_baseline = 56140000 # ResNet26 elif cfg.MODEL.DEPTH == 34: stats_baseline = 71480000 # ResNet34 elif cfg.MODEL.DEPTH == 38: stats_baseline = 86819000 # ResNet38 elif cfg.MODEL.DEPTH == 50: stats_baseline = 130000000 # ResNet50 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'tinyimagenet200': pre_repeat = 9 if cfg.MODEL.TYPE == 'resnet': if 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 18: # ResNet18 stats_baseline = 1820000000 elif 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 34: # ResNet34 stats_baseline = 3663761408 elif 'sep' in cfg.RESNET.TRANS_FUN: # ResNet34-sep stats_baseline = 553614592 elif 'bottleneck' in cfg.RESNET.TRANS_FUN: # ResNet50 stats_baseline = 4089184256 elif cfg.MODEL.TYPE == 'efficientnet': # EfficientNet stats_baseline = 385824092 elif cfg.MODEL.TYPE == 'cnn': # CNN if cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 166438912 elif cfg.TRAIN.DATASET == 'imagenet': pre_repeat = 9 if cfg.MODEL.TYPE == 'resnet': if 'basic' in cfg.RESNET.TRANS_FUN: # ResNet34 stats_baseline = 3663761408 elif 'sep' in cfg.RESNET.TRANS_FUN: # ResNet34-sep stats_baseline = 553614592 elif 'bottleneck' in cfg.RESNET.TRANS_FUN: # ResNet50 stats_baseline = 4089184256 elif cfg.MODEL.TYPE == 'efficientnet': # EfficientNet stats_baseline = 385824092 elif cfg.MODEL.TYPE == 'cnn': # CNN if cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 166438912 cfg.defrost() stats = model_builder.build_model_stats(mode) if stats != stats_baseline: # 1st round: set first stage dim for i in range(pre_repeat): scale = round(math.sqrt(stats_baseline / stats), 2) first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first = int(round(first * scale)) cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 step = 1 while True: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first += flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if stats == stats_baseline: break if flag != flag_init: if cfg.RGRAPH.UPPER == False: # make sure the stats is SMALLER than baseline if flag < 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break else: if flag > 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break # 2nd round: set other stage dim first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [int(round(dim / first)) for dim in cfg.RGRAPH.DIM_LIST] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 if 'share' not in cfg.RESNET.TRANS_FUN: for i in range(1, len(cfg.RGRAPH.DIM_LIST)): for j in range(ratio_list[i]): cfg.RGRAPH.DIM_LIST[i] += flag_init stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if flag_init != flag: cfg.RGRAPH.DIM_LIST[i] -= flag_init break stats = model_builder.build_model_stats(mode) print('FINAL', cfg.RGRAPH.GROUP_NUM, cfg.RGRAPH.DIM_LIST, stats, stats_baseline, stats < stats_baseline) # Build the model (before the loaders to ease debugging) model = model_builder.build_model() params, flops = log_model_info(model, writer_eval) if cfg.IS_INFERENCE and cfg.IS_DDP: model = torch.nn.parallel.DataParallel(model) # for name, param in model.named_parameters(): # print(name, param.shape) # Define the loss function loss_fun = losses.get_loss_fun() # Construct the optimizer optimizer = optim.construct_optimizer(model) # Load a checkpoint if applicable start_epoch = 0 if cu.had_checkpoint(): print("Checking for a checkpoint") last_checkpoint = cu.get_checkpoint_last() print("Last Checkpoint : ", last_checkpoint) checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model, optimizer) logger.info('Loaded checkpoint from: {}'.format(last_checkpoint)) if checkpoint_epoch == cfg.OPTIM.MAX_EPOCH: exit() start_epoch = checkpoint_epoch else: start_epoch = checkpoint_epoch + 1 print("Epoch = ", start_epoch) # Create data loaders test_loader = loader.construct_test_loader() # Create meters test_meter = TestMeter(len(test_loader)) if cfg.ONLINE_FLOPS: model_dummy = model_builder.build_model() IMAGE_SIZE = 224 n_flops, n_params = mu.measure_model(model_dummy, IMAGE_SIZE, IMAGE_SIZE) logger.info('FLOPs: %.2fM, Params: %.2fM' % (n_flops / 1e6, n_params / 1e6)) del (model_dummy) # Perform the training loop logger.info('Start epoch: {}'.format(start_epoch + 1)) if start_epoch == cfg.OPTIM.MAX_EPOCH: cur_epoch = start_epoch - 1 eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval, params, flops, is_master=is_master) noise_mode = ['gaussian', 'speckle', 's&p'] noise_var = [0.001, 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6] # change the variance values as desired. model.eval() accuracies_gaussian = [] accuracies_saltpepper = [] accuracies_speckle = [] for mode in noise_mode: for level in noise_var: print("Adding noise={} at level={} to images".format(mode, level)) ctr = 0 correct = 0 total = 0 for cur_iter, (inputs, labels) in enumerate(test_loader): if not 's&p' in mode: noisy_img = torch.tensor(random_noise(inputs, mode=mode, mean=0, var=level, clip=True)) else: noisy_img = torch.tensor(random_noise(inputs, mode=mode, salt_vs_pepper=0.5, clip=True)) noisy_img, labels = noisy_img.cuda(), labels.cuda(non_blocking=True) outputs = model(noisy_img.float()) _, predicted = torch.max(outputs.data, 1) ctr += 1 total += labels.size(0) correct += (predicted == labels).sum() if total > X: # replace X with the number of images to be generated for adversarial attacks. break acc = 100 * float(correct) / total print("acc =", round(acc, 2), "correct =", float(correct), "total =", total) if 'gaussian' in mode: print('Robust Accuracy = {:.3f} with level = {:.2f}'.format(acc, level)) accuracies_gaussian.append(round(acc, 2)) print("Guassian Accuracies after append :", accuracies_gaussian) elif 'speckle' in mode: print('Robust Accuracy = {:.3f} with level = {:.2f}'.format(acc, level)) accuracies_speckle.append(round(acc, 2)) print("Speckle Accuracies after append :", accuracies_speckle) elif 's&p' in mode: print('Robust Accuracy = {:.3f} with level = {:.2f}'.format(acc, level)) accuracies_saltpepper.append(round(acc, 2)) print("Salt&Pepper Accuracies after append :", accuracies_saltpepper) break else: print("noise mode not supported") # gaussian_noise(test_loader) # salt_pepper_noise(test_loader) # speckle_noise(test_loader) # Change the number of variable as desired number of outputs. gaus_001, gaus_01, gaus_05, gaus_1, gaus_2, gaus_3, gaus_4, gaus_5, gaus_6 = (items for items in accuracies_gaussian) speck_001, speck_01, speck_05, speck_1, speck_2, speck_3, speck_4, speck_5, speck_6 = (items for items in accuracies_speckle) saltpepper = accuracies_saltpepper[0] # load the top1 error and top5 error from the evaluation results f = open("{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH), "r") c_ids = [] for i in f.readlines(): sub_id = list(map(float, i.split(","))) c_ids.append(sub_id[3:5]) topK_errors = [sum(i) / len(c_ids) for i in zip(*c_ids)] top1_error, top5_error = topK_errors[0], topK_errors[1] result_gaussian = ', '.join( [str(cfg.RGRAPH.GROUP_NUM), str(cfg.RGRAPH.P), str(cfg.RGRAPH.SPARSITY), '{:.3f}'.format(top1_error), '{:.3f}'.format(top5_error), str(gaus_001), str(gaus_01), str(gaus_05), str(gaus_1), str(gaus_2), str(gaus_3), str(gaus_4), str(gaus_5), str(gaus_6)]) result_speck = ', '.join( [str(cfg.RGRAPH.GROUP_NUM), str(cfg.RGRAPH.P), str(cfg.RGRAPH.SPARSITY), '{:.3f}'.format(top1_error), '{:.3f}'.format(top5_error), str(speck_001), str(speck_01), str(speck_05), str(speck_1), str(speck_2), str(speck_3), str(speck_4), str(speck_5), str(speck_6)]) result_sp = ', '.join( [str(cfg.RGRAPH.GROUP_NUM), str(cfg.RGRAPH.P), str(cfg.RGRAPH.SPARSITY), '{:.3f}'.format(top1_error), '{:.3f}'.format(top5_error), str(saltpepper)]) with open("{}/gaus_noise_stats.txt".format(cfg.OUT_DIR), "a") as text_file: print(" Writing Text File with accuracies Gaussian:{} ".format(accuracies_gaussian)) text_file.write(result_gaussian + '\n') with open("{}/saltpepper_noise_stats.txt".format(cfg.OUT_DIR), "a") as text_file: print(" Writing Text File with accuracies Salt & Pepper:{} ".format(accuracies_saltpepper)) text_file.write(result_sp + '\n') with open("{}/speckle_noise_stats.txt".format(cfg.OUT_DIR), "a") as text_file: print(" Writing Text File with accuracies Speckle:{} ".format(accuracies_speckle)) text_file.write(result_speck + '\n') def single_proc_train(): """Performs single process training.""" # Setup logging lu.setup_logging() # Show the config logger.info('Config:\n{}'.format(cfg)) # Setup tensorboard if provided writer_train = None writer_eval = None # If use tensorboard if cfg.TENSORBOARD and du.is_master_proc() and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: comment = '' current_time = datetime.now().strftime('%b%d_%H-%M-%S') logdir_train = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_train') logdir_eval = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_eval') if not os.path.exists(logdir_train): os.makedirs(logdir_train) if not os.path.exists(logdir_eval): os.makedirs(logdir_eval) writer_train = SummaryWriter(logdir_train) writer_eval = SummaryWriter(logdir_eval) # Fix the RNG seeds (see RNG comment in core/config.py for discussion) np.random.seed(cfg.RGRAPH.SEED_TRAIN) torch.manual_seed(cfg.RGRAPH.SEED_TRAIN) # Configure the CUDNN backend torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK # Launch inference + adversarial run train_model(writer_eval, is_master=du.is_master_proc()) if writer_eval is not None: # writer_train.close() writer_eval.close() def check_seed_exists(i): fname = "{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH) if os.path.isfile(fname): with open(fname, 'r') as f: lines = f.readlines() if len(lines) > i: return True return False def main(): # Parse cmd line args args = parse_args() # Load config options cfg.merge_from_file(args.cfg_file) cfg.merge_from_list(args.opts) assert_cfg() # cfg.freeze() # Ensure that the output dir exists os.makedirs(cfg.OUT_DIR, exist_ok=True) # Save the config dump_cfg() for i, cfg.RGRAPH.SEED_TRAIN in enumerate(range(cfg.RGRAPH.SEED_TRAIN_START, cfg.RGRAPH.SEED_TRAIN_END)): # check if a seed has been run if not check_seed_exists(i): print("Launching inference for seed {}".format(i)) single_proc_train() else: print('Inference seed {} already exists, stopping inference'.format(cfg.RGRAPH.SEED_TRAIN)) if __name__ == '__main__': main()
23,864
42.708791
139
py
RobDanns
RobDanns-main/deep_learning/tools/train_net.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Train a classification model.""" import argparse import pickle import numpy as np import os import sys import torch import math # import torchvision # import time from pycls.config import assert_cfg from pycls.config import cfg from pycls.config import dump_cfg from pycls.datasets import loader from pycls.models import model_builder from pycls.utils.meters import TestMeter from pycls.utils.meters import TrainMeter import pycls.models.losses as losses import pycls.models.optimizer as optim import pycls.utils.checkpoint as cu import pycls.utils.distributed as du import pycls.utils.logging as lu import pycls.utils.metrics as mu import pycls.utils.multiprocessing as mpu import pycls.utils.net as nu from datetime import datetime from tensorboardX import SummaryWriter # import wandb logger = lu.get_logger(__name__) print("Let's use GPU :", torch.cuda.current_device()) def parse_args(): """Parses the arguments.""" parser = argparse.ArgumentParser( description='Train a classification model' ) parser.add_argument( '--cfg', dest='cfg_file', help='Config file', required=True, type=str ) parser.add_argument( 'opts', help='See pycls/core/config.py for all options', default=None, nargs=argparse.REMAINDER ) if len(sys.argv) == 1: parser.print_help() sys.exit(1) return parser.parse_args() def is_eval_epoch(cur_epoch): """Determines if the model should be evaluated at the current epoch.""" return ( (cur_epoch + 1) % cfg.TRAIN.EVAL_PERIOD == 0 or (cur_epoch + 1) == cfg.OPTIM.MAX_EPOCH ) def log_model_info(model, writer_eval=None): """Logs model info""" logger.info('Model:\n{}'.format(model)) params = mu.params_count(model) flops = mu.flops_count(model) logger.info('Params: {:,}'.format(params)) logger.info('Flops: {:,}'.format(flops)) logger.info('Number of node: {:,}'.format(cfg.RGRAPH.GROUP_NUM)) # logger.info('{}, {}'.format(params,flops)) if writer_eval is not None: writer_eval.add_scalar('Params', params, 1) writer_eval.add_scalar('Flops', flops, 1) return params, flops def train_epoch( train_loader, model, loss_fun, optimizer, train_meter, cur_epoch, writer_train=None, params=0, flops=0, is_master=False): """Performs one epoch of training.""" # Shuffle the data loader.shuffle(train_loader, cur_epoch) # Update the learning rate lr = optim.get_epoch_lr(cur_epoch) optim.set_lr(optimizer, lr) # Enable training mode model.train() train_meter.iter_tic() for cur_iter, (inputs, labels) in enumerate(train_loader): # Transfer the data to the current GPU device inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) # Perform the forward pass preds = model(inputs) # Compute the loss loss = loss_fun(preds, labels) # Perform the backward pass optimizer.zero_grad() loss.backward() # Update the parameters optimizer.step() # Compute the errors top1_err, top5_err = mu.topk_errors(preds, labels, [1, 5]) # Combine the stats across the GPUs if cfg.NUM_GPUS > 1: loss, top1_err, top5_err = du.scaled_all_reduce( [loss, top1_err, top5_err] ) # Copy the stats from GPU to CPU (sync point) loss, top1_err, top5_err = loss.item(), top1_err.item(), top5_err.item() train_meter.iter_toc() # Update and log stats train_meter.update_stats( top1_err, top5_err, loss, lr, inputs.size(0) * cfg.NUM_GPUS ) train_meter.log_iter_stats(cur_epoch, cur_iter) train_meter.iter_tic() # Log epoch stats train_meter.log_epoch_stats(cur_epoch, writer_train, params, flops, is_master=is_master) trg_stats = train_meter.get_epoch_stats(cur_epoch) train_meter.reset() return trg_stats @torch.no_grad() def eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval=None, params=0, flops=0, is_master=False): """Evaluates the model on the test set.""" # Enable eval mode model.eval() test_meter.iter_tic() for cur_iter, (inputs, labels) in enumerate(test_loader): # Transfer the data to the current GPU device inputs, labels = inputs.cuda(), labels.cuda(non_blocking=True) # Compute the predictions preds = model(inputs) # Compute the errors top1_err, top5_err = mu.topk_errors(preds, labels, [1, 5]) # Combine the errors across the GPUs if cfg.NUM_GPUS > 1: top1_err, top5_err = du.scaled_all_reduce([top1_err, top5_err]) # Copy the errors from GPU to CPU (sync point) top1_err, top5_err = top1_err.item(), top5_err.item() test_meter.iter_toc() # Update and log stats test_meter.update_stats( top1_err, top5_err, inputs.size(0) * cfg.NUM_GPUS ) test_meter.log_iter_stats(cur_epoch, cur_iter) test_meter.iter_tic() # Log epoch stats # test_meter.log_epoch_stats(cur_epoch,writer_eval,params,flops) test_meter.log_epoch_stats(cur_epoch, writer_eval, params, flops, model, is_master=is_master) stats = test_meter.get_epoch_stats(cur_epoch) test_meter.reset() if cfg.RGRAPH.SAVE_GRAPH: adj_dict = nu.model2adj(model) adj_dict = {**adj_dict, 'top1_err': stats['top1_err']} os.makedirs('{}/graphs/{}'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN), exist_ok=True) np.savez('{}/graphs/{}/{}.npz'.format(cfg.OUT_DIR, cfg.RGRAPH.SEED_TRAIN, cur_epoch), **adj_dict) return stats def train_model(writer_train=None, writer_eval=None, is_master=False): """Trains the model.""" # Fit flops/params if cfg.TRAIN.AUTO_MATCH and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: mode = 'flops' # flops or params if cfg.TRAIN.DATASET == 'cifar10': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': stats_baseline = 40813184 # ResNet20 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'cifar100': pre_repeat = 15 if cfg.MODEL.TYPE == 'resnet': if cfg.MODEL.DEPTH == 20: stats_baseline = 40813184 # ResNet20 elif cfg.MODEL.DEPTH == 26: stats_baseline = 56140000 # ResNet26 elif cfg.MODEL.DEPTH == 34: stats_baseline = 71480000 # ResNet34 elif cfg.MODEL.DEPTH == 38: stats_baseline = 86819000 # ResNet38 elif cfg.MODEL.DEPTH == 50: stats_baseline = 130000000 # ResNet50 elif cfg.MODEL.TYPE == 'mlpnet': # 5-layer MLP. cfg.MODEL.LAYERS exclude stem and head layers if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 256: stats_baseline = 985600 elif cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 2364416 elif cfg.RGRAPH.DIM_LIST[0] == 1024: stats_baseline = 6301696 elif cfg.MODEL.TYPE == 'cnn': if cfg.MODEL.LAYERS == 3: if cfg.RGRAPH.DIM_LIST[0] == 512: stats_baseline = 806884352 elif cfg.RGRAPH.DIM_LIST[0] == 16: stats_baseline = 1216672 elif cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 48957952 elif '16d' in cfg.OUT_DIR: stats_baseline = 3392128 elif cfg.TRAIN.DATASET == 'imagenet': pre_repeat = 9 if cfg.MODEL.TYPE == 'resnet': if 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 18: # ResNet18 stats_baseline = 1820000000 elif 'basic' in cfg.RESNET.TRANS_FUN and cfg.MODEL.DEPTH == 34: # ResNet34 stats_baseline = 3663761408 elif 'sep' in cfg.RESNET.TRANS_FUN: # ResNet34-sep stats_baseline = 553614592 elif 'bottleneck' in cfg.RESNET.TRANS_FUN: # ResNet50 stats_baseline = 4089184256 elif cfg.MODEL.TYPE == 'efficientnet': # EfficientNet stats_baseline = 385824092 elif cfg.MODEL.TYPE == 'cnn': # CNN if cfg.MODEL.LAYERS == 6: if '64d' in cfg.OUT_DIR: stats_baseline = 166438912 cfg.defrost() stats = model_builder.build_model_stats(mode) if stats != stats_baseline: # 1st round: set first stage dim for i in range(pre_repeat): scale = round(math.sqrt(stats_baseline / stats), 2) first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first = int(round(first * scale)) cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 step = 1 while True: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first += flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if stats == stats_baseline: break if flag != flag_init: if cfg.RGRAPH.UPPER == False: # make sure the stats is SMALLER than baseline if flag < 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break else: if flag > 0: first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [dim / first for dim in cfg.RGRAPH.DIM_LIST] first -= flag_init * step cfg.RGRAPH.DIM_LIST = [int(round(first * ratio)) for ratio in ratio_list] break # 2nd round: set other stage dim first = cfg.RGRAPH.DIM_LIST[0] ratio_list = [int(round(dim / first)) for dim in cfg.RGRAPH.DIM_LIST] stats = model_builder.build_model_stats(mode) flag_init = 1 if stats < stats_baseline else -1 if 'share' not in cfg.RESNET.TRANS_FUN: for i in range(1, len(cfg.RGRAPH.DIM_LIST)): for j in range(ratio_list[i]): cfg.RGRAPH.DIM_LIST[i] += flag_init stats = model_builder.build_model_stats(mode) flag = 1 if stats < stats_baseline else -1 if flag_init != flag: cfg.RGRAPH.DIM_LIST[i] -= flag_init break stats = model_builder.build_model_stats(mode) print('FINAL', cfg.RGRAPH.GROUP_NUM, cfg.RGRAPH.DIM_LIST, stats, stats_baseline, stats < stats_baseline) # Build the model (before the loaders to ease debugging) model = model_builder.build_model() params, flops = log_model_info(model, writer_eval) # Define the loss function loss_fun = losses.get_loss_fun() # Construct the optimizer optimizer = optim.construct_optimizer(model) # wandb.watch(model) # Load a checkpoint if applicable start_epoch = 0 if cfg.TRAIN.AUTO_RESUME and cu.has_checkpoint(): last_checkpoint = cu.get_checkpoint_last1() checkpoint_epoch = cu.load_checkpoint(last_checkpoint, model, optimizer) logger.info('Loaded checkpoint from: {}'.format(last_checkpoint)) if checkpoint_epoch == cfg.OPTIM.MAX_EPOCH: exit() start_epoch = checkpoint_epoch else: start_epoch = checkpoint_epoch + 1 # Create data loaders train_loader = loader.construct_train_loader() test_loader = loader.construct_test_loader() # Create meters train_meter = TrainMeter(len(train_loader)) test_meter = TestMeter(len(test_loader)) if cfg.ONLINE_FLOPS: model_dummy = model_builder.build_model() IMAGE_SIZE = 224 n_flops, n_params = mu.measure_model(model_dummy, IMAGE_SIZE, IMAGE_SIZE) logger.info('FLOPs: %.2fM, Params: %.2fM' % (n_flops / 1e6, n_params / 1e6)) del (model_dummy) # Perform the training loop logger.info('Start epoch: {}'.format(start_epoch + 1)) # do eval at initialization initial_eval_stats = eval_epoch(test_loader, model, test_meter, -1, writer_eval, params, flops, is_master=is_master) if start_epoch == cfg.OPTIM.MAX_EPOCH: cur_epoch = start_epoch - 1 last_epoch_eval_stats = eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval, params, flops, is_master=is_master) else: for cur_epoch in range(start_epoch, cfg.OPTIM.MAX_EPOCH): print('Epoch {} Started'.format(cur_epoch)) # Train for one epoch trg_stats = train_epoch( train_loader, model, loss_fun, optimizer, train_meter, cur_epoch, writer_train, is_master=is_master ) # Compute precise BN stats if cfg.BN.USE_PRECISE_STATS: nu.compute_precise_bn_stats(model, train_loader) # Save a checkpoint if cu.is_checkpoint_epoch(cur_epoch): checkpoint_file = cu.save_checkpoint(model, optimizer, cur_epoch) logger.info('Wrote checkpoint to: {}'.format(checkpoint_file)) # Evaluate the model if is_eval_epoch(cur_epoch): eval_stats = eval_epoch(test_loader, model, test_meter, cur_epoch, writer_eval, params, flops, is_master=is_master) # wandb.log({'Epoch': cur_epoch, 'Train top1_err': trg_stats['top1_err'], 'Test top1_err': eval_stats['top1_err']}) def single_proc_train(): """Performs single process training.""" # Setup logging lu.setup_logging() # Show the config logger.info('Config:\n{}'.format(cfg)) # Setup tensorboard if provided writer_train = None writer_eval = None ## If use tensorboard if cfg.TENSORBOARD and du.is_master_proc() and cfg.RGRAPH.SEED_TRAIN == cfg.RGRAPH.SEED_TRAIN_START: comment = '' current_time = datetime.now().strftime('%b%d_%H-%M-%S') logdir_train = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_train') logdir_eval = os.path.join(cfg.OUT_DIR, 'runs', current_time + comment + '_eval') if not os.path.exists(logdir_train): os.makedirs(logdir_train) if not os.path.exists(logdir_eval): os.makedirs(logdir_eval) writer_train = SummaryWriter(logdir_train) writer_eval = SummaryWriter(logdir_eval) # Fix the RNG seeds (see RNG comment in core/config.py for discussion) np.random.seed(cfg.RGRAPH.SEED_TRAIN) torch.manual_seed(cfg.RGRAPH.SEED_TRAIN) # Configure the CUDNN backend torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK # Train the model train_model(writer_train, writer_eval, is_master=du.is_master_proc()) if writer_train is not None and writer_eval is not None: writer_train.close() writer_eval.close() def check_seed_exists(i): fname = "{}/results_epoch{}.txt".format(cfg.OUT_DIR, cfg.OPTIM.MAX_EPOCH) if os.path.isfile(fname): with open(fname, 'r') as f: lines = f.readlines() if len(lines) > i: return True return False def main(): # wandb.init(project = 'Rob_G2NN', entity='rowanai-graph-robustness') # Parse cmd line args args = parse_args() # wandb.config.update(args) # Load config options cfg.merge_from_file(args.cfg_file) cfg.merge_from_list(args.opts) assert_cfg() # cfg.freeze() # Ensure that the output dir exists os.makedirs(cfg.OUT_DIR, exist_ok=True) # Save the config dump_cfg() for i, cfg.RGRAPH.SEED_TRAIN in enumerate(range(cfg.RGRAPH.SEED_TRAIN_START, cfg.RGRAPH.SEED_TRAIN_END)): # check if a seed has been run if not check_seed_exists(i): if cfg.NUM_GPUS > 1: mpu.multi_proc_run(num_proc=cfg.NUM_GPUS, fun=single_proc_train) else: single_proc_train() else: print('Seed {} exists, skip!'.format(cfg.RGRAPH.SEED_TRAIN)) if __name__ == '__main__': main()
18,692
39.113734
127
py
RobDanns
RobDanns-main/deep_learning/pycls/config.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Configuration file.""" import os from yacs.config import CfgNode as CN # Global config object _C = CN() # Example usage: # from core.config import cfg cfg = _C # ---------------------------------------------------------------------------- # # Model options # ---------------------------------------------------------------------------- # _C.MODEL = CN() # Model type to use _C.MODEL.TYPE = '' # Number of weight layers _C.MODEL.DEPTH = 0 # Number of classes _C.MODEL.NUM_CLASSES = 10 # Loss function (see pycls/models/loss.py for options) _C.MODEL.LOSS_FUN = 'cross_entropy' # Num layers, excluding the stem and head layers. Total layers used should +2 _C.MODEL.LAYERS = 3 # ---------------------------------------------------------------------------- # # ResNet options # ---------------------------------------------------------------------------- # _C.RESNET = CN() # Transformation function (see pycls/models/resnet.py for options) _C.RESNET.TRANS_FUN = 'basic_transform' # Number of groups to use (1 -> ResNet; > 1 -> ResNeXt) _C.RESNET.NUM_GROUPS = 1 # Width of each group (64 -> ResNet; 4 -> ResNeXt) _C.RESNET.WIDTH_PER_GROUP = 64 # Apply stride to 1x1 conv (True -> MSRA; False -> fb.torch) _C.RESNET.STRIDE_1X1 = False # Whether append 1x1 resblock _C.RESNET.APPEND1x1 = 0 # For group conv only _C.RESNET.GROUP_SIZE = 2 # ---------------------------------------------------------------------------- # # EfficientNet options # ---------------------------------------------------------------------------- # _C.EFFICIENT_NET = CN() # Stem width _C.EFFICIENT_NET.STEM_W = 32 # Depth for each stage (number of blocks in the stage) _C.EFFICIENT_NET.DEPTHS = [] # Width for each stage (width of each block in the stage) _C.EFFICIENT_NET.WIDTHS = [] # Expansion ratios for MBConv blocks in each stage _C.EFFICIENT_NET.EXP_RATIOS = [] # Squeeze-and-Excitation (SE) operation _C.EFFICIENT_NET.SE_ENABLED = True # Squeeze-and-Excitation (SE) ratio _C.EFFICIENT_NET.SE_RATIO = 0.25 # Linear projection _C.EFFICIENT_NET.LIN_PROJ = True # Strides for each stage (applies to the first block of each stage) _C.EFFICIENT_NET.STRIDES = [] # Kernel sizes for each stage _C.EFFICIENT_NET.KERNELS = [] # Head type ('conv_head' or 'simple_head') _C.EFFICIENT_NET.HEAD_TYPE = 'conv_head' # Head width (applies to 'conv_head') _C.EFFICIENT_NET.HEAD_W = 1280 # Ativation function _C.EFFICIENT_NET.ACT_FUN = 'swish' # Drop connect ratio _C.EFFICIENT_NET.DC_RATIO = 0.0 # Drop connect implementation _C.EFFICIENT_NET.DC_IMP = 'tf' # Dropout ratio _C.EFFICIENT_NET.DROPOUT_RATIO = 0.0 # ---------------------------------------------------------------------------- # # Relational graph options # ---------------------------------------------------------------------------- # _C.RGRAPH = CN() # dim for first layer. NOTE: this is fixed when matching FLOPs _C.RGRAPH.DIM_FIRST = 16 # dim for each stage _C.RGRAPH.DIM_LIST = [] # wide stem module _C.RGRAPH.STEM_MODE = 'default' # How to message exchange: dense, hier (deprecated) _C.RGRAPH.TALK_MODE = 'dense' # Num of nodes _C.RGRAPH.GROUP_NUM = 32 # Size of nodes in Stage 1 _C.RGRAPH.GROUP_SIZE = 1 # The type of message passing used _C.RGRAPH.MESSAGE_TYPE = 'ws' # Whether use directed graph _C.RGRAPH.DIRECTED = False # Graph sparsity _C.RGRAPH.SPARSITY = 0.5 # Graph Randomness _C.RGRAPH.P = 0.0 # Graph seed _C.RGRAPH.SEED_GRAPH = 1 # training seed used _C.RGRAPH.SEED_TRAIN = 1 # training seed, start, end _C.RGRAPH.SEED_TRAIN_START = 1 _C.RGRAPH.SEED_TRAIN_END = 2 # Keep graph across the network _C.RGRAPH.KEEP_GRAPH = True # Append additaion 1x1 layers for additional talks _C.RGRAPH.ADD_1x1 = 0 # Match upper computational bound _C.RGRAPH.UPPER = True # Auto match computational budget _C.RGRAPH.AUTO_MATCH = True # AGG func. Only sum is supported in current mask-based implementation _C.RGRAPH.AGG_FUNC = 'sum' # Save weight matrices as graphs. Warning: the saved matrices can be huge _C.RGRAPH.SAVE_GRAPH = False # ---------------------------------------------------------------------------- # # Batch norm options # ---------------------------------------------------------------------------- # _C.BN = CN() # BN epsilon _C.BN.EPS = 1e-5 # BN momentum (BN momentum in PyTorch = 1 - BN momentum in Caffe2) _C.BN.MOM = 0.1 # Precise BN stats _C.BN.USE_PRECISE_STATS = True _C.BN.NUM_SAMPLES_PRECISE = 1024 # Initialize the gamma of the final BN of each block to zero _C.BN.ZERO_INIT_FINAL_GAMMA = False # ---------------------------------------------------------------------------- # # Optimizer options # ---------------------------------------------------------------------------- # _C.OPTIM = CN() # Base learning rate _C.OPTIM.BASE_LR = 0.1 # Learning rate policy select from {'cos', 'exp', 'steps'} _C.OPTIM.LR_POLICY = 'cos' # Exponential decay factor _C.OPTIM.GAMMA = 0.1 # Step size for 'exp' and 'cos' policies (in epochs) _C.OPTIM.STEP_SIZE = 1 # Steps for 'steps' policy (in epochs) _C.OPTIM.STEPS = [] # Learning rate multiplier for 'steps' policy _C.OPTIM.LR_MULT = 0.1 # Maximal number of epochs _C.OPTIM.MAX_EPOCH = 200 # Momentum _C.OPTIM.MOMENTUM = 0.9 # Momentum dampening _C.OPTIM.DAMPENING = 0.0 # Nesterov momentum _C.OPTIM.NESTEROV = True # L2 regularization _C.OPTIM.WEIGHT_DECAY = 5e-4 # Start the warm up from OPTIM.BASE_LR * OPTIM.WARMUP_FACTOR _C.OPTIM.WARMUP_FACTOR = 0.1 # Gradually warm up the OPTIM.BASE_LR over this number of epochs _C.OPTIM.WARMUP_EPOCHS = 0 # ---------------------------------------------------------------------------- # # Training options # ---------------------------------------------------------------------------- # _C.TRAIN = CN() # Dataset and split _C.TRAIN.DATASET = '' _C.TRAIN.SPLIT = 'train' # Total mini-batch size _C.TRAIN.BATCH_SIZE = 128 # Evaluate model on test data every eval period epochs _C.TRAIN.EVAL_PERIOD = 1 # Save model checkpoint every checkpoint period epochs _C.TRAIN.CHECKPOINT_PERIOD = 50 # Resume training from the latest checkpoint in the output directory _C.TRAIN.AUTO_RESUME = True # Checkpoint to start training from (if no automatic checkpoint saved) _C.TRAIN.START_CHECKPOINT = '' _C.TRAIN.AUTO_MATCH = False # ---------------------------------------------------------------------------- # # Testing options # ---------------------------------------------------------------------------- # _C.TEST = CN() # Dataset and split _C.TEST.DATASET = '' _C.TEST.SPLIT = 'val' # Total mini-batch size _C.TEST.BATCH_SIZE = 200 # ---------------------------------------------------------------------------- # # Common train/test data loader options # ---------------------------------------------------------------------------- # _C.DATA_LOADER = CN() # Number of data loader workers per training process _C.DATA_LOADER.NUM_WORKERS = 4 # Load data to pinned host memory _C.DATA_LOADER.PIN_MEMORY = True # ---------------------------------------------------------------------------- # # Memory options # ---------------------------------------------------------------------------- # _C.MEM = CN() # Perform ReLU inplace _C.MEM.RELU_INPLACE = True # ---------------------------------------------------------------------------- # # CUDNN options # ---------------------------------------------------------------------------- # _C.CUDNN = CN() # Perform benchmarking to select the fastest CUDNN algorithms to use # Note that this may increase the memory usage and will likely not result # in overall speedups when variable size inputs are used (e.g. COCO training) _C.CUDNN.BENCHMARK = False # ---------------------------------------------------------------------------- # # Misc options # ---------------------------------------------------------------------------- # # Number of GPUs to use (applies to both training and testing) _C.NUM_GPUS = 1 # Output directory _C.OUT_DIR = '/tmp' # Checkpoint directory for inference _C.CHECKPT_DIR = '/tmp' _C.IS_INFERENCE = False _C.IS_DDP = False # Config destination (in OUT_DIR) _C.CFG_DEST = 'config.yaml' # Note that non-determinism may still be present due to non-deterministic # operator implementations in GPU operator libraries _C.RNG_SEED = 1 # Log destination ('stdout' or 'file') _C.LOG_DEST = 'file' # Log period in iters _C.LOG_PERIOD = 10 # Distributed backend _C.DIST_BACKEND = 'nccl' # Hostname and port for initializing multi-process groups _C.HOST = 'localhost' _C.PORT = 12002 # Computing flops by online foward pass _C.ONLINE_FLOPS = False # Whether use Tensorboard _C.TENSORBOARD = False def assert_cfg(): """Checks config values invariants.""" assert not _C.OPTIM.STEPS or _C.OPTIM.STEPS[0] == 0, \ 'The first lr step must start at 0' assert _C.TRAIN.SPLIT in ['train', 'val', 'test'], \ 'Train split \'{}\' not supported'.format(_C.TRAIN.SPLIT) assert _C.TRAIN.BATCH_SIZE % _C.NUM_GPUS == 0, \ 'Train mini-batch size should be a multiple of NUM_GPUS.' assert _C.TEST.SPLIT in ['train', 'val', 'test'], \ 'Test split \'{}\' not supported'.format(_C.TEST.SPLIT) assert _C.TEST.BATCH_SIZE % _C.NUM_GPUS == 0, \ 'Test mini-batch size should be a multiple of NUM_GPUS.' # assert not _C.BN.USE_PRECISE_STATS or _C.NUM_GPUS == 1, \ # 'Precise BN stats computation not verified for > 1 GPU' assert _C.LOG_DEST in ['stdout', 'file'], \ 'Log destination \'{}\' not supported'.format(_C.LOG_DEST) def dump_cfg(): """Dumps the config to the output directory.""" cfg_file = os.path.join(_C.OUT_DIR, _C.CFG_DEST) with open(cfg_file, 'w') as f: _C.dump(stream=f) def load_cfg(out_dir, cfg_dest='config.yaml'): """Loads config from specified output directory.""" cfg_file = os.path.join(out_dir, cfg_dest) _C.merge_from_file(cfg_file)
10,201
24.378109
107
py
RobDanns
RobDanns-main/deep_learning/pycls/models/losses.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Loss functions.""" import torch.nn as nn from pycls.config import cfg # Supported losses _LOSS_FUNS = { 'cross_entropy': nn.CrossEntropyLoss, } def get_loss_fun(): """Retrieves the loss function.""" assert cfg.MODEL.LOSS_FUN in _LOSS_FUNS.keys(), \ 'Loss function \'{}\' not supported'.format(cfg.TRAIN.LOSS) return _LOSS_FUNS[cfg.MODEL.LOSS_FUN]().cuda()
730
26.074074
107
py
RobDanns
RobDanns-main/deep_learning/pycls/models/efficientnet.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """EfficientNet models.""" import math import torch import torch.nn as nn from pycls.config import cfg import pycls.utils.net as nu import pycls.utils.logging as logging from .relation_graph import * logger = logging.get_logger(__name__) def get_conv(name): """Retrieves the transformation function by name.""" trans_funs = { 'mbconv_transform': MBConv, 'mbtalkconv_transform': MBTalkConv, } assert name in trans_funs.keys(), \ 'Transformation function \'{}\' not supported'.format(name) return trans_funs[name] def drop_connect_tf(x, drop_ratio): """Drop connect (tensorflow port).""" keep_ratio = 1.0 - drop_ratio rt = torch.rand([x.shape[0], 1, 1, 1], dtype=x.dtype, device=x.device) rt.add_(keep_ratio) bt = torch.floor(rt) x.div_(keep_ratio) x.mul_(bt) return x def drop_connect_pt(x, drop_ratio): """Drop connect (pytorch version).""" keep_ratio = 1.0 - drop_ratio mask = torch.empty([x.shape[0], 1, 1, 1], dtype=x.dtype, device=x.device) mask.bernoulli_(keep_ratio) x.div_(keep_ratio) x.mul_(mask) return x def get_act_fun(act_type): """Retrieves the activations function.""" act_funs = { 'swish': Swish, 'relu': nn.ReLU, } assert act_type in act_funs.keys(), \ 'Activation function \'{}\' not supported'.format(act_type) return act_funs[act_type] class SimpleHead(nn.Module): """Simple head.""" def __init__(self, dim_in, num_classes): super(SimpleHead, self).__init__() # AvgPool self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) # Dropout if cfg.EFFICIENT_NET.DROPOUT_RATIO > 0.0: self.dropout = nn.Dropout(p=cfg.EFFICIENT_NET.DROPOUT_RATIO) # FC self.fc = nn.Linear(dim_in, num_classes, bias=True) def forward(self, x): x = self.avg_pool(x) x = x.view(x.size(0), -1) x = self.dropout(x) if hasattr(self, 'dropout') else x x = self.fc(x) return x class ConvHead(nn.Module): """EfficientNet conv head.""" def __init__(self, in_w, out_w, num_classes, act_fun): super(ConvHead, self).__init__() self._construct_class(in_w, out_w, num_classes, act_fun) def _construct_class(self, in_w, out_w, num_classes, act_fun): # 1x1, BN, Swish self.conv = nn.Conv2d( in_w, out_w, kernel_size=1, stride=1, padding=0, bias=False ) self.conv_bn = nn.BatchNorm2d( out_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.conv_swish = act_fun() # AvgPool self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) # Dropout if cfg.EFFICIENT_NET.DROPOUT_RATIO > 0.0: self.dropout = nn.Dropout(p=cfg.EFFICIENT_NET.DROPOUT_RATIO) # FC self.fc = nn.Linear(out_w, num_classes, bias=True) def forward(self, x): # 1x1, BN, Swish x = self.conv_swish(self.conv_bn(self.conv(x))) # AvgPool x = self.avg_pool(x) x = x.view(x.size(0), -1) # Dropout x = self.dropout(x) if hasattr(self, 'dropout') else x # FC x = self.fc(x) return x class LinearHead(nn.Module): """EfficientNet linear head.""" def __init__(self, in_w, out_w, num_classes, act_fun): super(LinearHead, self).__init__() self._construct_class(in_w, out_w, num_classes, act_fun) def _construct_class(self, in_w, out_w, num_classes, act_fun): # AvgPool self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) # FC0 self.fc0 = nn.Linear(in_w, out_w, bias=False) self.fc0_bn = nn.BatchNorm1d( out_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.fc0_swish = act_fun() # FC self.fc = nn.Linear(out_w, num_classes, bias=True) def forward(self, x): # AvgPool x = self.avg_pool(x) x = x.view(x.size(0), -1) # Linear, BN, Swish x = self.fc0_swish(self.fc0_bn(self.fc0(x))) # FC x = self.fc(x) return x class MBConv(nn.Module): """Mobile inverted bottleneck block with SE (MBConv).""" def __init__(self, in_w, exp_r, kernel, stride, se_r, out_w, act_fun, seed=None, exp_w=None): super(MBConv, self).__init__() self._construct_class(in_w, exp_r, kernel, stride, se_r, out_w, act_fun) def _construct_class(self, in_w, exp_r, kernel, stride, se_r, out_w, act_fun): # Expansion: 1x1, BN, Swish self.expand = None exp_w = int(in_w * exp_r) # Include exp ops only if the exp ratio is different from 1 if exp_w != in_w: self.expand = nn.Conv2d( in_w, exp_w, kernel_size=1, stride=1, padding=0, bias=False ) self.expand_bn = nn.BatchNorm2d( exp_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.expand_swish = act_fun() # Depthwise: 3x3 dwise, BN, Swish self.dwise = nn.Conv2d( exp_w, exp_w, kernel_size=kernel, stride=stride, groups=exp_w, bias=False, # Hacky padding to preserve res (supports only 3x3 and 5x5) padding=(1 if kernel == 3 else 2) ) self.dwise_bn = nn.BatchNorm2d( exp_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.dwise_swish = act_fun() # SE: x * F_ex(x) if cfg.EFFICIENT_NET.SE_ENABLED: se_w = int(in_w * se_r) self.se = SE(exp_w, se_w, act_fun) # Linear projection: 1x1, BN self.lin_proj = nn.Conv2d( exp_w, out_w, kernel_size=1, stride=1, padding=0, bias=False ) self.lin_proj_bn = nn.BatchNorm2d( out_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) # Nonlinear projection if not cfg.EFFICIENT_NET.LIN_PROJ: self.lin_proj_swish = act_fun() # Skip connections on blocks w/ same in and out shapes (MN-V2, Fig. 4) self.has_skip = (stride == 1) and (in_w == out_w) def forward(self, x): f_x = x # Expansion if self.expand: f_x = self.expand_swish(self.expand_bn(self.expand(f_x))) # Depthwise f_x = self.dwise_swish(self.dwise_bn(self.dwise(f_x))) # SE if cfg.EFFICIENT_NET.SE_ENABLED: f_x = self.se(f_x) # Linear projection f_x = self.lin_proj_bn(self.lin_proj(f_x)) # Nonlinear projection if not cfg.EFFICIENT_NET.LIN_PROJ: f_x = self.lin_proj_swish(f_x) # Skip connection if self.has_skip: # Drop connect if self.training and cfg.EFFICIENT_NET.DC_RATIO > 0.0: if cfg.EFFICIENT_NET.DC_IMP == 'tf': f_x = drop_connect_tf(f_x, cfg.EFFICIENT_NET.DC_RATIO) else: f_x = drop_connect_pt(f_x, cfg.EFFICIENT_NET.DC_RATIO) f_x = x + f_x return f_x class MBTalkConv(nn.Module): """Mobile inverted bottleneck block with SE (MBConv).""" def __init__(self, in_w, exp_r, kernel, stride, se_r, out_w, act_fun, seed=None, exp_w=None): super(MBTalkConv, self).__init__() self.seed=seed self._construct_class(in_w, exp_r, kernel, stride, se_r, out_w, act_fun, exp_w) def _construct_class(self, in_w, exp_r, kernel, stride, se_r, out_w, act_fun, exp_w): # Expansion: 1x1, BN, Swish self.expand = None if int(exp_r)==1: exp_w = in_w else: self.expand = TalkConv2d( in_w, exp_w, cfg.RGRAPH.GROUP_NUM, kernel_size=1, stride=1, padding=0, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.expand_bn = nn.BatchNorm2d( exp_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.expand_swish = act_fun() # Depthwise: 3x3 dwise, BN, Swish self.dwise = nn.Conv2d( exp_w, exp_w, kernel_size=kernel, stride=stride, groups=exp_w, bias=False, # Hacky padding to preserve res (supports only 3x3 and 5x5) padding=(1 if kernel == 3 else 2) ) self.dwise_bn = nn.BatchNorm2d( exp_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.dwise_swish = act_fun() # SE: x * F_ex(x) if cfg.EFFICIENT_NET.SE_ENABLED: se_w = int(in_w * se_r) self.se = SE(exp_w, se_w, act_fun) # Linear projection: 1x1, BN self.lin_proj = TalkConv2d( exp_w, out_w, cfg.RGRAPH.GROUP_NUM, kernel_size=1, stride=1, padding=0, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.lin_proj_bn = nn.BatchNorm2d( out_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) # Nonlinear projection if not cfg.EFFICIENT_NET.LIN_PROJ: self.lin_proj_swish = act_fun() # Skip connections on blocks w/ same in and out shapes (MN-V2, Fig. 4) self.has_skip = (stride == 1) and (in_w == out_w) def forward(self, x): f_x = x # Expansion if self.expand: f_x = self.expand_swish(self.expand_bn(self.expand(f_x))) # Depthwise f_x = self.dwise_swish(self.dwise_bn(self.dwise(f_x))) # SE if cfg.EFFICIENT_NET.SE_ENABLED: f_x = self.se(f_x) # Linear projection f_x = self.lin_proj_bn(self.lin_proj(f_x)) # Nonlinear projection if not cfg.EFFICIENT_NET.LIN_PROJ: f_x = self.lin_proj_swish(f_x) # Skip connection if self.has_skip: # Drop connect if self.training and cfg.EFFICIENT_NET.DC_RATIO > 0.0: if cfg.EFFICIENT_NET.DC_IMP == 'tf': f_x = drop_connect_tf(f_x, cfg.EFFICIENT_NET.DC_RATIO) else: f_x = drop_connect_pt(f_x, cfg.EFFICIENT_NET.DC_RATIO) f_x = x + f_x return f_x class Stage(nn.Module): """EfficientNet stage.""" def __init__(self, in_w, exp_r, kernel, stride, se_r, out_w, d, act_fun, exp_w=None): super(Stage, self).__init__() self._construct_class(in_w, exp_r, kernel, stride, se_r, out_w, d, act_fun, exp_w) def _construct_class(self, in_w, exp_r, kernel, stride, se_r, out_w, d, act_fun, exp_w): if cfg.RGRAPH.KEEP_GRAPH: seed = cfg.RGRAPH.SEED_GRAPH else: seed = int(cfg.RGRAPH.SEED_GRAPH*100) # Construct a sequence of blocks for i in range(d): trans_fun = get_conv(cfg.RESNET.TRANS_FUN) # Stride and input width apply to the first block of the stage stride_b = stride if i == 0 else 1 in_w_b = in_w if i == 0 else out_w # Construct the block self.add_module( 'b{}'.format(i + 1), trans_fun(in_w_b, exp_r, kernel, stride_b, se_r, out_w, act_fun, seed=seed, exp_w=exp_w) ) if not cfg.RGRAPH.KEEP_GRAPH: seed += 1 def forward(self, x): for block in self.children(): x = block(x) return x class StemIN(nn.Module): """EfficientNet stem for ImageNet.""" def __init__(self, in_w, out_w, act_fun): super(StemIN, self).__init__() self._construct_class(in_w, out_w, act_fun) def _construct_class(self, in_w, out_w, act_fun): self.conv = nn.Conv2d( in_w, out_w, kernel_size=3, stride=2, padding=1, bias=False ) self.bn = nn.BatchNorm2d( out_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.swish = act_fun() def forward(self, x): for layer in self.children(): x = layer(x) return x class EfficientNet(nn.Module): """EfficientNet model.""" def __init__(self): assert cfg.TRAIN.DATASET in ['imagenet'], \ 'Training on {} is not supported'.format(cfg.TRAIN.DATASET) assert cfg.TEST.DATASET in ['imagenet'], \ 'Testing on {} is not supported'.format(cfg.TEST.DATASET) assert cfg.TRAIN.DATASET == cfg.TEST.DATASET, \ 'Train and test dataset must be the same for now' assert cfg.EFFICIENT_NET.HEAD_TYPE in ['conv_head', 'simple_head', 'linear_head'], \ 'Unsupported head type: {}'.format(cfg.EFFICIENT_NET.HEAD_TYPE) super(EfficientNet, self).__init__() self._construct_class( stem_w=cfg.EFFICIENT_NET.STEM_W, ds=cfg.EFFICIENT_NET.DEPTHS, ws=cfg.EFFICIENT_NET.WIDTHS, exp_rs=cfg.EFFICIENT_NET.EXP_RATIOS, se_r=cfg.EFFICIENT_NET.SE_RATIO, ss=cfg.EFFICIENT_NET.STRIDES, ks=cfg.EFFICIENT_NET.KERNELS, head_type=cfg.EFFICIENT_NET.HEAD_TYPE, head_w=cfg.EFFICIENT_NET.HEAD_W, act_type=cfg.EFFICIENT_NET.ACT_FUN, nc=cfg.MODEL.NUM_CLASSES ) self.apply(nu.init_weights) def _construct_class( self, stem_w, ds, ws, exp_rs, se_r, ss, ks, head_type, head_w, act_type, nc ): """Constructs imagenet models.""" # Group params by stage stage_params = list(zip(ds, ws, exp_rs, ss, ks)) # Activation function act_fun = get_act_fun(act_type) # Set dim for each stage dim_list = cfg.RGRAPH.DIM_LIST expdim_list = [int(cfg.EFFICIENT_NET.WIDTHS[i]*cfg.EFFICIENT_NET.EXP_RATIOS[i]) for i in range(len(cfg.EFFICIENT_NET.WIDTHS))] # Construct the stems self.stem = StemIN(3, stem_w, act_fun) prev_w = stem_w # Construct the stages for i, (d, w, exp_r, stride, kernel) in enumerate(stage_params): if cfg.RESNET.TRANS_FUN != 'mbconv_transform': w = dim_list[i] exp_w = expdim_list[i] self.add_module( 's{}'.format(i + 1), Stage(prev_w, exp_r, kernel, stride, se_r, w, d, act_fun, exp_w=exp_w) ) prev_w = w # Construct the head if head_type == 'conv_head': self.head = ConvHead(prev_w, head_w, nc, act_fun) elif head_type == 'linear_head': self.head = LinearHead(prev_w, head_w, nc, act_fun) else: self.head = SimpleHead(prev_w, nc) def forward(self, x): for module in self.children(): x = module(x) return x
15,385
33.809955
108
py
RobDanns
RobDanns-main/deep_learning/pycls/models/resnet.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """ResNet or ResNeXt model.""" import torch.nn as nn import torch from pycls.config import cfg import pycls.utils.logging as lu import pycls.utils.net as nu from .relation_graph import * import time import pdb logger = lu.get_logger(__name__) # Stage depths for an ImageNet model {model depth -> (d2, d3, d4, d5)} _IN_MODEL_STAGE_DS = { 18: (2, 2, 2, 2), 34: (3, 4, 6, 3), 50: (3, 4, 6, 3), 101: (3, 4, 23, 3), 152: (3, 8, 36, 3), } def get_trans_fun(name): """Retrieves the transformation function by name.""" trans_funs = { ############ Res-34 'channelbasic_transform': ChannelBasicTransform, 'groupbasictalk_transform': GroupBasicTalkTransform, ############ Res-34-sep 'channelsep_transform': ChannelSepTransform, 'groupseptalk_transform': GroupSepTalkTransform, ############ Res-50 'bottleneck_transform': BottleneckTransform, 'talkbottleneck_transform': TalkBottleneckTransform, } assert name in trans_funs.keys(), \ 'Transformation function \'{}\' not supported'.format(name) return trans_funs[name] ############ Res-34 class ChannelBasicTransform(nn.Module): """Basic transformation: 3x3, 3x3""" def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): super(ChannelBasicTransform, self).__init__() self._construct_class(dim_in, dim_out, stride) def _construct_class(self, dim_in, dim_out, stride): # 3x3, BN, ReLU self.a = nn.Conv2d( dim_in, dim_out, kernel_size=3, stride=stride, padding=1, bias=False ) self.a_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # 3x3, BN self.b = nn.Conv2d( dim_out, dim_out, kernel_size=3, stride=1, padding=1, bias=False ) self.b_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.b_bn.final_bn = True def forward(self, x): for layer in self.children(): x = layer(x) return x class GroupBasicTalkTransform(nn.Module): """Basic transformation: 3x3, 3x3, relational graph""" def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): self.seed = seed super(GroupBasicTalkTransform, self).__init__() self._construct_class(dim_in, dim_out, stride) def _construct_class(self, dim_in, dim_out, stride): # 3x3, BN, ReLU self.a = TalkConv2d( dim_in, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=3, stride=stride, padding=1, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.a_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # 3x3, BN self.b = TalkConv2d( dim_out, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=3, stride=1, padding=1, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.b_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.b_bn.final_bn = True def forward(self, x): for layer in self.children(): x = layer(x) return x ############ Res-34-sep class ChannelSepTransform(nn.Module): """Separable transformation: 3x3, 3x3""" def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): super(ChannelSepTransform, self).__init__() self._construct_class(dim_in, dim_out, stride) def _construct_class(self, dim_in, dim_out, stride): # ReLU, 3x3, BN, 1x1, BN self.a_3x3 = nn.Conv2d( dim_in, dim_in, kernel_size=3, stride=stride, padding=1, bias=False, groups=dim_in ) self.a_1x1 = nn.Conv2d( dim_in, dim_out, kernel_size=1, stride=1, padding=0, bias=False ) self.a_1x1_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # ReLU, 3x3, BN, 1x1, BN self.b_3x3 = nn.Conv2d( dim_out, dim_out, kernel_size=3, stride=1, padding=1, bias=False, groups=dim_out ) self.b_1x1 = nn.Conv2d( dim_out, dim_out, kernel_size=1, stride=1, padding=0, bias=False ) self.b_1x1_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.b_1x1_bn.final_bn = True def forward(self, x): for layer in self.children(): x = layer(x) return x class GroupSepTalkTransform(nn.Module): """Separable transformation: 3x3, 3x3, relational graph""" def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): self.seed = seed super(GroupSepTalkTransform, self).__init__() self._construct_class(dim_in, dim_out, stride) def _construct_class(self, dim_in, dim_out, stride): # ReLU, 3x3, BN, 1x1, BN self.a_3x3 = nn.Conv2d( dim_in, dim_in, kernel_size=3, stride=stride, padding=1, bias=False, groups=dim_in ) self.a_1x1 = TalkConv2d( dim_in, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=1, stride=1, padding=0, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.a_1x1_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # ReLU, 3x3, BN, 1x1, BN self.b_3x3 = nn.Conv2d( dim_out, dim_out, kernel_size=3, stride=1, padding=1, bias=False, groups=dim_out ) self.b_1x1 = TalkConv2d( dim_out, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=1, stride=1, padding=0, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.b_1x1_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.b_1x1_bn.final_bn = True def forward(self, x): for layer in self.children(): x = layer(x) return x ############ Res-50 class BottleneckTransform(nn.Module): """Bottleneck transformation: 1x1, 3x3, 1x1""" def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): super(BottleneckTransform, self).__init__() dim_inner = int(round(dim_out / 4)) self._construct_class(dim_in, dim_out, stride, dim_inner, num_gs, seed) def _construct_class(self, dim_in, dim_out, stride, dim_inner, num_gs, seed): # MSRA -> stride=2 is on 1x1; TH/C2 -> stride=2 is on 3x3 # (str1x1, str3x3) = (stride, 1) if cfg.RESNET.STRIDE_1X1 else (1, stride) (str1x1, str3x3) = (1, stride) # 1x1, BN, ReLU self.a = nn.Conv2d( dim_in, dim_inner, kernel_size=1, stride=str1x1, padding=0, bias=False ) self.a_bn = nn.BatchNorm2d( dim_inner, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # 3x3, BN, ReLU self.b = nn.Conv2d( dim_inner, dim_inner, kernel_size=3, stride=str3x3, padding=1, groups=num_gs, bias=False ) self.b_bn = nn.BatchNorm2d( dim_inner, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.b_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # 1x1, BN self.c = nn.Conv2d( dim_inner, dim_out, kernel_size=1, stride=1, padding=0, bias=False ) self.c_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.c_bn.final_bn = True def forward(self, x): for layer in self.children(): x = layer(x) return x class TalkBottleneckTransform(nn.Module): """Bottleneck transformation: 1x1, 3x3, 1x1, relational graph""" def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): super(TalkBottleneckTransform, self).__init__() dim_inner = int(round(dim_out / 4)) self.seed = seed self._construct_class(dim_in, dim_out, stride, dim_inner, num_gs, seed) def _construct_class(self, dim_in, dim_out, stride, dim_inner, num_gs, seed): # MSRA -> stride=2 is on 1x1; TH/C2 -> stride=2 is on 3x3 # (str1x1, str3x3) = (stride, 1) if cfg.RESNET.STRIDE_1X1 else (1, stride) (str1x1, str3x3) = (1, stride) # 1x1, BN, ReLU self.a = TalkConv2d( dim_in, dim_inner, cfg.RGRAPH.GROUP_NUM, kernel_size=1, stride=str1x1, padding=0, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.a_bn = nn.BatchNorm2d( dim_inner, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # 3x3, BN, ReLU self.b = TalkConv2d( dim_inner, dim_inner, cfg.RGRAPH.GROUP_NUM, kernel_size=3, stride=str3x3, padding=1, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.b_bn = nn.BatchNorm2d( dim_inner, eps=cfg.BN.EPS, momentum=cfg.BN.MOM ) self.b_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # 1x1, BN self.c = TalkConv2d( dim_inner, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=1, stride=1, padding=0, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.c_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.c_bn.final_bn = True def forward(self, x): for layer in self.children(): x = layer(x) return x ##### Remaining ResNet code class ResBlock(nn.Module): """Residual block: x + F(x)""" def __init__( self, dim_in, dim_out, stride, trans_fun, dim_inner=None, num_gs=1, seed=None): super(ResBlock, self).__init__() self.seed = seed self._construct_class(dim_in, dim_out, stride, trans_fun, dim_inner, num_gs, seed) def _add_skip_proj(self, dim_in, dim_out, stride): if 'group' in cfg.RESNET.TRANS_FUN and 'share' not in cfg.RESNET.TRANS_FUN: self.proj = TalkConv2d( dim_in, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=1, stride=stride, padding=0, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) else: self.proj = nn.Conv2d( dim_in, dim_out, kernel_size=1, stride=stride, padding=0, bias=False ) self.bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) def _construct_class(self, dim_in, dim_out, stride, trans_fun, dim_inner, num_gs, seed): # Use skip connection with projection if dim or res change self.proj_block = (dim_in != dim_out) or (stride != 1) if self.proj_block: self._add_skip_proj(dim_in, dim_out, stride) self.f = trans_fun(dim_in, dim_out, stride, dim_inner, num_gs, seed) self.act = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) def forward(self, x): if self.proj_block: x = self.bn(self.proj(x)) + self.f(x) else: x = x + self.f(x) x = self.act(x) return x class ResStage(nn.Module): """Stage of ResNet.""" def __init__( self, dim_in, dim_out, stride, num_bs, dim_inner=None, num_gs=1): super(ResStage, self).__init__() self._construct_class(dim_in, dim_out, stride, num_bs, dim_inner, num_gs) def _construct_class(self, dim_in, dim_out, stride, num_bs, dim_inner, num_gs): if cfg.RGRAPH.KEEP_GRAPH: seed = cfg.RGRAPH.SEED_GRAPH else: seed = int(cfg.RGRAPH.SEED_GRAPH * 100) for i in range(num_bs): # Stride and dim_in apply to the first block of the stage b_stride = stride if i == 0 else 1 b_dim_in = dim_in if i == 0 else dim_out # Retrieve the transformation function trans_fun = get_trans_fun(cfg.RESNET.TRANS_FUN) # Construct the block res_block = ResBlock( b_dim_in, dim_out, b_stride, trans_fun, dim_inner, num_gs, seed=seed ) if not cfg.RGRAPH.KEEP_GRAPH: seed += 1 self.add_module('b{}'.format(i + 1), res_block) for j in range(cfg.RGRAPH.ADD_1x1): trans_fun = get_trans_fun(cfg.RESNET.TRANS_FUN + '1x1') # Construct the block res_block = ResBlock( dim_out, dim_out, 1, trans_fun, dim_inner, num_gs, seed=seed ) if not cfg.RGRAPH.KEEP_GRAPH: seed += 1 self.add_module('b{}_{}1x1'.format(i + 1, j + 1), res_block) def forward(self, x): for block in self.children(): x = block(x) return x class ResStem(nn.Module): """Stem of ResNet.""" def __init__(self, dim_in, dim_out): assert cfg.TRAIN.DATASET == cfg.TEST.DATASET, \ 'Train and test dataset must be the same for now' super(ResStem, self).__init__() if cfg.TRAIN.DATASET == 'cifar10': self._construct_cifar(dim_in, dim_out) else: self._construct_imagenet(dim_in, dim_out) def _construct_cifar(self, dim_in, dim_out): # 3x3, BN, ReLU # self.conv = nn.Conv2d( # dim_in, dim_out, kernel_size=3, # stride=1, padding=1, bias=False # ) self.conv = nn.Conv2d( dim_in, dim_out, kernel_size=7, stride=1, padding=3, bias=False ) self.bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) def _construct_imagenet(self, dim_in, dim_out): # 7x7, BN, ReLU, pool self.conv = nn.Conv2d( dim_in, dim_out, kernel_size=7, stride=2, padding=3, bias=False ) self.bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) def forward(self, x): for layer in self.children(): x = layer(x) return x class ResHead(nn.Module): """ResNet head.""" def __init__(self, dim_in, num_classes): super(ResHead, self).__init__() self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) self.fc = nn.Linear(dim_in, num_classes, bias=True) def forward(self, x): x = self.avg_pool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x class ResNet(nn.Module): """ResNet model.""" def __init__(self): assert cfg.TRAIN.DATASET in ['cifar10', 'cifar100', 'tinyimagenet200', 'imagenet'], \ 'Training ResNet on {} is not supported'.format(cfg.TRAIN.DATASET) assert cfg.TEST.DATASET in ['cifar10', 'cifar100', 'tinyimagenet200', 'imagenet'], \ 'Testing ResNet on {} is not supported'.format(cfg.TEST.DATASET) assert cfg.TRAIN.DATASET == cfg.TEST.DATASET, \ 'Train and test dataset must be the same for now' super(ResNet, self).__init__() if cfg.TRAIN.DATASET == 'cifar10': self._construct_cifar() elif cfg.TRAIN.DATASET == 'cifar100': self._construct_cifar() else: self._construct_imagenet() self.apply(nu.init_weights) # # ##### basic transform def _construct_cifar(self): assert (cfg.MODEL.DEPTH - 2) % 6 == 0, \ 'Model depth should be of the format 6n + 2 for cifar' logger.info('Constructing: ResNet-{}, cifar'.format(cfg.MODEL.DEPTH)) # Each stage has the same number of blocks for cifar num_blocks = int((cfg.MODEL.DEPTH - 2) / 6) # length = num of stages (excluding stem and head) dim_list = cfg.RGRAPH.DIM_LIST # Stage 1: (N, 3, 32, 32) -> (N, 16, 32, 32)*8 # self.s1 = ResStem(dim_in=3, dim_out=16) self.s1 = ResStem(dim_in=3, dim_out=64) # Stage 2: (N, 16, 32, 32) -> (N, 16, 32, 32) # self.s2 = ResStage(dim_in=16, dim_out=dim_list[0], stride=1, num_bs=num_blocks) self.s2 = ResStage(dim_in=64, dim_out=dim_list[0], stride=1, num_bs=num_blocks) # Stage 3: (N, 16, 32, 32) -> (N, 32, 16, 16) self.s3 = ResStage(dim_in=dim_list[0], dim_out=dim_list[1], stride=2, num_bs=num_blocks) # Stage 4: (N, 32, 16, 16) -> (N, 64, 8, 8) self.s4 = ResStage(dim_in=dim_list[1], dim_out=dim_list[2], stride=2, num_bs=num_blocks) # Head: (N, 64, 8, 8) -> (N, num_classes) self.head = ResHead(dim_in=dim_list[2], num_classes=cfg.MODEL.NUM_CLASSES) # smaller imagenet def _construct_imagenet(self): logger.info('Constructing: ResNet-{}, Imagenet'.format(cfg.MODEL.DEPTH)) # Retrieve the number of blocks per stage (excluding base) (d2, d3, d4, d5) = _IN_MODEL_STAGE_DS[cfg.MODEL.DEPTH] # Compute the initial inner block dim dim_list = cfg.RGRAPH.DIM_LIST print(dim_list) # Stage 1: (N, 3, 224, 224) -> (N, 64, 56, 56) self.s1 = ResStem(dim_in=3, dim_out=64) # Stage 2: (N, 64, 56, 56) -> (N, 256, 56, 56) self.s2 = ResStage( dim_in=64, dim_out=dim_list[0], stride=1, num_bs=d2 ) # Stage 3: (N, 256, 56, 56) -> (N, 512, 28, 28) self.s3 = ResStage( dim_in=dim_list[0], dim_out=dim_list[1], stride=2, num_bs=d3 ) # Stage 4: (N, 512, 56, 56) -> (N, 1024, 14, 14) self.s4 = ResStage( dim_in=dim_list[1], dim_out=dim_list[2], stride=2, num_bs=d4 ) # Stage 5: (N, 1024, 14, 14) -> (N, 2048, 7, 7) self.s5 = ResStage( dim_in=dim_list[2], dim_out=dim_list[3], stride=2, num_bs=d5 ) # Head: (N, 2048, 7, 7) -> (N, num_classes) self.head = ResHead(dim_in=dim_list[3], num_classes=cfg.MODEL.NUM_CLASSES) def forward(self, x): for module in self.children(): x = module(x) return x
20,015
37.198473
108
py
RobDanns
RobDanns-main/deep_learning/pycls/models/cnn.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """CNN model.""" import torch.nn as nn import torch from pycls.config import cfg import pycls.utils.logging as lu import pycls.utils.net as nu from .relation_graph import * logger = lu.get_logger(__name__) def get_trans_fun(name): """Retrieves the transformation function by name.""" trans_funs = { ##### (1) Level 1: channel ### (1.1) Basic Conv 'convbasic_transform': ConvBasicTransform, 'symconvbasic_transform': SymConvBasicTransform, 'convtalk_transform': ConvTalkTransform, # relational graph } assert name in trans_funs.keys(), \ 'Transformation function \'{}\' not supported'.format(name) return trans_funs[name] ##### (1) Level 1: channel ### (1.1) Basic Conv class ConvBasicTransform(nn.Module): """Basic transformation: 3x3""" def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): super(ConvBasicTransform, self).__init__() self._construct_class(dim_in, dim_out, stride) def _construct_class(self, dim_in, dim_out, stride): # 3x3, BN, ReLU self.a = nn.Conv2d( dim_in, dim_out, kernel_size=3, stride=stride, padding=1, bias=False ) self.a_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) # self.a_bn.final_bn = True self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) def forward(self, x): for layer in self.children(): x = layer(x) return x class SymConvBasicTransform(nn.Module): """Basic transformation: 3x3 conv, symmetric""" def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): super(SymConvBasicTransform, self).__init__() self._construct_class(dim_in, dim_out, stride) def _construct_class(self, dim_in, dim_out, stride): # 3x3, BN, ReLU self.a = SymConv2d( dim_in, dim_out, kernel_size=3, stride=stride, padding=1, bias=False ) self.a_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) # self.a_bn.final_bn = True self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) def forward(self, x): for layer in self.children(): x = layer(x) return x class ConvTalkTransform(nn.Module): """Basic transformation: 3x3 conv, relational graph""" def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): self.seed = seed super(ConvTalkTransform, self).__init__() self._construct_class(dim_in, dim_out, stride) def _construct_class(self, dim_in, dim_out, stride): # 3x3, BN, ReLU self.a = TalkConv2d( dim_in, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=3, stride=stride, padding=1, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.a_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) # self.a_bn.final_bn = True self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) def forward(self, x): for layer in self.children(): x = layer(x) return x ##### Remaining CNN code class CNNStage(nn.Module): """Stage of CNN.""" def __init__( self, dim_in, dim_out, stride, num_bs, dim_inner=None, num_gs=1): super(CNNStage, self).__init__() self._construct_class(dim_in, dim_out, stride, num_bs, dim_inner, num_gs) def _construct_class(self, dim_in, dim_out, stride, num_bs, dim_inner, num_gs): if cfg.RGRAPH.KEEP_GRAPH: seed = cfg.RGRAPH.SEED_GRAPH else: seed = int(cfg.RGRAPH.SEED_GRAPH * 100) for i in range(num_bs): # Stride and dim_in apply to the first block of the stage b_stride = stride if i == 0 else 1 b_dim_in = dim_in if i == 0 else dim_out # Retrieve the transformation function trans_fun = get_trans_fun(cfg.RESNET.TRANS_FUN) # Construct the block res_block = trans_fun( b_dim_in, dim_out, b_stride, dim_inner, num_gs, seed=seed ) if not cfg.RGRAPH.KEEP_GRAPH: seed += 1 self.add_module('b{}'.format(i + 1), res_block) def forward(self, x): for block in self.children(): x = block(x) return x class CNNStem(nn.Module): """Stem of CNN.""" def __init__(self, dim_in, dim_out): assert cfg.TRAIN.DATASET == cfg.TEST.DATASET, \ 'Train and test dataset must be the same for now' super(CNNStem, self).__init__() if cfg.TRAIN.DATASET == 'cifar10': self._construct_cifar(dim_in, dim_out) elif cfg.TRAIN.DATASET == 'cifar100': self._construct_cifar(dim_in, dim_out) else: self._construct_imagenet(dim_in, dim_out) def _construct_cifar(self, dim_in, dim_out): # 3x3, BN, ReLU if cfg.RGRAPH.STEM_MODE == 'default': self.conv = nn.Conv2d( dim_in, dim_out, kernel_size=3, stride=1, padding=1, bias=False ) self.bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) elif cfg.RGRAPH.STEM_MODE == 'downsample': self.conv = nn.Conv2d( dim_in, dim_out, kernel_size=3, stride=1, padding=1, bias=False ) self.bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) def _construct_imagenet(self, dim_in, dim_out): # 3x3, BN, ReLU, pool self.conv = nn.Conv2d( dim_in, dim_out, kernel_size=3, stride=2, padding=1, bias=False ) self.bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) def forward(self, x): for layer in self.children(): x = layer(x) return x class CNNHead(nn.Module): """CNN head.""" def __init__(self, dim_in, num_classes): super(CNNHead, self).__init__() self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) self.dropout = nn.Dropout(p=0.15) self.fc = nn.Linear(dim_in, num_classes, bias=True) def forward(self, x): x = self.avg_pool(x) x = x.view(x.size(0), -1) x = self.dropout(x) x = self.fc(x) return x class CNN(nn.Module): """CNN model.""" def __init__(self): assert cfg.TRAIN.DATASET in ['cifar10', 'cifar100', 'tinyimagenet200', 'imagenet'], \ 'Training CNN on {} is not supported'.format(cfg.TRAIN.DATASET) assert cfg.TEST.DATASET in ['cifar10', 'cifar100', 'tinyimagenet200', 'imagenet'], \ 'Testing CNN on {} is not supported'.format(cfg.TEST.DATASET) assert cfg.TRAIN.DATASET == cfg.TEST.DATASET, \ 'Train and test dataset must be the same for now' super(CNN, self).__init__() self._construct() self.apply(nu.init_weights) # # ##### basic transform def _construct(self): # Each stage has the same number of blocks for cifar dim_list = cfg.RGRAPH.DIM_LIST num_bs = cfg.MODEL.LAYERS // 3 self.s1 = CNNStem(dim_in=3, dim_out=cfg.RGRAPH.DIM_FIRST) self.s2 = CNNStage(dim_in=cfg.RGRAPH.DIM_FIRST, dim_out=dim_list[0], stride=2, num_bs=num_bs) self.s3 = CNNStage(dim_in=dim_list[0], dim_out=dim_list[1], stride=2, num_bs=num_bs) self.s4 = CNNStage(dim_in=dim_list[1], dim_out=dim_list[2], stride=2, num_bs=num_bs) # self.s5 = CNNStage(dim_in=dim_list[2], dim_out=dim_list[3], stride=2, num_bs=num_bs) self.head = CNNHead(dim_in=dim_list[2], num_classes=cfg.MODEL.NUM_CLASSES) def forward(self, x): for module in self.children(): x = module(x) return x # #!/usr/bin/env python3 # # Copyright (c) Facebook, Inc. and its affiliates. # # # # This source code is licensed under the MIT license found in the # # LICENSE file in the root directory of this source tree. # """CNN model.""" # import torch.nn as nn # import torch # from pycls.config import cfg # import pycls.utils.logging as lu # import pycls.utils.net as nu # from .relation_graph import * # logger = lu.get_logger(__name__) # def get_trans_fun(name): # """Retrieves the transformation function by name.""" # trans_funs = { # ##### (1) Level 1: channel # ### (1.1) Basic Conv # 'convbasic_transform': ConvBasicTransform, # 'symconvbasic_transform': SymConvBasicTransform, # 'convtalk_transform': ConvTalkTransform, # relational graph # } # assert name in trans_funs.keys(), \ # 'Transformation function \'{}\' not supported'.format(name) # return trans_funs[name] # ##### (1) Level 1: channel # ### (1.1) Basic Conv # class ConvBasicTransform(nn.Module): # """Basic transformation: 3x3""" # def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): # super(ConvBasicTransform, self).__init__() # self._construct_class(dim_in, dim_out, stride) # def _construct_class(self, dim_in, dim_out, stride): # # 3x3, BN, ReLU # self.a = nn.Conv2d( # dim_in, dim_out, kernel_size=3, # stride=stride, padding=1, bias=False # ) # self.a_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) # # self.a_bn.final_bn = True # self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # def forward(self, x): # for layer in self.children(): # x = layer(x) # return x # class SymConvBasicTransform(nn.Module): # """Basic transformation: 3x3 conv, symmetric""" # def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): # super(SymConvBasicTransform, self).__init__() # self._construct_class(dim_in, dim_out, stride) # def _construct_class(self, dim_in, dim_out, stride): # # 3x3, BN, ReLU # self.a = SymConv2d( # dim_in, dim_out, kernel_size=3, # stride=stride, padding=1, bias=False # ) # self.a_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) # # self.a_bn.final_bn = True # self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # def forward(self, x): # for layer in self.children(): # x = layer(x) # return x # class ConvTalkTransform(nn.Module): # """Basic transformation: 3x3 conv, relational graph""" # def __init__(self, dim_in, dim_out, stride, dim_inner=None, num_gs=1, seed=None): # self.seed = seed # super(ConvTalkTransform, self).__init__() # self._construct_class(dim_in, dim_out, stride) # def _construct_class(self, dim_in, dim_out, stride): # # 3x3, BN, ReLU # self.a = TalkConv2d( # dim_in, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=3, # stride=stride, padding=1, bias=False, # message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, # sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed # ) # self.a_bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) # # self.a_bn.final_bn = True # self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE) # def forward(self, x): # for layer in self.children(): # x = layer(x) # return x # ##### Remaining CNN code # class CNNStage(nn.Module): # """Stage of CNN.""" # def __init__( # self, dim_in, dim_out, stride, num_bs, dim_inner=None, num_gs=1): # super(CNNStage, self).__init__() # self._construct_class(dim_in, dim_out, stride, num_bs, dim_inner, num_gs) # def _construct_class(self, dim_in, dim_out, stride, num_bs, dim_inner, num_gs): # if cfg.RGRAPH.KEEP_GRAPH: # seed = cfg.RGRAPH.SEED_GRAPH # else: # seed = int(cfg.RGRAPH.SEED_GRAPH * 100) # for i in range(num_bs): # # Stride and dim_in apply to the first block of the stage # b_stride = stride if i == 0 else 1 # b_dim_in = dim_in if i == 0 else dim_out # # Retrieve the transformation function # trans_fun = get_trans_fun(cfg.RESNET.TRANS_FUN) # # Construct the block # res_block = trans_fun( # b_dim_in, dim_out, b_stride, dim_inner, num_gs, seed=seed # ) # if not cfg.RGRAPH.KEEP_GRAPH: # seed += 1 # self.add_module('b{}'.format(i + 1), res_block) # def forward(self, x): # for block in self.children(): # x = block(x) # return x # class CNNStem(nn.Module): # """Stem of CNN.""" # def __init__(self, dim_in, dim_out): # assert cfg.TRAIN.DATASET == cfg.TEST.DATASET, \ # 'Train and test dataset must be the same for now' # super(CNNStem, self).__init__() # if cfg.TRAIN.DATASET == 'cifar10': # self._construct_cifar(dim_in, dim_out) # else: # self._construct_imagenet(dim_in, dim_out) # def _construct_cifar(self, dim_in, dim_out): # # 3x3, BN, ReLU # if cfg.RGRAPH.STEM_MODE == 'default': # self.conv = nn.Conv2d( # dim_in, dim_out, kernel_size=3, # stride=1, padding=1, bias=False # ) # self.bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, # momentum=cfg.BN.MOM) # self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) # elif cfg.RGRAPH.STEM_MODE == 'downsample': # self.conv = nn.Conv2d( # dim_in, dim_out, kernel_size=3, # stride=1, padding=1, bias=False # ) # self.bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, # momentum=cfg.BN.MOM) # self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) # self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) # def _construct_imagenet(self, dim_in, dim_out): # # 3x3, BN, ReLU, pool # self.conv = nn.Conv2d( # dim_in, dim_out, kernel_size=3, # stride=2, padding=1, bias=False # ) # self.bn = nn.BatchNorm2d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) # self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) # self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) # def forward(self, x): # for layer in self.children(): # x = layer(x) # return x # class CNNHead(nn.Module): # """CNN head.""" # def __init__(self, dim_in, num_classes): # super(CNNHead, self).__init__() # self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) # self.fc = nn.Linear(dim_in, num_classes, bias=True) # def forward(self, x): # x = self.avg_pool(x) # x = x.view(x.size(0), -1) # x = self.fc(x) # return x # class CNN(nn.Module): # """CNN model.""" # def __init__(self): # assert cfg.TRAIN.DATASET in ['cifar10', 'imagenet'], \ # 'Training ResNet on {} is not supported'.format(cfg.TRAIN.DATASET) # assert cfg.TEST.DATASET in ['cifar10', 'imagenet'], \ # 'Testing ResNet on {} is not supported'.format(cfg.TEST.DATASET) # assert cfg.TRAIN.DATASET == cfg.TEST.DATASET, \ # 'Train and test dataset must be the same for now' # super(CNN, self).__init__() # self._construct() # self.apply(nu.init_weights) # # # ##### basic transform # def _construct(self): # # Each stage has the same number of blocks for cifar # dim_list = cfg.RGRAPH.DIM_LIST # num_bs = cfg.MODEL.LAYERS // 3 # self.s1 = CNNStem(dim_in=3, dim_out=cfg.RGRAPH.DIM_FIRST) # self.s2 = CNNStage(dim_in=cfg.RGRAPH.DIM_FIRST, dim_out=dim_list[0], stride=2, num_bs=num_bs) # self.s3 = CNNStage(dim_in=dim_list[0], dim_out=dim_list[1], stride=2, num_bs=num_bs) # self.s4 = CNNStage(dim_in=dim_list[1], dim_out=dim_list[2], stride=2, num_bs=num_bs) # self.head = CNNHead(dim_in=dim_list[2], num_classes=cfg.MODEL.NUM_CLASSES) # def forward(self, x): # for module in self.children(): # x = module(x) # return x
17,388
34.779835
107
py
RobDanns
RobDanns-main/deep_learning/pycls/models/vgg.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """VGG example""" import torch.nn as nn import torch.nn.functional as F from pycls.config import cfg import pycls.utils.net as nu from .relation_graph import * class VGG(nn.Module): def __init__(self, num_classes=1024): super(VGG, self).__init__() self.seed = cfg.RGRAPH.SEED_GRAPH def conv_bn(dim_in, dim_out, stride, stem=False): if stem: conv = get_conv('convbasic_transform', dim_in, dim_out, stride) else: conv = get_conv(cfg.RESNET.TRANS_FUN, dim_in, dim_out, stride) return nn.Sequential( conv, nn.BatchNorm2d(dim_out), nn.ReLU(inplace=True) ) def get_conv(name, dim_in, dim_out, stride=1): if not cfg.RGRAPH.KEEP_GRAPH: self.seed += 1 if name == 'convbasic_transform': return nn.Conv2d(dim_in, dim_out, kernel_size=3, stride=stride, padding=1, bias=False) elif name == 'convtalk_transform': return TalkConv2d( dim_in, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=3, stride=stride, padding=1, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) self.dim_list = cfg.RGRAPH.DIM_LIST # print(self.dim_list) self.model = nn.Sequential( conv_bn(3, 64, 1, stem=True), conv_bn(64, self.dim_list[0], 1), nn.MaxPool2d(kernel_size=2, stride=2), conv_bn(self.dim_list[0], self.dim_list[1], 1), conv_bn(self.dim_list[1], self.dim_list[1], 1), nn.MaxPool2d(kernel_size=2, stride=2), conv_bn(self.dim_list[1], self.dim_list[2], 1), conv_bn(self.dim_list[2], self.dim_list[2], 1), nn.MaxPool2d(kernel_size=2, stride=2), conv_bn(self.dim_list[2], self.dim_list[3], 1), conv_bn(self.dim_list[3], self.dim_list[3], 1), nn.MaxPool2d(kernel_size=2, stride=2), conv_bn(self.dim_list[3], self.dim_list[3], 1), conv_bn(self.dim_list[3], self.dim_list[3], 1), ) self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) self.fc = nn.Linear(self.dim_list[3], num_classes) self.apply(nu.init_weights) def forward(self, x): x = self.model(x) x = self.avg_pool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x
3,097
35.880952
107
py
RobDanns
RobDanns-main/deep_learning/pycls/models/mlp.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """MLP model.""" import torch.nn as nn import torch from pycls.config import cfg import pycls.utils.logging as lu import pycls.utils.net as nu from .relation_graph import * import time import pdb logger = lu.get_logger(__name__) def get_trans_fun(name): """Retrieves the transformation function by name.""" trans_funs = { ##### (1) Level 1: channel 'linear_transform': LinearTransform, 'symlinear_transform': SymLinearTransform, 'grouplinear_transform': GroupLinearTransform, 'groupshufflelinear_transform': GroupShuffleLinearTransform, 'talklinear_transform': TalkLinearTransform, # relational graph } assert name in trans_funs.keys(), \ 'Transformation function \'{}\' not supported'.format(name) return trans_funs[name] ##### (0) Basic class LinearTransform(nn.Module): """Basic transformation: linear""" def __init__(self, dim_in, dim_out, seed=None): super(LinearTransform, self).__init__() self._construct_class(dim_in, dim_out) def _construct_class(self, dim_in, dim_out): # 3x3, BN, ReLU self.a = nn.Linear( dim_in, dim_out, bias=False ) self.a_bn = nn.BatchNorm1d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.a_bn.final_bn = True self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) def forward(self, x): for layer in self.children(): x = layer(x) return x class SymLinearTransform(nn.Module): """Basic transformation: linear, symmetric""" def __init__(self, dim_in, dim_out, seed=None): super(SymLinearTransform, self).__init__() self._construct_class(dim_in, dim_out) def _construct_class(self, dim_in, dim_out): # 3x3, BN, ReLU self.a = SymLinear( dim_in, dim_out, bias=False ) self.a_bn = nn.BatchNorm1d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.a_bn.final_bn = True self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) def forward(self, x): for layer in self.children(): x = layer(x) return x class GroupLinearTransform(nn.Module): """Basic transformation: linear, group""" def __init__(self, dim_in, dim_out, seed=None): super(GroupLinearTransform, self).__init__() self._construct_class(dim_in, dim_out) def _construct_class(self, dim_in, dim_out): # 3x3, BN, ReLU self.a = GroupLinear( dim_in, dim_out, bias=False, group_size=cfg.RGRAPH.GROUP_SIZE ) self.a_bn = nn.BatchNorm1d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.a_bn.final_bn = True self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) def forward(self, x): for layer in self.children(): x = layer(x) return x class GroupShuffleLinearTransform(nn.Module): """Basic transformation: linear, shuffle""" def __init__(self, dim_in, dim_out, seed=None): super(GroupShuffleLinearTransform, self).__init__() self._construct_class(dim_in, dim_out) def _construct_class(self, dim_in, dim_out): # 3x3, BN, ReLU self.a = GroupLinear( dim_in, dim_out, bias=False, group_size=cfg.RGRAPH.GROUP_SIZE ) self.a_bn = nn.BatchNorm1d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.a_bn.final_bn = True self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) self.shuffle_shape = (dim_out // cfg.RGRAPH.GROUP_NUM, cfg.RGRAPH.GROUP_NUM) def forward(self, x): x = self.a(x) x = x.view(x.shape[0], self.shuffle_shape[0], self.shuffle_shape[1]).permute(0, 2, 1).contiguous() x = x.view(x.shape[0], x.shape[1] * x.shape[2]) x = self.a_bn(x) x = self.relu(x) return x class TalkLinearTransform(nn.Module): """Basic transformation: linear, relational graph""" def __init__(self, dim_in, dim_out, seed=None): self.seed = seed super(TalkLinearTransform, self).__init__() self._construct_class(dim_in, dim_out) def _construct_class(self, dim_in, dim_out): self.a = TalkLinear( dim_in, dim_out, cfg.RGRAPH.GROUP_NUM, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed) self.a_bn = nn.BatchNorm1d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.a_bn.final_bn = True self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) def forward(self, x): for layer in self.children(): x = layer(x) return x class MLPStage(nn.Module): """Stage of MLPNet.""" def __init__( self, dim_in, dim_out, num_bs): super(MLPStage, self).__init__() self._construct_class(dim_in, dim_out, num_bs) def _construct_class(self, dim_in, dim_out, num_bs): if cfg.RGRAPH.KEEP_GRAPH: seed = cfg.RGRAPH.SEED_GRAPH else: seed = int(dim_out * 100 * cfg.RGRAPH.SPARSITY) for i in range(num_bs): b_dim_in = dim_in if i == 0 else dim_out trans_fun = get_trans_fun(cfg.RESNET.TRANS_FUN) res_block = trans_fun( b_dim_in, dim_out, seed=seed ) if not cfg.RGRAPH.KEEP_GRAPH: seed += 1 self.add_module('b{}'.format(i + 1), res_block) def forward(self, x): for block in self.children(): x = block(x) return x class MLPStem(nn.Module): """Stem of MLPNet.""" def __init__(self, dim_in, dim_out): super(MLPStem, self).__init__() if cfg.TRAIN.DATASET == 'cifar10': self._construct_cifar(dim_in, dim_out) else: raise NotImplementedError def _construct_cifar(self, dim_in, dim_out): self.linear = nn.Linear( dim_in, dim_out, bias=False ) self.bn = nn.BatchNorm1d(dim_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM) self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE) def forward(self, x): x = x.view(x.size(0), -1) for layer in self.children(): x = layer(x) return x class MLPHead(nn.Module): """MLPNet head.""" def __init__(self, dim_in, num_classes): super(MLPHead, self).__init__() self.fc = nn.Linear(dim_in, num_classes, bias=True) def forward(self, x): x = self.fc(x) return x class MLPNet(nn.Module): """MLPNet model.""" def __init__(self): assert cfg.TRAIN.DATASET in ['cifar10'], \ 'Training MLPNet on {} is not supported'.format(cfg.TRAIN.DATASET) assert cfg.TEST.DATASET in ['cifar10'], \ 'Testing MLPNet on {} is not supported'.format(cfg.TEST.DATASET) assert cfg.TRAIN.DATASET == cfg.TEST.DATASET, \ 'Train and test dataset must be the same for now' super(MLPNet, self).__init__() if cfg.TRAIN.DATASET == 'cifar10': self._construct_cifar() else: raise NotImplementedError self.apply(nu.init_weights) # ##### basic transform def _construct_cifar(self): num_layers = cfg.MODEL.LAYERS dim_inner = cfg.RGRAPH.DIM_LIST[0] dim_first = cfg.RGRAPH.DIM_FIRST self.s1 = MLPStem(dim_in=3072, dim_out=dim_first) self.s2 = MLPStage(dim_in=dim_first, dim_out=dim_inner, num_bs=num_layers) self.head = MLPHead(dim_in=dim_inner, num_classes=cfg.MODEL.NUM_CLASSES) def forward(self, x): for module in self.children(): x = module(x) return x
8,012
30.300781
107
py
RobDanns
RobDanns-main/deep_learning/pycls/models/model_builder.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Model construction functions.""" import torch from pycls.config import cfg from pycls.models.resnet import ResNet from pycls.models.mlp import MLPNet from pycls.models.cnn import CNN from pycls.models.mobilenet import MobileNetV1 from pycls.models.efficientnet import EfficientNet from pycls.models.vgg import VGG import pycls.utils.logging as lu import pycls.utils.metrics as mu logger = lu.get_logger(__name__) # Supported model types _MODEL_TYPES = { 'resnet': ResNet, 'mlpnet': MLPNet, 'cnn': CNN, 'mobilenet': MobileNetV1, 'efficientnet': EfficientNet, 'vgg': VGG, } def build_model(): """Builds the model.""" assert cfg.MODEL.TYPE in _MODEL_TYPES.keys(), \ 'Model type \'{}\' not supported'.format(cfg.MODEL.TYPE) assert cfg.NUM_GPUS <= torch.cuda.device_count(), \ 'Cannot use more GPU devices than available' # Construct the model model = _MODEL_TYPES[cfg.MODEL.TYPE]() # Determine the GPU used by the current process cur_device = torch.cuda.current_device() # Transfer the model to the current GPU device model = model.cuda(device=cur_device) # Use multi-process data parallel model in the multi-gpu setting if cfg.NUM_GPUS > 1: # Make model replica operate on the current device model = torch.nn.parallel.DistributedDataParallel( module=model, device_ids=[cur_device], output_device=cur_device ) return model ## auto match flop def build_model_stats(mode='flops'): """Builds the model.""" assert cfg.MODEL.TYPE in _MODEL_TYPES.keys(), \ 'Model type \'{}\' not supported'.format(cfg.MODEL.TYPE) assert cfg.NUM_GPUS <= torch.cuda.device_count(), \ 'Cannot use more GPU devices than available' # Construct the model model = _MODEL_TYPES[cfg.MODEL.TYPE]() if mode == 'flops': flops = mu.flops_count(model) return flops else: params = mu.params_count(model) return params
2,355
30
107
py
RobDanns
RobDanns-main/deep_learning/pycls/models/mobilenet.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """MobileNet example""" import torch.nn as nn import torch.nn.functional as F from pycls.config import cfg import pycls.utils.net as nu from .relation_graph import * class MobileNetV1(nn.Module): def __init__(self, num_classes=1024): super(MobileNetV1, self).__init__() if cfg.RGRAPH.KEEP_GRAPH: self.seed = cfg.RGRAPH.SEED_GRAPH else: self.seed = int(cfg.RGRAPH.SEED_GRAPH * 100) def conv_bn(dim_in, dim_out, stride): return nn.Sequential( nn.Conv2d(dim_in, dim_out, 3, stride, 1, bias=False), nn.BatchNorm2d(dim_out), nn.ReLU(inplace=True) ) def get_conv(name, dim_in, dim_out): if not cfg.RGRAPH.KEEP_GRAPH: self.seed += 1 if name == 'channelbasic_transform': return nn.Conv2d(dim_in, dim_out, 1, 1, 0, bias=False) elif name == 'groupbasictalk_transform': return TalkConv2d( dim_in, dim_out, cfg.RGRAPH.GROUP_NUM, kernel_size=1, stride=1, padding=0, bias=False, message_type=cfg.RGRAPH.MESSAGE_TYPE, directed=cfg.RGRAPH.DIRECTED, agg=cfg.RGRAPH.AGG_FUNC, sparsity=cfg.RGRAPH.SPARSITY, p=cfg.RGRAPH.P, talk_mode=cfg.RGRAPH.TALK_MODE, seed=self.seed ) def conv_dw(dim_in, dim_out, stride): conv1x1 = get_conv(cfg.RESNET.TRANS_FUN, dim_in, dim_out) return nn.Sequential( nn.Conv2d(dim_in, dim_in, 3, stride, 1, groups=dim_in, bias=False), nn.BatchNorm2d(dim_in), nn.ReLU(inplace=True), conv1x1, nn.BatchNorm2d(dim_out), nn.ReLU(inplace=True), ) self.dim_list = cfg.RGRAPH.DIM_LIST # print(self.dim_list) self.model = nn.Sequential( conv_bn(3, 32, 2), conv_dw(32, self.dim_list[1], 1), conv_dw(self.dim_list[1], self.dim_list[2], 2), conv_dw(self.dim_list[2], self.dim_list[2], 1), conv_dw(self.dim_list[2], self.dim_list[3], 2), conv_dw(self.dim_list[3], self.dim_list[3], 1), conv_dw(self.dim_list[3], self.dim_list[4], 2), conv_dw(self.dim_list[4], self.dim_list[4], 1), conv_dw(self.dim_list[4], self.dim_list[4], 1), conv_dw(self.dim_list[4], self.dim_list[4], 1), conv_dw(self.dim_list[4], self.dim_list[4], 1), conv_dw(self.dim_list[4], self.dim_list[4], 1), conv_dw(self.dim_list[4], self.dim_list[5], 2), conv_dw(self.dim_list[5], self.dim_list[5], 1), ) self.fc = nn.Linear(self.dim_list[5], num_classes) self.apply(nu.init_weights) def forward(self, x): x = self.model(x) x = F.avg_pool2d(x, 7) x = x.view(-1, self.dim_list[5]) x = self.fc(x) return x
3,404
35.223404
107
py
RobDanns
RobDanns-main/deep_learning/pycls/models/optimizer.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Optimizer.""" import torch from pycls.config import cfg import pycls.utils.lr_policy as lr_policy def construct_optimizer(model): """Constructs the optimizer. Note that the momentum update in PyTorch differs from the one in Caffe2. In particular, Caffe2: V := mu * V + lr * g p := p - V PyTorch: V := mu * V + g p := p - lr * V where V is the velocity, mu is the momentum factor, lr is the learning rate, g is the gradient and p are the parameters. Since V is defined independently of the learning rate in PyTorch, when the learning rate is changed there is no need to perform the momentum correction by scaling V (unlike in the Caffe2 case). """ return torch.optim.SGD( model.parameters(), lr=cfg.OPTIM.BASE_LR, momentum=cfg.OPTIM.MOMENTUM, weight_decay=cfg.OPTIM.WEIGHT_DECAY, dampening=cfg.OPTIM.DAMPENING, nesterov=cfg.OPTIM.NESTEROV ) def get_epoch_lr(cur_epoch): """Retrieves the lr for the given epoch (as specified by the lr policy).""" return lr_policy.get_epoch_lr(cur_epoch) def set_lr(optimizer, new_lr): """Sets the optimizer lr to the specified value.""" for param_group in optimizer.param_groups: param_group['lr'] = new_lr
1,678
27.457627
107
py
RobDanns
RobDanns-main/deep_learning/pycls/models/relation_graph.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Relational graph modules""" import math import torch import torch.nn as nn from torch.nn.parameter import Parameter import torch.nn.functional as F import torch.nn.init as init import networkx as nx import numpy as np from torch.nn.modules.utils import _pair from torch.nn.modules.conv import _ConvNd from torch.autograd import Function from itertools import repeat from networkx.utils import py_random_state from pycls.datasets.load_graph import load_graph import pdb import time import random def compute_count(channel, group): divide = channel // group remain = channel % group out = np.zeros(group, dtype=int) out[:remain] = divide + 1 out[remain:] = divide return out @py_random_state(3) def ws_graph(n, k, p, seed=1): """Returns a ws-flex graph, k can be real number in [2,n] """ assert k >= 2 and k <= n # compute number of edges: edge_num = int(round(k * n / 2)) count = compute_count(edge_num, n) # print(count) G = nx.Graph() for i in range(n): source = [i] * count[i] target = range(i + 1, i + count[i] + 1) target = [node % n for node in target] # print(source, target) G.add_edges_from(zip(source, target)) # rewire edges from each node nodes = list(G.nodes()) for i in range(n): u = i target = range(i + 1, i + count[i] + 1) target = [node % n for node in target] for v in target: if seed.random() < p: w = seed.choice(nodes) # Enforce no self-loops or multiple edges while w == u or G.has_edge(u, w): w = seed.choice(nodes) if G.degree(u) >= n - 1: break # skip this rewiring else: G.remove_edge(u, v) G.add_edge(u, w) return G @py_random_state(4) def connected_ws_graph(n, k, p, tries=100, seed=1): """Returns a connected ws-flex graph. """ for i in range(tries): # seed is an RNG so should change sequence each call G = ws_graph(n, k, p, seed) if nx.is_connected(G): return G raise nx.NetworkXError('Maximum number of tries exceeded') def nx_to_edge(graph, directed=False, add_self_loops=True, shuffle_id=False, seed=1): '''nx graph to edge index''' graph.remove_edges_from(graph.selfloop_edges()) # relabel graphs keys = list(graph.nodes) vals = list(range(graph.number_of_nodes())) # shuffle node id assignment if shuffle_id: random.seed(seed) random.shuffle(vals) mapping = dict(zip(keys, vals)) graph = nx.relabel_nodes(graph, mapping, copy=True) # get edges edge_index = np.array(list(graph.edges)) if not directed: edge_index = np.concatenate((edge_index, edge_index[:, ::-1]), axis=0) if add_self_loops: edge_self = np.arange(graph.number_of_nodes())[:, np.newaxis] edge_self = np.tile(edge_self, (1, 2)) edge_index = np.concatenate((edge_index, edge_self), axis=0) # sort edges idx = np.argsort(edge_index[:, 0]) edge_index = edge_index[idx, :] return edge_index # edge index generator def generate_index(message_type='ba', n=16, sparsity=0.5, p=0.2, directed=False, seed=123): degree = n * sparsity known_names = ['mcwhole', 'mcwholeraw', 'mcvisual', 'mcvisualraw', 'cat', 'catraw'] if message_type == 'er': graph = nx.gnm_random_graph(n=n, m=n * degree // 2, seed=seed) elif message_type == 'random': edge_num = int(n * n * sparsity) edge_id = np.random.choice(n * n, edge_num, replace=False) edge_index = np.zeros((edge_num, 2), dtype=int) for i in range(edge_num): edge_index[i, 0] = edge_id[i] // n edge_index[i, 1] = edge_id[i] % n elif message_type == 'ws': graph = connected_ws_graph(n=n, k=degree, p=p, seed=seed) elif message_type == 'ba': graph = nx.barabasi_albert_graph(n=n, m=degree // 2, seed=seed) elif message_type == 'hypercube': graph = nx.hypercube_graph(n=int(np.log2(n))) elif message_type == 'grid': m = degree n = n // degree graph = nx.grid_2d_graph(m=m, n=n) elif message_type == 'cycle': graph = nx.cycle_graph(n=n) elif message_type == 'tree': graph = nx.random_tree(n=n, seed=seed) elif message_type == 'regular': graph = nx.connected_watts_strogatz_graph(n=n, k=degree, p=0, seed=seed) elif message_type in known_names: graph = load_graph(message_type) edge_index = nx_to_edge(graph, directed=True, seed=seed) else: raise NotImplementedError if message_type != 'random' and message_type not in known_names: edge_index = nx_to_edge(graph, directed=directed, seed=seed) return edge_index def compute_size(channel, group, seed=1): np.random.seed(seed) divide = channel // group remain = channel % group out = np.zeros(group, dtype=int) out[:remain] = divide + 1 out[remain:] = divide out = np.random.permutation(out) return out def compute_densemask(in_channels, out_channels, group_num, edge_index): repeat_in = compute_size(in_channels, group_num) repeat_out = compute_size(out_channels, group_num) mask = np.zeros((group_num, group_num)) mask[edge_index[:, 0], edge_index[:, 1]] = 1 mask = np.repeat(mask, repeat_out, axis=0) mask = np.repeat(mask, repeat_in, axis=1) return mask def get_mask(in_channels, out_channels, group_num, message_type='ba', directed=False, sparsity=0.5, p=0.2, talk_mode='dense', seed=123): assert group_num <= in_channels and group_num <= out_channels # high-level graph edge index edge_index_high = generate_index(message_type=message_type, n=group_num, sparsity=sparsity, p=p, directed=directed, seed=seed) # get in/out size for each high-level node in_sizes = compute_size(in_channels, group_num) out_sizes = compute_size(out_channels, group_num) # decide low-level node num group_num_low = int(min(np.min(in_sizes), np.min(out_sizes))) # decide how to fill each node mask_high = compute_densemask(in_channels, out_channels, group_num, edge_index_high) return mask_high ############## Linear model class TalkLinear(nn.Linear): '''Relational graph version of Linear. Neurons "talk" according to the graph structure''' def __init__(self, in_channels, out_channels, group_num, bias=False, message_type='ba', directed=False, sparsity=0.5, p=0.2, talk_mode='dense', seed=None): group_num_max = min(in_channels, out_channels) if group_num > group_num_max: group_num = group_num_max # print(group_num, in_channels, out_channels, kernel_size, stride) super(TalkLinear, self).__init__( in_channels, out_channels, bias) self.mask = get_mask(in_channels, out_channels, group_num, message_type, directed, sparsity, p, talk_mode, seed) nonzero = np.sum(self.mask) self.mask = torch.from_numpy(self.mask).float().cuda() self.flops_scale = nonzero / (in_channels * out_channels) self.params_scale = self.flops_scale self.init_scale = torch.sqrt(out_channels / torch.sum(self.mask.cpu(), dim=0, keepdim=True)) def forward(self, x): weight = self.weight * self.mask # pdb.set_trace() return F.linear(x, weight, self.bias) class SymLinear(nn.Module): '''Linear with symmetric weight matrices''' def __init__(self, in_features, out_features, bias=True): super(SymLinear, self).__init__() self.in_features = in_features self.out_features = out_features self.weight = Parameter(torch.Tensor(out_features, in_features)) if bias: self.bias = Parameter(torch.Tensor(out_features)) else: self.register_parameter('bias', None) self.reset_parameters() def reset_parameters(self): init.kaiming_uniform_(self.weight, a=math.sqrt(5)) if self.bias is not None: fan_in, _ = init._calculate_fan_in_and_fan_out(self.weight) bound = 1 / math.sqrt(fan_in) init.uniform_(self.bias, -bound, bound) def forward(self, input): weight = self.weight + self.weight.permute(1, 0) return F.linear(input, weight, self.bias) def extra_repr(self): return 'in_features={}, out_features={}, bias={}'.format( self.in_features, self.out_features, self.bias is not None ) ############## Conv model class TalkConv2d(_ConvNd): '''Relational graph version of Conv2d. Neurons "talk" according to the graph structure''' def __init__(self, in_channels, out_channels, group_num, kernel_size, stride=1, padding=0, dilation=1, bias=False, message_type='ba', directed=False, agg='sum', sparsity=0.5, p=0.2, talk_mode='dense', seed=None): group_num_max = min(in_channels, out_channels) if group_num > group_num_max: group_num = group_num_max kernel_size = _pair(kernel_size) stride = _pair(stride) padding = _pair(padding) dilation = _pair(dilation) super(TalkConv2d, self).__init__( in_channels, out_channels, kernel_size, stride, padding, dilation, False, _pair(0), 1, bias, 'zeros') self.mask = get_mask(in_channels, out_channels, group_num, message_type, directed, sparsity, p, talk_mode, seed) nonzero = np.sum(self.mask) self.mask = torch.from_numpy(self.mask[:, :, np.newaxis, np.newaxis]).float().cuda() self.init_scale = torch.sqrt(out_channels / torch.sum(self.mask.cpu(), dim=0, keepdim=True)) self.flops_scale = nonzero / (in_channels * out_channels) self.params_scale = self.flops_scale def forward(self, input): weight = self.weight * self.mask return F.conv2d(input, weight, self.bias, self.stride, self.padding, self.dilation, 1) class SymConv2d(_ConvNd): '''Conv2d with symmetric weight matrices''' def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros'): kernel_size = _pair(kernel_size) stride = _pair(stride) padding = _pair(padding) dilation = _pair(dilation) super(SymConv2d, self).__init__( in_channels, out_channels, kernel_size, stride, padding, dilation, False, _pair(0), groups, bias, padding_mode) def forward(self, input): weight = self.weight + self.weight.permute(1, 0, 2, 3) if self.padding_mode == 'circular': expanded_padding = ((self.padding[1] + 1) // 2, self.padding[1] // 2, (self.padding[0] + 1) // 2, self.padding[0] // 2) return F.conv2d(F.pad(input, expanded_padding, mode='circular'), weight, self.bias, self.stride, _pair(0), self.dilation, self.groups) return F.conv2d(input, weight, self.bias, self.stride, self.padding, self.dilation, self.groups) ########### Other OPs class Swish(nn.Module): """Swish activation function: x * sigmoid(x)""" def __init__(self): super(Swish, self).__init__() def forward(self, x): return x * torch.sigmoid(x) class SE(nn.Module): """Squeeze-and-Excitation (SE) block w/ Swish activation fun.""" def __init__(self, in_w, se_w, act_fun): super(SE, self).__init__() self._construct_class(in_w, se_w, act_fun) def _construct_class(self, in_w, se_w, act_fun): # AvgPool self.avg_pool = nn.AdaptiveAvgPool2d((1, 1)) # FC, Swish, FC, Sigmoid self.f_ex = nn.Sequential( nn.Conv2d(in_w, se_w, kernel_size=1, bias=True), act_fun(), nn.Conv2d(se_w, in_w, kernel_size=1, bias=True), nn.Sigmoid() ) def forward(self, x): return x * self.f_ex(self.avg_pool(x)) class SparseLinear(nn.Linear): '''Sparse Linear layer''' def __init__(self, group_num, in_scale, out_scale, bias=False, edge_index=None, flops_scale=0.5, params_scale=0.5): # mask is used for reset to zero mask_one = np.ones((out_scale, in_scale), dtype=bool) mask_zero = np.zeros((out_scale, in_scale), dtype=bool) mask_list = [[mask_one for i in range(group_num)] for j in range(group_num)] for i in range(edge_index.shape[0]): mask_list[edge_index[i, 0]][edge_index[i, 1]] = mask_zero self.mask = np.block(mask_list) self.edge_index = edge_index # todo: update to pytorch 1.2.0, then use bool() dtype self.mask = torch.from_numpy(self.mask).byte().cuda() self.flops_scale = flops_scale self.params_scale = params_scale super(SparseLinear, self).__init__( group_num * in_scale, group_num * out_scale, bias) def forward(self, x): weight = self.weight.clone().masked_fill_(self.mask, 0) # pdb.set_trace() return F.linear(x, weight, self.bias) class GroupLinear(nn.Module): '''Group conv style linear layer''' def __init__(self, in_channels, out_channels, bias=False, group_size=1): super(GroupLinear, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.group_size = group_size self.group_num = in_channels // group_size self.in_scale = in_channels // self.group_num self.out_scale = out_channels // self.group_num assert in_channels % self.group_num == 0 assert out_channels % self.group_num == 0 assert in_channels % self.group_size == 0 # Note: agg_fun is always sum self.edge_index = np.arange(self.group_num)[:, np.newaxis].repeat(2, axis=1) self.edge_num = self.edge_index.shape[0] flops_scale = self.edge_num / (self.group_num * self.group_num) params_scale = self.edge_num / (self.group_num * self.group_num) self.linear = SparseLinear(self.group_num, self.in_scale, self.out_scale, bias, edge_index=self.edge_index, flops_scale=flops_scale, params_scale=params_scale) def forward(self, x): x = self.linear(x) return x
15,045
35.877451
114
py
RobDanns
RobDanns-main/deep_learning/pycls/datasets/cifar100.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """CIFAR100 dataset.""" import numpy as np import os import pickle import torch import torch.utils.data import pycls.datasets.transforms as transforms from torchvision import datasets import pycls.utils.logging as lu logger = lu.get_logger(__name__) # Per-channel mean and SD values in BGR order _MEAN = [129.3, 124.1, 112.4] _SD = [68.2, 65.4, 70.4] class Cifar100(torch.utils.data.Dataset): """CIFAR-100 dataset.""" def __init__(self, data_path, split, batch_size): assert os.path.exists(data_path), \ 'Data path \'{}\' not found'.format(data_path) assert split in ['train', 'test'], \ 'Split \'{}\' not supported for cifar'.format(split) logger.info('Constructing CIFAR-100 {}...'.format(split)) self._data_path = data_path self._split = split self._batch_size = batch_size # Data format: # self._inputs - (split_size, 3, 32, 32) ndarray # self._labels - split_size list self._inputs, self._labels = self._load_data() def _load_batch(self, batch_path): with open(batch_path, 'rb') as f: d = pickle.load(f, encoding='bytes') return d[b'data'], d[b'fine_labels'] # return d[b'data'], d[b'labels'] def _load_data(self): """Loads data in memory.""" logger.info('{} data path: {}'.format(self._split, self._data_path)) # Compute data batch names if self._split == 'train': batch_names = ['train'] # datasets.CIFAR100(self._data_path, train=True) # batch_names = ['data_batch_{}'.format(i) for i in range(1, 6)] else: batch_names = ['test'] # Load data batches inputs, labels = [], [] for batch_name in batch_names: batch_path = os.path.join(self._data_path, batch_name) inputs_batch, labels_batch = self._load_batch(batch_path) inputs.append(inputs_batch) labels += labels_batch # Combine and reshape the inputs inputs = np.vstack(inputs).astype(np.float32) inputs = inputs.reshape((-1, 3, 32, 32)) return inputs, labels def _transform_image(self, image): """Transforms the image for network input.""" if self._batch_size != 1: image = transforms.color_normalization(image, _MEAN, _SD) if self._split == 'train': image = transforms.horizontal_flip(image=image, prob=0.5) image = transforms.random_crop(image=image, size=32, pad_size=4) return image def __getitem__(self, index): image, label = self._inputs[index, ...], self._labels[index] image = self._transform_image(image) return image, label def __len__(self): return self._inputs.shape[0]
3,163
34.155556
107
py
RobDanns
RobDanns-main/deep_learning/pycls/datasets/cifar10.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """CIFAR10 dataset.""" import numpy as np import os import pickle import torch import torch.utils.data import pycls.datasets.transforms as transforms import pycls.utils.logging as lu from pycls.config import cfg logger = lu.get_logger(__name__) # Per-channel mean and SD values in BGR order _MEAN = [125.3, 123.0, 113.9] _SD = [63.0, 62.1, 66.7] class Cifar10(torch.utils.data.Dataset): """CIFAR-10 dataset.""" def __init__(self, data_path, split, batch_size): assert os.path.exists(data_path), \ 'Data path \'{}\' not found'.format(data_path) assert split in ['train', 'test'], \ 'Split \'{}\' not supported for cifar'.format(split) logger.info('Constructing CIFAR-10 {}...'.format(split)) self._data_path = data_path self._split = split self._batch_size = batch_size # Data format: # self._inputs - (split_size, 3, 32, 32) ndarray # self._labels - split_size list self._inputs, self._labels = self._load_data() def _load_batch(self, batch_path): with open(batch_path, 'rb') as f: d = pickle.load(f, encoding='bytes') return d[b'data'], d[b'labels'] def _load_data(self): """Loads data in memory.""" logger.info('{} data path: {}'.format(self._split, self._data_path)) # Compute data batch names if self._split == 'train': batch_names = ['data_batch_{}'.format(i) for i in range(1, 6)] else: batch_names = ['test_batch'] # Load data batches inputs, labels = [], [] for batch_name in batch_names: batch_path = os.path.join(self._data_path, batch_name) inputs_batch, labels_batch = self._load_batch(batch_path) inputs.append(inputs_batch) labels += labels_batch # Combine and reshape the inputs inputs = np.vstack(inputs).astype(np.float32) inputs = inputs.reshape((-1, 3, 32, 32)) return inputs, labels def _transform_image(self, image): """Transforms the image for network input.""" if self._batch_size != 1: # Normalizing input images image = transforms.color_normalization(image, _MEAN, _SD) if self._split == 'train': image = transforms.horizontal_flip(image=image, prob=0.5) image = transforms.random_crop(image=image, size=32, pad_size=4) return image def __getitem__(self, index): image, label = self._inputs[index, ...], self._labels[index] image = self._transform_image(image) return image, label def __len__(self): return self._inputs.shape[0]
3,048
33.647727
107
py
RobDanns
RobDanns-main/deep_learning/pycls/datasets/loader.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Data loader.""" from torch.utils.data.distributed import DistributedSampler from torch.utils.data.sampler import RandomSampler import torch from pycls.config import cfg from pycls.datasets.cifar10 import Cifar10 from pycls.datasets.cifar100 import Cifar100 from pycls.datasets.tinyimagenet200 import TinyImageNet200 from pycls.datasets.imagenet import ImageNet import pycls.datasets.paths as dp # Supported datasets _DATASET_CATALOG = { 'cifar10': Cifar10, 'cifar100': Cifar100, 'tinyimagenet200': TinyImageNet200, 'imagenet': ImageNet } def _construct_loader(dataset_name, split, batch_size, shuffle, drop_last): """Constructs the data loader for the given dataset.""" assert dataset_name in _DATASET_CATALOG.keys(), \ 'Dataset \'{}\' not supported'.format(dataset_name) assert dp.has_data_path(dataset_name), \ 'Dataset \'{}\' has no data path'.format(dataset_name) # Retrieve the data path for the dataset data_path = dp.get_data_path(dataset_name) # Construct the dataset dataset = _DATASET_CATALOG[dataset_name](data_path, split, batch_size) # Create a sampler for multi-process training sampler = DistributedSampler(dataset) if cfg.NUM_GPUS > 1 else None # Create a loader loader = torch.utils.data.DataLoader( dataset, batch_size=batch_size, shuffle=(False if sampler else shuffle), sampler=sampler, num_workers=cfg.DATA_LOADER.NUM_WORKERS, pin_memory=cfg.DATA_LOADER.PIN_MEMORY, drop_last=drop_last ) return loader def construct_train_loader(): """Train loader wrapper.""" return _construct_loader( dataset_name=cfg.TRAIN.DATASET, split=cfg.TRAIN.SPLIT, batch_size=int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS), shuffle=True, drop_last=True ) def construct_test_loader(): """Test loader wrapper.""" return _construct_loader( dataset_name=cfg.TEST.DATASET, split=cfg.TEST.SPLIT, batch_size=int(cfg.TEST.BATCH_SIZE / cfg.NUM_GPUS), shuffle=False, drop_last=False ) def construct_test_loader_adv(): """Test loader wrapper.""" return _construct_loader( dataset_name=cfg.TEST.DATASET, split=cfg.TEST.SPLIT, batch_size=1, shuffle=False, drop_last=False ) def shuffle(loader, cur_epoch): """"Shuffles the data.""" assert isinstance(loader.sampler, (RandomSampler, DistributedSampler)), \ 'Sampler type \'{}\' not supported'.format(type(loader.sampler)) # RandomSampler handles shuffling automatically if isinstance(loader.sampler, DistributedSampler): # DistributedSampler shuffles data based on epoch loader.sampler.set_epoch(cur_epoch)
3,131
30.009901
107
py
RobDanns
RobDanns-main/deep_learning/pycls/datasets/imagenet.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """ImageNet dataset.""" import cv2 import numpy as np import os import torch import torch.utils.data import pycls.datasets.transforms as transforms import pycls.utils.logging as lu logger = lu.get_logger(__name__) # Per-channel mean and SD values in BGR order _MEAN = [0.406, 0.456, 0.485] _SD = [0.225, 0.224, 0.229] # Eig vals and vecs of the cov mat _EIG_VALS = [0.2175, 0.0188, 0.0045] _EIG_VECS = np.array([ [-0.5675, 0.7192, 0.4009], [-0.5808, -0.0045, -0.8140], [-0.5836, -0.6948, 0.4203] ]) class ImageNet(torch.utils.data.Dataset): """ImageNet dataset.""" def __init__(self, data_path, split, batch_size): assert os.path.exists(data_path), \ 'Data path \'{}\' not found'.format(data_path) assert split in ['train', 'val'], \ 'Split \'{}\' not supported for ImageNet'.format(split) logger.info('Constructing ImageNet {}...'.format(split)) self._data_path = data_path self._split = split self._batch_size = batch_size self._construct_imdb() def _construct_imdb(self): """Constructs the imdb.""" # Compile the split data path split_path = os.path.join(self._data_path, self._split) logger.info('{} data path: {}'.format(self._split, split_path)) # Map ImageNet class ids to contiguous ids self._class_ids = os.listdir(split_path) self._class_id_cont_id = {v: i for i, v in enumerate(self._class_ids)} # Construct the image db self._imdb = [] for class_id in self._class_ids: cont_id = self._class_id_cont_id[class_id] im_dir = os.path.join(split_path, class_id) for im_name in os.listdir(im_dir): self._imdb.append({ 'im_path': os.path.join(im_dir, im_name), 'class': cont_id, }) logger.info('Number of images: {}'.format(len(self._imdb))) logger.info('Number of classes: {}'.format(len(self._class_ids))) def _prepare_im(self, im): """Prepares the image for network input.""" # Train and test setups differ if self._split == 'train': # Scale and aspect ratio im = transforms.random_sized_crop( image=im, size=224, area_frac=0.08 ) # Horizontal flip im = transforms.horizontal_flip(image=im, prob=0.5, order='HWC') else: # Scale and center crop im = transforms.scale(256, im) im = transforms.center_crop(224, im) # HWC -> CHW im = transforms.HWC2CHW(im) # [0, 255] -> [0, 1] im = im / 255.0 # PCA jitter if self._split == 'train': im = transforms.lighting(im, 0.1, _EIG_VALS, _EIG_VECS) # Color normalization if self._batch_size != 1: im = transforms.color_normalization(im, _MEAN, _SD) return im def __getitem__(self, index): # Load the image im = cv2.imread(self._imdb[index]['im_path']) im = im.astype(np.float32, copy=False) # Prepare the image for training / testing im = self._prepare_im(im) # Retrieve the label label = self._imdb[index]['class'] return im, label def __len__(self): return len(self._imdb) # class ImageNet(torch.utils.data.Dataset): # """ImageNet dataset.""" # def __init__(self, data_path, split): # assert os.path.exists(data_path), \ # 'Data path \'{}\' not found'.format(data_path) # assert split in ['train', 'val'], \ # 'Split \'{}\' not supported for ImageNet'.format(split) # logger.info('Constructing ImageNet {}...'.format(split)) # self._data_path = data_path # self._split = split # self._construct_imdb() # def _construct_imdb(self): # """Constructs the imdb.""" # # Compile the split data path # split_path = os.path.join(self._data_path, self._split) # logger.info('{} data path: {}'.format(self._split, split_path)) # # Map ImageNet class ids to contiguous ids # self._class_ids = os.listdir(split_path) # self._class_id_cont_id = {v: i for i, v in enumerate(self._class_ids)} # # Construct the image db # self._imdb = [] # counter = 1 # for class_id in self._class_ids: # print('progress: {}/{}'.format(counter,len(self._class_ids))) # counter += 1 # cont_id = self._class_id_cont_id[class_id] # im_dir = os.path.join(split_path, class_id) # for im_name in os.listdir(im_dir): # self._imdb.append({ # 'im_path': os.path.join(im_dir, im_name), # 'class': cont_id, # 'img': cv2.imread(os.path.join(im_dir, im_name)).astype(np.float32, copy=False) # }) # logger.info('Number of images: {}'.format(len(self._imdb))) # logger.info('Number of classes: {}'.format(len(self._class_ids))) # def _prepare_im(self, im): # """Prepares the image for network input.""" # # Train and test setups differ # if self._split == 'train': # # Scale and aspect ratio # im = transforms.random_sized_crop( # image=im, size=224, area_frac=0.08 # ) # # Horizontal flip # im = transforms.horizontal_flip(image=im, prob=0.5, order='HWC') # else: # # Scale and center crop # im = transforms.scale(256, im) # im = transforms.center_crop(224, im) # # HWC -> CHW # im = transforms.HWC2CHW(im) # # [0, 255] -> [0, 1] # im = im / 255.0 # # PCA jitter # if self._split == 'train': # im = transforms.lighting(im, 0.1, _EIG_VALS, _EIG_VECS) # # Color normalization # im = transforms.color_normalization(im, _MEAN, _SD) # return im # def __getitem__(self, index): # # Load the image # im = self._imdb[index]['img'] # # Prepare the image for training / testing # im = self._prepare_im(im) # # Retrieve the label # label = self._imdb[index]['class'] # return im, label # def __len__(self): # return len(self._imdb)
6,759
35.344086
107
py
RobDanns
RobDanns-main/deep_learning/pycls/utils/checkpoint.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Functions that handle saving and loading of checkpoints.""" import os import torch from collections import OrderedDict from pycls.config import cfg import pycls.utils.distributed as du # Common prefix for checkpoint file names _NAME_PREFIX = 'model_epoch_' # Checkpoints directory name _DIR_NAME = 'checkpoints' def get_checkpoint_dir(): """Get location for storing checkpoints.""" return os.path.join(cfg.OUT_DIR, _DIR_NAME) def got_checkpoint_dir(): """Get location for storing checkpoints for inference time.""" return os.path.join(cfg.CHECKPT_DIR, _DIR_NAME) def get_checkpoint(epoch): """Get the full path to a checkpoint file.""" name = '{}{:04d}.pyth'.format(_NAME_PREFIX, epoch) return os.path.join(get_checkpoint_dir(), name) def got_checkpoint(epoch): """Get the full path to a checkpoint file for inference time.""" name = '{}{:04d}.pyth'.format(_NAME_PREFIX, epoch) return os.path.join(got_checkpoint_dir(), name) def get_checkpoint_last(): d = get_checkpoint_dir() names = os.listdir(d) if os.path.exists(d) else [] names = [f for f in names if _NAME_PREFIX in f] assert len(names), 'No checkpoints found in \'{}\'.'.format(d) name = sorted(names)[-1] return os.path.join(d, name) def got_checkpoint_last(): d = got_checkpoint_dir() names = os.listdir(d) if os.path.exists(d) else [] names = [f for f in names if _NAME_PREFIX in f] assert len(names), 'No checkpoints found in \'{}\'.'.format(d) name = sorted(names)[-1] return os.path.join(d, name) def has_checkpoint(): """Determines if the given directory contains a checkpoint.""" d = get_checkpoint_dir() print("checkpoint directory =", d) files = os.listdir(d) if os.path.exists(d) else [] return any(_NAME_PREFIX in f for f in files) def had_checkpoint(): """Determines if the given directory contains a checkpoint for inference time.""" d = got_checkpoint_dir() print("checkpoint directory =", d) files = os.listdir(d) if os.path.exists(d) else [] return any(_NAME_PREFIX in f for f in files) def is_checkpoint_epoch(cur_epoch): """Determines if a checkpoint should be saved on current epoch.""" return (cur_epoch + 1) % cfg.TRAIN.CHECKPOINT_PERIOD == 0 def save_checkpoint(model, optimizer, epoch): """Saves a checkpoint.""" # Save checkpoints only from the master process if not du.is_master_proc(): return os.makedirs(get_checkpoint_dir(), exist_ok=True) checkpoint = { 'epoch': epoch, 'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict(), 'cfg': cfg.dump() } checkpoint_file = get_checkpoint(epoch + 1) torch.save(checkpoint, checkpoint_file) return checkpoint_file def load_checkpoint(checkpoint_file, model, optimizer=None): """Loads the checkpoint from the given file.""" assert os.path.exists(checkpoint_file), \ 'Checkpoint \'{}\' not found'.format(checkpoint_file) # if cfg.IS_INFERENCE and cfg.IS_DDP: # state_dict = torch.load(checkpoint_file, map_location='cpu') # new_state_dict = OrderedDict() # print("state_dict.items() :", state_dict) # for k, v in state_dict.items(): # name = k[7:] # remove `module.` # new_state_dict[name] = v # # load params # epoch = state_dict['epoch'] # model.load_state_dict(new_state_dict['model_state']) # if optimizer: # optimizer.load_state_dict(new_state_dict['optimizer_state']) if cfg.IS_INFERENCE: print("Mapping model to CPU") checkpoint = torch.load(checkpoint_file, map_location='cpu') # print(checkpoint) else: checkpoint = torch.load(checkpoint_file) epoch = checkpoint['epoch'] print("Epochs from checkpoint = ", epoch) model.load_state_dict(checkpoint['model_state'], strict=False) if optimizer: optimizer.load_state_dict(checkpoint['optimizer_state']) return epoch
4,392
31.540741
107
py
RobDanns
RobDanns-main/deep_learning/pycls/utils/net.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Functions for manipulating networks.""" import itertools import math import torch import torch.nn as nn from pycls.config import cfg from ..models.relation_graph import * def init_weights(m): """Performs ResNet style weight initialization.""" if isinstance(m, nn.Conv2d) or isinstance(m, SymConv2d): # Note that there is no bias due to BN fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(mean=0.0, std=math.sqrt(2.0 / fan_out)) elif isinstance(m, TalkConv2d): # Note that there is no bias due to BN ### uniform init fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels * m.params_scale ### node specific init # fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(mean=0.0, std=math.sqrt(2.0 / fan_out)) # m.weight.data = m.weight.data*m.init_scale elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.BatchNorm1d): zero_init_gamma = ( hasattr(m, 'final_bn') and m.final_bn and cfg.BN.ZERO_INIT_FINAL_GAMMA ) m.weight.data.fill_(0.0 if zero_init_gamma else 1.0) m.bias.data.zero_() elif isinstance(m, nn.Linear) or isinstance(m, TalkLinear) or isinstance(m, SymLinear): m.weight.data.normal_(mean=0.0, std=0.01) if m.bias is not None: m.bias.data.zero_() @torch.no_grad() def compute_precise_bn_stats(model, loader): """Computes precise BN stats on training data.""" # Compute the number of minibatches to use num_iter = min(cfg.BN.NUM_SAMPLES_PRECISE // loader.batch_size, len(loader)) # Retrieve the BN layers bns = [m for m in model.modules() if isinstance(m, torch.nn.BatchNorm2d)] # Initialize stats storage mus = [torch.zeros_like(bn.running_mean) for bn in bns] sqs = [torch.zeros_like(bn.running_var) for bn in bns] # Remember momentum values moms = [bn.momentum for bn in bns] # Disable momentum for bn in bns: bn.momentum = 1.0 # Accumulate the stats across the data samples for inputs, _labels in itertools.islice(loader, num_iter): model(inputs.cuda()) # Accumulate the stats for each BN layer for i, bn in enumerate(bns): m, v = bn.running_mean, bn.running_var sqs[i] += (v + m * m) / num_iter mus[i] += m / num_iter # Set the stats and restore momentum values for i, bn in enumerate(bns): bn.running_var = sqs[i] - mus[i] * mus[i] bn.running_mean = mus[i] bn.momentum = moms[i] def get_flat_weights(model): """Gets all model weights as a single flat vector.""" return torch.cat([p.data.view(-1, 1) for p in model.parameters()], 0) def set_flat_weights(model, flat_weights): """Sets all model weights from a single flat vector.""" k = 0 for p in model.parameters(): n = p.data.numel() p.data.copy_(flat_weights[k:(k + n)].view_as(p.data)) k += n assert k == flat_weights.numel() def model2adj(model): adj_dict = {} i = 0 for n, m in model.named_modules(): if isinstance(m, nn.Linear) or isinstance(m, nn.Conv2d): adj_dict['weight_{}'.format(i)] = m.weight.data.squeeze().cpu().numpy() i += 1 elif isinstance(m, SymLinear): weight = m.weight.data + m.weight.data.permute(1, 0) adj_dict['weight_{}'.format(i)] = weight.squeeze().cpu().numpy() i += 1 elif isinstance(m, SymConv2d): weight = m.weight.data + m.weight.data.permute(1, 0, 2, 3) adj_dict['weight_{}'.format(i)] = weight.squeeze().cpu().numpy() i += 1 elif isinstance(m, TalkLinear) or isinstance(m, TalkConv2d): adj_dict['weight_{}'.format(i)] = m.weight.data.squeeze().cpu().numpy() adj_dict['mask_{}'.format(i)] = m.mask.data.squeeze().cpu().numpy() i += 1 return adj_dict
4,360
37.59292
107
py
RobDanns
RobDanns-main/deep_learning/pycls/utils/distributed.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Distributed helpers.""" import torch from pycls.config import cfg def is_master_proc(): """Determines if the current process is the master process. Master process is responsible for logging, writing and loading checkpoints. In the multi GPU setting, we assign the master role to the rank 0 process. When training using a single GPU, there is only one training processes which is considered the master processes. """ return cfg.NUM_GPUS == 1 or torch.distributed.get_rank() == 0 def init_process_group(proc_rank, world_size): """Initializes the default process group.""" # Set the GPU to use torch.cuda.set_device(proc_rank) # Initialize the process group # print('--rank{},world{}--'.format(proc_rank, world_size)) # torch.distributed.init_process_group( # backend=cfg.DIST_BACKEND, # init_method="tcp://{}:{}".format(cfg.HOST, cfg.PORT), # world_size=world_size, # rank=proc_rank # ) torch.distributed.init_process_group( backend=cfg.DIST_BACKEND, init_method='env://', world_size=world_size, rank=proc_rank ) def destroy_process_group(): """Destroys the default process group.""" torch.distributed.destroy_process_group() def scaled_all_reduce(tensors): """Performs the scaled all_reduce operation on the provided tensors. The input tensors are modified in-place. Currently supports only the sum reduction operator. The reduced values are scaled by the inverse size of the process group (equivalent to cfg.NUM_GPUS). """ # Queue the reductions reductions = [] for tensor in tensors: reduction = torch.distributed.all_reduce(tensor, async_op=True) reductions.append(reduction) # Wait for reductions to finish for reduction in reductions: reduction.wait() # Scale the results for tensor in tensors: tensor.mul_(1.0 / cfg.NUM_GPUS) return tensors
2,323
31.277778
107
py
RobDanns
RobDanns-main/deep_learning/pycls/utils/metrics.py
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the original graph2nn github repo. # File modifications and additions by Rowan AI Lab, licensed under the Creative Commons Zero v1.0 Universal # LICENSE file in the root directory of this source tree. """Functions for computing metrics.""" import numpy as np import torch import torch.nn as nn import pdb from pycls.config import cfg from functools import reduce import operator from ..models.relation_graph import * # Number of bytes in a megabyte _B_IN_MB = 1024 * 1024 def topks_correct(preds, labels, ks): """Computes the number of top-k correct predictions for each k.""" assert preds.size(0) == labels.size(0), \ 'Batch dim of predictions and labels must match' # Find the top max_k predictions for each sample _top_max_k_vals, top_max_k_inds = torch.topk( preds, max(ks), dim=1, largest=True, sorted=True ) # (batch_size, max_k) -> (max_k, batch_size) top_max_k_inds = top_max_k_inds.t() # (batch_size, ) -> (max_k, batch_size) rep_max_k_labels = labels.view(1, -1).expand_as(top_max_k_inds) # (i, j) = 1 if top i-th prediction for the j-th sample is correct top_max_k_correct = top_max_k_inds.eq(rep_max_k_labels) # Compute the number of topk correct predictions for each k topks_correct = [ top_max_k_correct[:k, :].view(-1).float().sum() for k in ks ] return topks_correct def topk_errors(preds, labels, ks): """Computes the top-k error for each k.""" num_topks_correct = topks_correct(preds, labels, ks) return [(1.0 - x / preds.size(0)) * 100.0 for x in num_topks_correct] def topk_accuracies(preds, labels, ks): """Computes the top-k accuracy for each k.""" num_topks_correct = topks_correct(preds, labels, ks) return [(x / preds.size(0)) * 100.0 for x in num_topks_correct] def params_count(model): """Computes the number of parameters.""" count = 0 for n,m in model.named_modules(): if isinstance(m, TalkConv2d) or isinstance(m, TalkLinear): count += np.sum([p.numel()*m.params_scale for p in m.parameters(recurse=False)]).item() else: count += np.sum([p.numel() for p in m.parameters(recurse=False)]).item() return int(count) def flops_count(model): """Computes the number of flops.""" assert cfg.TRAIN.DATASET in ['cifar10', 'cifar100', 'tinyimagenet200', 'imagenet'], \ 'Computing flops for {} is not supported'.format(cfg.TRAIN.DATASET) # im_size = 32 if cfg.TRAIN.DATASET == 'cifar10' else 224 if cfg.TRAIN.DATASET == 'cifar10': im_size = 32 elif cfg.TRAIN.DATASET == 'cifar100': im_size = 32 elif cfg.TRAIN.DATASET == 'tinyimagenet200': im_size = 64 else: im_size = 224 h, w = im_size, im_size count = 0 for n, m in model.named_modules(): if isinstance(m, nn.Conv2d): if '.se' in n: count += m.in_channels * m.out_channels + m.bias.numel() continue h_out = (h + 2 * m.padding[0] - m.kernel_size[0]) // m.stride[0] + 1 w_out = (w + 2 * m.padding[1] - m.kernel_size[1]) // m.stride[1] + 1 count += np.prod([ m.weight.numel(), h_out, w_out ]) if 'proj' not in n: h, w = h_out, w_out elif isinstance(m, TalkConv2d): h_out = (h + 2 * m.padding[0] - m.kernel_size[0]) // m.stride[0] + 1 w_out = (w + 2 * m.padding[1] - m.kernel_size[1]) // m.stride[1] + 1 count += int(np.prod([ m.weight.numel()*m.flops_scale, h_out, w_out ])) if 'proj' not in n and 'pool' not in n: h, w = h_out, w_out elif isinstance(m, nn.MaxPool2d): h = (h + 2 * m.padding - m.kernel_size) // m.stride + 1 w = (w + 2 * m.padding - m.kernel_size) // m.stride + 1 elif isinstance(m, TalkLinear): count += int(m.in_features * m.out_features * m.flops_scale) elif isinstance(m, nn.Linear): count += m.in_features * m.out_features return count def gpu_mem_usage(): """Computes the GPU memory usage for the current device (MB).""" mem_usage_bytes = torch.cuda.max_memory_allocated() return mem_usage_bytes / _B_IN_MB # Online FLOPs/Params calculation from CondenseNet codebase count_ops = 0 count_params = 0 def get_num_gen(gen): return sum(1 for x in gen) def is_pruned(layer): try: layer.mask return True except AttributeError: return False def is_leaf(model): return get_num_gen(model.children()) == 0 def get_layer_info(layer): layer_str = str(layer) type_name = layer_str[:layer_str.find('(')].strip() return type_name def get_layer_param(model): return sum([reduce(operator.mul, i.size(), 1) for i in model.parameters()]) ### The input batch size should be 1 to call this function def measure_layer(layer, x): global count_ops, count_params delta_ops = 0 delta_params = 0 multi_add = 1 type_name = get_layer_info(layer) ### ops_conv if type_name in ['Conv2d']: out_h = int((x.size()[2] + 2 * layer.padding[0] - layer.kernel_size[0]) / layer.stride[0] + 1) out_w = int((x.size()[3] + 2 * layer.padding[1] - layer.kernel_size[1]) / layer.stride[1] + 1) delta_ops = layer.in_channels * layer.out_channels * layer.kernel_size[0] * \ layer.kernel_size[1] * out_h * out_w / layer.groups * multi_add print(layer) print('out_h: ', out_h, 'out_w:', out_w) delta_params = get_layer_param(layer) ### ops_nonlinearity elif type_name in ['ReLU']: delta_ops = x.numel() delta_params = get_layer_param(layer) ### ops_pooling elif type_name in ['AvgPool2d', 'MaxPool2d']: in_w = x.size()[2] kernel_ops = layer.kernel_size * layer.kernel_size out_w = int((in_w + 2 * layer.padding - layer.kernel_size) / layer.stride + 1) out_h = int((in_w + 2 * layer.padding - layer.kernel_size) / layer.stride + 1) delta_ops = x.size()[0] * x.size()[1] * out_w * out_h * kernel_ops delta_params = get_layer_param(layer) elif type_name in ['AdaptiveAvgPool2d']: delta_ops = x.size()[0] * x.size()[1] * x.size()[2] * x.size()[3] delta_params = get_layer_param(layer) ### ops_linear elif type_name in ['Linear']: weight_ops = layer.weight.numel() * multi_add bias_ops = layer.bias.numel() delta_ops = x.size()[0] * (weight_ops + bias_ops) delta_params = get_layer_param(layer) elif type_name in ['WeightedSumTransform']: weight_ops = layer.weight.numel() * multi_add delta_ops = x.size()[0] * (weight_ops) delta_params = get_layer_param(layer) ### ops_nothing elif type_name in ['BatchNorm2d', 'Dropout2d', 'DropChannel', 'Dropout', 'Sigmoid', 'DirichletWeightedSumTransform', 'Softmax', 'Identity', 'Sequential']: delta_params = get_layer_param(layer) ### unknown layer type else: raise TypeError('unknown layer type: %s' % type_name) count_ops += delta_ops count_params += delta_params return def measure_model(model, H, W): global count_ops, count_params count_ops = 0 count_params = 0 data = torch.zeros(1, 3, H, W).cuda() def should_measure(x): return is_leaf(x) or is_pruned(x) def modify_forward(model): for child in model.children(): if should_measure(child): def new_forward(m): def lambda_forward(x): measure_layer(m, x) return m.old_forward(x) return lambda_forward child.old_forward = child.forward child.forward = new_forward(child) else: modify_forward(child) def restore_forward(model): for child in model.children(): # leaf node if is_leaf(child) and hasattr(child, 'old_forward'): child.forward = child.old_forward child.old_forward = None else: restore_forward(child) modify_forward(model) model.forward(data) restore_forward(model) return count_ops, count_params
8,557
33.095618
158
py
kge_ecotox_regression
kge_ecotox_regression-main/main.py
""" TODO: - Train embedding model. - Apply embeddings to data. - Encode data. - Train,valid,test model """ from autoencoder import create_auto_encoder from model import create_model, CorrelelatedFeatures, ApproxKerasSVM, coeff_determination import numpy as np import pandas as pd from sklearn.model_selection import train_test_split, GridSearchCV, RandomizedSearchCV from sklearn.preprocessing import OneHotEncoder from tensorflow.keras.callbacks import EarlyStopping from sklearn.preprocessing import MinMaxScaler from sklearn.model_selection import KFold from random import shuffle from collections import defaultdict import tensorflow as tf from sklearn.svm import SVR from sklearn.neural_network import MLPRegressor from sklearn.preprocessing import StandardScaler from sklearn.linear_model import LogisticRegression, LinearRegression, HuberRegressor, BayesianRidge from sklearn.ensemble import RandomForestRegressor, AdaBoostRegressor, VotingRegressor, BaggingRegressor, ExtraTreesRegressor, GradientBoostingRegressor from sklearn.compose import TransformedTargetRegressor from sklearn.preprocessing import QuantileTransformer, RobustScaler from sklearn.tree import DecisionTreeRegressor from itertools import product from random import choice, choices from sklearn.pipeline import Pipeline from tqdm import tqdm from matplotlib import pyplot as plt from sklearn.decomposition import PCA,FastICA from sklearn.cluster import FeatureAgglomeration from sklearn.feature_selection import RFE from sklearn.metrics import r2_score from sklearn.isotonic import IsotonicRegression from sklearn.feature_selection import VarianceThreshold from sklearn.dummy import DummyRegressor from sklearn.experimental import enable_hist_gradient_boosting # noqa from sklearn.ensemble import HistGradientBoostingRegressor from sklearn.neighbors import KNeighborsRegressor from sklearn.model_selection import cross_val_score, LeaveOneOut MAX_ENCODER_EPOCHS = 1000 MAX_EPOCHS = 1000 EPSILON = 1e-10 MODEL = 'ComplEx' hidden_dim = (128,) SEED = 42 np.random.seed(SEED) import tensorflow as tf tf.get_logger().setLevel('ERROR') import warnings warnings.filterwarnings('ignore') def load_fingerprints(filename): df = pd.read_csv(filename,index_col='chemical') l = len(df.iloc[0]['fingerprint']) out = {} for c in df.index: fp = df.loc[c]['fingerprint'] v = [int(f) for f in fp] out[c] = np.asarray(v) return out def load_features(filename): df = pd.read_csv(filename,index_col='chemical') df = df.dropna() columns = df.columns out = {} for c in df.index: v = [df.loc[c][col] for col in columns] out[c] = np.asarray(v) return out def load_one_hot(entities): all_entities = list(set(entities)) out = {} for e in entities: v = np.zeros((len(all_entities),)) v[all_entities.index(e)] = 1 out[e] = np.asarray(v) return out def load_embeddings(filename,filename_ids): df = np.load(filename) ids = dict(np.load(filename_ids)) return {k:df[int(ids[k])] for k in ids} def load_data(filename,filter_chemicals=None, filter_species=None): df = pd.read_csv(filename) X,y = [],[] if filter_chemicals: to_drop = set(df.chemical) - filter_chemicals for c in to_drop: df = df.drop(df[df.chemical == c].index) if filter_species: to_drop = set(df.species) - filter_species for s in to_drop: df = df.drop(df[df.species == s].index) df = df.drop(df[df.study_duration > 24*14].index) df = df.groupby(['chemical','species'],as_index=False).mean() X = list(zip(df['chemical'],df['species'])) y = np.log(df.concentration+EPSILON) tmp = np.asarray(df.study_duration).reshape((-1,1)) mms = StandardScaler() tmp = mms.fit_transform(tmp) experimental_features = dict(zip(X,tmp.reshape(-1,1))) y = np.asarray(y).reshape((-1,1)) #y = MinMaxScaler().fit_transform(y) return X, y, experimental_features def data_split(X,Y,restrictions=None,method = 1, variant = 1, prop=0.33): """ C_x - chemical set S_x - species set t,v - training,validation 1. C_t \cap C_v == Ø and S_t \cap S_v != Ø, 2. C_t \cap C_v == Ø and S_t \cap S_v == Ø, 3. C_t \cap C_v != Ø and S_t \cap S_v != Ø, 4. C_t \cap C_v != Ø and S_t \cap S_v == Ø, Variants where C_t \cap C_v != Ø (same for S_x): 1. C_t == C_v 2. |C_t \cap C_v| < |C_t \cup C_v| Restrictions: Retriction of a set. eg. s_1 \in S_v and |S_v|=1, {'S_v':{'content:[s_1],'max_len',1}} """ C_t,C_v,S_t,S_v=map(set,[[]]*4) restrictions = {**{'C_t':{},'C_v':{},'S_t':{},'S_v':{}},**restrictions} def filter_restrictions(C_t,C_v,S_t,S_v): for _set,_inv_set,k in zip([C_t,C_v,S_t,S_v],[C_v,C_t,S_v,S_t],['C_t','C_v','S_t','S_v']): if k in restrictions: if 'content' in restrictions[k]: _set |= restrictions[k]['content'] if 'not content' in restrictions[k]: _set -= restrictions[k]['not content'] if 'max_len' in restrictions[k]: while restrictions[k]['max_len'] < len(_set): entity = choice(list(_set)) if not ('content' in restrictions[k] and entity in restrictions[k]['content']): _set.remove(entity) return C_t,C_v,S_t,S_v def check_restrictions(C_t,C_v,S_t,S_v): for _set,k,inv_k in zip([C_t,C_v,S_t,S_v],['C_t','C_v','S_t','S_v'],['C_v','C_t','S_v','S_t']): if k in restrictions: if 'content' in restrictions[k] and 'not content' in restrictions[k]: try: assert len(restrictions[k]['content'].intersection(restrictions[k]['not content'])) < 1 except AssertionError: raise AssertionError('Set %s content conflict.' % k) if 'content' in restrictions[k] and 'max_len' in restrictions[k]: try: assert len(restrictions[k]['content']) <= restrictions[k]['max_len'] except AssertionError: raise AssertionError('Set %s content is longer than max length' % k) if ((method == 1 and 'C' in k) or (method == 4 and 'S' in k) or method == 2) and 'content' in restrictions[inv_k]: try: assert restrictions[k]['content'].intersection(restrictions[inv_k]['content']) == set() except AssertionError: raise AssertionError('Intersection in %s content is not allowed in method %s.' % ('chemical' if method==1 else 'species',str(method))) if method == 3 and 'content' in restrictions[inv_k]: try: assert restrictions[k]['content'].intersection(restrictions[inv_k]['content']) == set() except AssertionError: raise AssertionError('Intersection in set content is not allowed in method 3.') C,S = map(set,zip(*X)) if method == 1: C_t,C_v = train_test_split(list(C),test_size=prop) if variant == 1: S_t,S_v = S, S else: S_t = choices(list(S),k=int((1-prop)*len(S))) S_v = choices(list(S),k=int(prop*len(S))) if method == 2: S_t,S_v = train_test_split(list(S),test_size=prop) C_t,C_v = train_test_split(list(C),test_size=prop) if method == 3: X_t, X_v = train_test_split(X,test_size=prop) C_t,S_t = map(set,zip(*X_t)) C_v,S_v = map(set,zip(*X_v)) if method == 4: S_t,S_v = train_test_split(list(S),test_size=prop) if variant == 1: C_t,C_v = C, C else: C_t = choices(list(C),k=int((1-prop)*len(C))) C_v = choices(list(C),k=int(prop*len(C))) C_t,C_v,S_t,S_v = map(set,[C_t,C_v,S_t,S_v]) C_t,C_v,S_t,S_v = filter_restrictions(C_t,C_v,S_t,S_v) if method == 1: C_t -= C_v if method == 2: C_t -= C_v S_t -= S_v if method == 4: S_t -= S_v if method == 1: assert C_t.intersection(C_v) == set() if variant == 1: S_t = S_v assert S_t == S_v else: assert len(S_t.intersection(S_v)) < len(S_t.union(S_v)) if method == 2: assert C_t.intersection(C_v) == set() and S_t.intersection(S_v) == set() if method == 3: assert len(C_t.intersection(C_v)) > 0 and len(S_t.intersection(S_v)) > 0 if method == 4: assert S_t.intersection(S_v) == set() if variant == 1: C_t = C_v assert C_t == C_v else: assert len(C_t.intersection(C_v)) < len(C_t.union(C_v)) check_restrictions(C_t,C_v,S_t,S_v) Xtr = [] Xte = [] ytr = [] yte = [] for x,y in zip(X,Y): c,s = x if c in C_t and s in S_t: Xtr.append(x) ytr.append(y) if c in C_v and s in S_v: Xte.append(x) yte.append(y) return Xtr,Xte,ytr,yte class FilterFingerprints: def __init__(self): pass def fit(self,X): idx = [] for i,a in enumerate(X.T): if len(np.unique(a)) > 1: idx.append(i) self.idx = idx def transform(self,X): if len(X.shape) > 1: return X[:,self.idx] else: return X[self.idx] def fit_transform(self,X): self.fit(X) return self.transform(X) def compile_model(model): model.compile(optimizer='adagrad',loss='log_cosh',metrics=['mae','mse',R2(name='r2')]) import math def lcm(a, b): return abs(a*b) // math.gcd(a, b) def combine(Xs): n = map(len,Xs) l = max(*map(lambda x: lcm(len(x[0]),len(x[1])),product(Xs,Xs))) r = [l//a for a in n] tmp = [] for X,a in zip(Xs,r): tmp.append(np.repeat(X,a,axis=0)) return np.concatenate(tmp,axis=1) def list_duplicates(seq): tally = defaultdict(list) for i,item in enumerate(seq): tally[item].append(i) return ((key,locs) for key,locs in tally.items() if len(locs)>1) def run_model(C_t,C_v,S_t,S_v,y, experimental_features, fingerprints, chemical_embedding, species_embedding, chemical_features, merge_species=False): """ Take four classes of chemicals, two pairs of siblings, test these on one-two species, combine siblings, combine cusins, see performance drop. Repeat on species side. Repeat with embeddings for chemicals and species and see the same performance on lower levels, but imporved over baseline on higher levels. """ """ 5-fold validation + 1-fold test set """ keys = set(y.keys()) keys_t = keys.intersection(set(product(C_t,S_t))) keys_v = keys.intersection(set(product(C_v,S_v))) ytr,yte = map(lambda x:np.asarray([y[i] for i in x]),[keys_t,keys_v]) if len(yte) < 1 or len(ytr) < 1: return None,None,None fingerprints_train,fingerprints_test = map(lambda x:np.asarray([fingerprints[i] for i,_ in x]),[keys_t,keys_v]) chemical_embedding_train,chemical_embedding_test = map(lambda x:np.asarray([chemical_embedding[i] for i,_ in x]),[keys_t,keys_v]) chemical_features_train,chemical_features_test = map(lambda x:np.asarray([chemical_features[i] for i,_ in x]),[keys_t,keys_v]) species_embedding_train,species_embedding_test = map(lambda x:np.asarray([species_embedding[i] for _,i in x]),[keys_t,keys_v]) experimental_features_train,experimental_features_test = map(lambda x:np.asarray([experimental_features[i] for i in x]),[keys_t,keys_v]) species_one_hot_encoder = OneHotEncoder(sparse=False) sp_t = set(list(zip(*keys_t))[1]) sp_v = set(list(zip(*keys_v))[1]) sp = np.asarray(list(sp_t|sp_v)).reshape((-1,1)) species_one_hot_encoder.fit(sp) species_one_hot_train,species_one_hot_test = map(lambda x:species_one_hot_encoder.transform(np.asarray(list(zip(*x))[1]).reshape((-1,1))),[keys_t,keys_v]) if merge_species: for array in [species_embedding_train,species_one_hot_train,ytr]: for elem,loc in list_duplicates([c for c,_ in keys_t]): #i.e. mean where c is the same array[loc] = np.mean(array[loc]) for array in [species_embedding_test,species_one_hot_test,yte]: for elem,loc in list_duplicates([c for c,_ in keys_v]): array[loc] = np.mean(array[loc]) n_tr = ytr.shape[1] n_te = yte.shape[1] train_1 = combine([fingerprints_train,chemical_features_train,species_one_hot_train,experimental_features_train,ytr]) train_2 = combine([fingerprints_train,chemical_features_train,species_embedding_train,chemical_embedding_train,experimental_features_train,ytr]) test_1 = combine([fingerprints_test,chemical_features_test,species_one_hot_test,experimental_features_test,yte]) test_2 = combine([fingerprints_test,chemical_features_test,species_embedding_test,chemical_embedding_test,experimental_features_test,yte]) Xtr_1,ytr = train_1[:,:-n_tr],train_1[:,-n_tr:] Xtr_2,ytr = train_2[:,:-n_tr],train_2[:,-n_tr:] Xte_1,yte = test_1[:,:-n_te],test_1[:,-n_te:] Xte_2,yte = test_2[:,:-n_te],test_2[:,-n_te:] res1 = np.zeros(yte.ravel().shape) res2 = np.zeros(yte.ravel().shape) params = {'n_neighbors':[2,5,10,25,50,100], 'weights':['uniform','distance']} n = min(len(ytr),5) FOLDS = 10 for Xtr,Xte,res in zip([Xtr_1,Xtr_2],[Xte_1,Xte_2],[res1,res2]): for _ in range(FOLDS): regr = AdaBoostRegressor(n_estimators=10,loss='square') regr.fit(Xtr,ytr.ravel()) res += regr.predict(Xte)/FOLDS return res1,res2,yte from SPARQLWrapper import SPARQLWrapper, JSON sparql = SPARQLWrapper("https://query.wikidata.org/sparql") sparql.setReturnFormat(JSON) def get_species_name(ncbi_id): q = """ select ?label where { ?s wdt:P685 "%s" ; wdt:P225 ?label . } """ % ncbi_id sparql.setQuery(q) try: results = sparql.query().convert() for result in results["results"]["bindings"]: out = result["label"]["value"] return out except: return ncbi_id def encode_fingerprints(fingerprints_all): fingerprint_encoder, fingerprint_ae = create_auto_encoder(input_size=len(fingerprints_all[0]),dense_layers=(128,),noise=0.1) fingerprint_ae.compile(optimizer='adagrad',loss='binary_crossentropy') fingerprint_ae.fit(fingerprints_all,fingerprints_all, epochs=MAX_ENCODER_EPOCHS, callbacks=[EarlyStopping('loss',min_delta=1e-5)], verbose=0) return fingerprint_encoder.predict(fingerprints_all) from sklearn.cluster import KMeans # function returns WSS score for k values from 1 to kmax def calculate_WSS(points, kmax): sse = [] for k in range(1, kmax+1): kmeans = KMeans(n_clusters = k).fit(points) centroids = kmeans.cluster_centers_ pred_clusters = kmeans.predict(points) curr_sse = 0 # calculate square of Euclidean distance of each point from its cluster center and add to current WSS for i in range(len(points)): curr_center = centroids[pred_clusters[i]] curr_sse += (points[i, 0] - curr_center[0]) ** 2 + (points[i, 1] - curr_center[1]) ** 2 sse.append(curr_sse) return sse def define_chemical_clusters(fingerprints,k=15,use_pca=True): if not isinstance(fingerprints,list): fingerprints = [fingerprints] keys = set.intersection(*[set(f.keys()) for f in fingerprints]) array = np.concatenate([np.asarray([v[k] for k in keys]) for v in fingerprints],axis=1) if use_pca: array = PCA(2).fit_transform(array) if k < 0: sse = calculate_WSS(array,25) k = np.argmin(sse) + 1 plt.plot(sse) plt.show() clusters = defaultdict(set) kmeans = KMeans(n_clusters = k).fit(array) cp = kmeans.predict(array) for k,v in zip(keys,cp): clusters[v].add(k) return clusters, kmeans.cluster_centers_ def merge_closest(clusters,cluster_centers,ord=2): dist = {} for i,cc1 in enumerate(cluster_centers): for j,cc2 in enumerate(cluster_centers): if i == j: continue dist[(i,j)] = np.linalg.norm(cc1-cc2,ord=ord) if len(dist) > 1: merge,_ = sorted(dist.items(),key=lambda x:x[1])[0] else: merge = (i,j) k1,k2 = merge cluster_centers[k1] = np.mean([cluster_centers[k1],cluster_centers[k2]],axis=0) cluster_centers = np.delete(cluster_centers,k2,axis=0) clusters[k1] |= clusters[k2] clusters.pop(k2,None) return clusters, cluster_centers def filter_data(X,Y,C_t,C_v,S_t,S_v): Xtr,Xte,ytr,yte = [],[],[],[] for x,y in zip(X,Y): c,s = x if c in C_t and s in S_t: Xtr.append(x) ytr.append(y) if c in C_v and s in S_v: Xte.append(x) yte.append(y) return Xtr,Xte,ytr,yte import sys # insert at 1, 0 is the script path (or '' in REPL) sys.path.insert(1, '/media/erik/Mass/Dropbox/NIVA_GITLAB/pySMIfp') from smiles_fingerprints import smiles_fingerprint def load_smiles_fingerprints(): q = """ select ?chembl ?smiles where { ?c wdt:P233 ?smiles ; wdt:P592 ?chembl . } """ converter = {} sparql.setQuery(q) results = sparql.query().convert() for result in results["results"]["bindings"]: ch = result["chembl"]["value"] smi = result['smiles']['value'] smifp = smiles_fingerprint(smi) converter['http://rdf.ebi.ac.uk/resource/chembl/molecule/'+ch] = smifp return converter def save_smiles_fingerprints(fp,filename='data/smiles_fingerprints.csv'): a = {} for i in range(len(smiles_fingerprint('C'))): a['sig%s'%str(i)] = [array[i] for _,array in fp.items()] df = pd.DataFrame(data={'chemical':list(fp.keys()),**a}) df.to_csv(filename) def read_smiles_fingerprints(filename): df = pd.read_csv(filename) cols = [c for c in df.columns if 'sig' in c] chemicals = df['chemical'].values arrays = df[cols].values return dict(zip(chemicals,np.asarray(arrays))) def chemical_similarities(fingerprints): keys = fingerprints.keys() array = np.asarray([i for k,i in fingerprints.items()]) sim = [] for a in array: v = a @ array.T w = np.sum(a) + np.sum(array,axis=1) sim_score = 2*v/w sim.append(sim_score) return {k:s for k,s in zip(keys,sim)} def main(): """ organic = obo['CHEBI_50860'] inorganic = obo['CHEBI_24835'] """ model = 'ComplEx' g1_parts = [[0],[0,1],[0,1,2]] g2_parts = [[0],[0,1]] p = list(product(g1_parts,g2_parts)) p += [p[-1]] ul = (False,False) f1,f2=[],[] for g1p,g2p,in p: for lit,gp,fs,name in zip([*ul],[g1p,g2p],[f1,f2],['_chemical_','_taxonomy_']): fs.append(model+name+str(hash((lit,*gp)))) if (g1p,g2p) == p[-1]: ul = (True,True) organic_chemicals = set() inorganic_chemicals = set() salts = set() for i in range(1,10): df = pd.read_csv('./data/chemical_group_%s.csv' % str(i),index_col='parent') try: organic_chemicals |= set(df.loc['http://purl.obolibrary.org/obo/CHEBI_50860','children'].split(',')) except: pass try: inorganic_chemicals |= set(df.loc['http://purl.obolibrary.org/obo/CHEBI_24835','children'].split(',')) except: pass try: salts |= set(df.loc['http://purl.obolibrary.org/obo/CHEBI_24866','children'].split(',')) except: pass print('Num organic chemicals',len(organic_chemicals)) print('Num inorganic chemicals',len(inorganic_chemicals)) print('Num salts',len(salts)) C = organic_chemicals try: smiles_fingerprints = read_smiles_fingerprints('./data/smiles_fingerprints.csv') except FileNotFoundError: smiles_fingerprints = load_smiles_fingerprints() save_smiles_fingerprints(smiles_fingerprints,'./data/smiles_fingerprints.csv') mms = MinMaxScaler().fit_transform(np.asarray([smiles_fingerprints[k] for k in smiles_fingerprints])) smiles_fingerprints = dict(zip(smiles_fingerprints,mms)) X,Y,experimental_features = load_data('./data/experiments.csv',filter_chemicals=None, filter_species=None) pubchem_fingerprints = load_fingerprints('./data/chemicals_fingerprints.csv') Y = {k:y for k,y in zip(X,Y)} pubchem_fingerprints = chemical_similarities(pubchem_fingerprints) chemical_embedding = load_embeddings('./data/embeddings/%s_entity_embeddings.npy' % f1[0], './data/embeddings/%s_entity_ids.npy' % f1[0]) species_embedding = load_embeddings('./data/embeddings/%s_entity_embeddings.npy' % f2[0], './data/embeddings/%s_entity_ids.npy' % f2[0]) chemical_features = load_features('./data/chemicals_features.csv') chemical_features = dict(zip(chemical_features,MinMaxScaler().fit_transform(np.asarray([chemical_features[k] for k in chemical_features])))) for cf in [QuantileTransformer(n_quantiles=100,output_distribution='normal')]: chemical_embedding = dict(zip(chemical_embedding,cf.fit_transform(np.asarray([chemical_embedding[k] for k in chemical_embedding])))) for cf in [QuantileTransformer(n_quantiles=100,output_distribution='normal')]: species_embedding = dict(zip(species_embedding,cf.fit_transform(np.asarray([species_embedding[k] for k in species_embedding])))) species_divisions = defaultdict(set) for k in range(1,2): df = pd.read_csv('./data/species_groups_%s.csv' % str(k), index_col='parent') for s in df.index: species_divisions[s] |= set(df.loc[s,'children'].split(',')) species_divisions = dict(filter(lambda x:len(x[1])>5,species_divisions.items())) #for k in species_divisions: #print(get_species_name(k.split('/')[-1])) #species_divisions = defaultdict(set) #df = pd.read_csv('./data/species_divisions.csv', index_col='parent') #for s in df.index: #species_divisions[s] |= set(df.loc[s,'children'].split(',')) C = set.intersection(*map(lambda k:set(k.keys()),[smiles_fingerprints,pubchem_fingerprints,chemical_features,chemical_embedding])) for d in [smiles_fingerprints,pubchem_fingerprints,chemical_embedding,chemical_features]: for c in set(d.keys()): if not c in C: d.pop(c,None) n = 7 clusters, cluster_centers = define_chemical_clusters([smiles_fingerprints],k=max(-1,n),use_pca=False) print(*map(lambda x:len(x[1]),clusters.items())) data = {} all_runs = {} TOP_K = 10 while True: for C,S in tqdm(product(clusters,species_divisions),total=len(clusters)*len(species_divisions)): k = [C,S] C = list(clusters[C]) S = species_divisions[S] k[1] = get_species_name(k[1].split('/')[-1]) loo = LeaveOneOut() predictions = [] y_true = [] for train_index, test_index in loo.split(C): C_t = [C[i] for i in train_index] C_v = [C[i] for i in test_index] r1,r2,yte = run_model(C_t,C_v,S,S,Y, experimental_features, pubchem_fingerprints, chemical_embedding, species_embedding, chemical_features, merge_species=True) if r1 is None and r2 is None: continue r1 = np.mean(r1) r2 = np.mean(r2) y_true.append(np.mean(yte)) predictions.append((r1,r2)) y_true, predictions = map(np.asarray,[y_true,predictions]) if len(predictions) < 10: continue try: if len(predictions.shape) < 2: predictions = np.expand_dims(predictions,axis=1) rsq_1 = r2_score(y_true,predictions[:,0]) rsq_2 = r2_score(y_true,predictions[:,1]) all_runs[tuple(k)] = (rsq_1,rsq_2) except ValueError: pass all_runs = dict(sorted(all_runs.items(),key=lambda x: sum(x[1])/2,reverse=True)) print(all_runs) data[len(cluster_centers)] = all_runs if len(cluster_centers) > 0: clusters, cluster_centers = merge_closest(clusters,cluster_centers) for k in list(all_runs.keys())[:TOP_K]: _,s = k species_divisions.pop(k,None) else: break pd.to_pickle(data,'chemical_cluster_merging.pkl') exit() ks = set() for k in species_divisions: S = species_divisions[k] still_true = True for k_c in clusters: C = clusters[k_c] Xtr,Xte,ytr,yte = filter_data(X,Y,C,C,S,S) if count(Xtr,Xte) > 100: ks.add(k) for k in tqdm(ks): n=6 clusters, cluster_centers = define_chemical_clusters([smiles_fingerprints],k=max(-1,n)) S = species_divisions[k] sn = get_species_name(k.split('/')[-1]) results = defaultdict(list) i = 0 while True: k_c = sorted(clusters,key=lambda x:len(clusters[x]),reverse=True)[0] C_t = clusters[k_c] if len(C_t) < 1: continue C_t,C_v = train_test_split(list(C_t),test_size=0.25) S_t = S S_v = S Xtr,Xte,ytr,yte = filter_data(X,Y,C_t,C_v,S_t,S_v) try: assert count(Xtr,Xte) > 20 r1,r2 = run_model(Xtr, Xte, ytr, yte, experimental_features, pubchem_fingerprints, chemical_embedding, species_embedding, chemical_features, merge_species=True) except AssertionError: r1,r2 = float('nan'), float('nan') except np.AxisError: r1,r2 = float('nan'), float('nan') results[i].append((r1,r2)) clusters, cluster_centers = merge_closest(clusters,cluster_centers) if len(cluster_centers) < 1: break i += 1 v0 = [[v[0] for v in results[k]] for k in results] v1 = [[v[1] for v in results[k]] for k in results] fig, ax = plt.subplots() for x,color,ran in zip([v0,v1],['red','green'],[np.arange(0,len(v0)*2,2),np.arange(1,len(v1)*2,2)]): mins = [np.nanmin(a) for a in x] maxes = [np.nanmax(a) for a in x] means = [np.nanmean(a) for a in x] std = [np.nanstd(a) for a in x] mins,maxes,means,std = map(np.asarray,[mins,maxes,means,std]) ax.bar(ran,maxes,width=0.5,color=color) #plt.ylim(-1,1) ax.set_xticks(np.arange(0.5,len(v0)*2,2)) ax.set_xticklabels(('%s Clusters' % str(abs(i)) for i in range(-n,0))) plt.savefig('./plots/chemical_clusters_taxon_%s.png' % sn) exit() #def tqdm(x,**params): #return x for filter_chemicals,string,TOP_K in tqdm(zip([inorganic_chemicals | salts],['organic'],[4]),total=1,desc='Chemical Groups'): #if string=='organic': continue for division in tqdm(S_v,total=len(S_v),desc='Divisions'): if not len(S_v[division]) > 1: continue model_params={'encode':False,'train_ae_fingerprints':False,'train_ae_species':False} results = [[]]*TOP_K f = lambda _s: sum([1 for c,s in X if (s == _s and c in C-filter_chemicals)]) tmp_division = list(sorted(S_v[division],key=f,reverse=True))[:TOP_K] for i,s_v in tqdm(enumerate(tmp_division),desc='Species in division %s' % division,leave=False,total=len(tmp_division)): C_restriction = {'C_v':{'not content':filter_chemicals},'C_t':{'not content':filter_chemicals}} configs = [] #Method 1 configs.append((1, 1, {'S_v':{'content':set([s_v]),'max_len':1}})) configs.append((1, 2, {'S_v':{'content':set([s_v]),'max_len':1}})) #Method 2 configs.append((2, 1, {'S_v':{'content':set([s_v]),'max_len':1}})) #Method 3 configs.append((3, 1, {'S_v':{'content':set([s_v]),'max_len':1}})) configs.append((3, 2, {'S_v':{'content':set([s_v]),'max_len':1}})) #Method 4 configs.append((4, 1, {'S_v':{'content':set([s_v]),'max_len':1}})) configs.append((4, 2, {'S_v':{'content':set([s_v]),'max_len':1}})) tmp_res = np.zeros((len(configs),2)) for j,config in tqdm(enumerate(configs),total=len(configs),leave=False,desc='Configs'): m,v,res = config r1_tmp = [] r2_tmp = [] for _ in range(10): tf.keras.backend.clear_session() prop = 0.3 Xtr,Xte,ytr,yte = data_split(X,Y,restrictions={**res,**C_restriction},method=m,variant=v,prop=prop) try: r1,r2 = run_model(Xtr, Xte, ytr, yte, experimental_features, fingerprints, chemical_embedding, species_embedding, model_params=model_params) except: r1,r2=0,0 r1_tmp.append(r1) r2_tmp.append(r2) tmp_res[j,0] = np.mean(r1_tmp) tmp_res[j,1] = np.mean(r2_tmp) results[i] = tmp_res fig, axs = plt.subplots(1,len(results),figsize=(40, 10)) for i,ax in enumerate(axs): ms = results[i] baseline = ms[:,0] over = ms[:,1] baseline = np.nan_to_num(baseline, nan=0.0,posinf=0.0, neginf=0.0) over = np.nan_to_num(over, nan=0.0,posinf=0.0, neginf=0.0) width = 0.4 ax.bar(np.arange(0,len(baseline)*2,2),baseline,width,color='red') ax.bar(np.arange(1,len(baseline)*2,2),over,width,color='green') ax.set_title(get_species_name(tmp_division[i].split('/')[-1])) ax.set_xticks(np.arange(0.5,len(baseline)*2,2)) ax.set_xticklabels((str(i) for i in range(len(configs)))) ax.set_ylim(0,max(*over,*baseline)+0.1) plt.savefig('plots/division_%s_%s.png' % (division,string)) if __name__ == '__main__': main()
32,681
35.394209
169
py
kge_ecotox_regression
kge_ecotox_regression-main/embedding_model.py
from tensorflow.keras import Model, Sequential from tensorflow.keras.layers import Input, Embedding, Dense, Dropout, Conv2D, Flatten, Concatenate, Multiply import tensorflow as tf def min_distance_loss(w,epsilon=1.0): r = tf.reduce_sum(w*w, 1) r = tf.reshape(r, [-1, 1]) D = r - 2*tf.matmul(w, tf.transpose(w)) + tf.transpose(r) D = D + tf.linalg.diag(epsilon * tf.ones(D.shape[0])) return tf.reduce_sum(tf.where(D<epsilon,1.0,0.0))/tf.cast(w.shape[1],tf.float32) def TransE(entities,relations,dim=200,bias=1,lamb=1,norm_size=0.0,mdl=0.0): inp = Input((3,)) inp_label = Input(()) s,p,o = tf.unstack(inp,axis=-1) entity_embedding = Embedding(len(entities),dim,name='entity_embedding') relation_embedding = Embedding(len(relations),dim,name='relation_embedding') h,r,t = entity_embedding(s),relation_embedding(p),entity_embedding(o) score = bias - tf.norm(h+r-t, ord=2, axis=-1) loss = lamb - inp_label * score loss = tf.where(loss>0,loss,0) + \ norm_size * tf.norm(entity_embedding.weights[0],ord=2)**2 + \ min_distance_loss(entity_embedding.weights[0]) * mdl model = Model(inputs=[inp,inp_label],outputs=score) model.add_loss(loss) model.compile(optimizer='adam',loss=None) return model def DistMult(entities,relations,dim=200,norm_size=0.0,mdl=0.0): inp = Input((3,)) inp_label = Input(()) s,p,o = tf.unstack(inp,axis=-1) entity_embedding = Embedding(len(entities),dim,name='entity_embedding') relation_embedding = Embedding(len(relations),dim,name='relation_embedding') h,r,t = entity_embedding(s),relation_embedding(p),entity_embedding(o) score = tf.keras.layers.Activation('linear')(tf.reduce_sum(h*r*t,axis=-1)) model = Model(inputs=[inp,inp_label],outputs=score) loss = lambda true,pred: tf.reduce_sum(tf.math.log(1+tf.math.exp(-true*pred))) + \ norm_size * tf.norm(entity_embedding.weights[0],ord=2)**2 + \ min_distance_loss(entity_embedding.weights[0],mdl) * mdl model.compile(optimizer='adam',loss=loss) return model def ComplEx(entities,relations,dim=200,norm_size=0.0,mdl=0.0): inp = Input((3,)) inp_label = Input(()) s,p,o = tf.unstack(inp,axis=-1) entity_embedding = Embedding(len(entities),dim,name='entity_embedding') relation_embedding = Embedding(len(relations),dim,name='relation_embedding') h,r,t = entity_embedding(s),relation_embedding(p),entity_embedding(o) h_real,h_img = tf.split(h,2,axis=-1) r_real,r_img = tf.split(r,2,axis=-1) t_real,t_img = tf.split(t,2,axis=-1) score = tf.reduce_sum(r_real*h_real*t_real,axis=-1) + \ tf.reduce_sum(r_real*h_img*t_img,axis=-1) + \ tf.reduce_sum(r_img*h_real*t_img,axis=-1) - \ tf.reduce_sum(r_img*h_img*t_real,axis=-1) model = Model(inputs=[inp,inp_label],outputs=score) loss = lambda true,pred: tf.reduce_sum(tf.math.log(1+tf.math.exp(-true*pred))) + \ norm_size * tf.norm(entity_embedding.weights[0],ord=2)**2 + \ min_distance_loss(entity_embedding.weights[0]) * mdl model.compile(optimizer='adam',loss=loss) return model def ConvE(entities,relations): dim = 200 inp = Input((3,)) inp_label = Input(()) s,p,o = tf.unstack(inp,axis=-1) entity_embedding = Embedding(len(entities),dim,name='entity_embedding') relation_embedding = Embedding(len(relations),dim,name='relation_embedding') h,r,t = entity_embedding(s),relation_embedding(p),entity_embedding(o) h = tf.reshape(h,(-1,20,10,1)) r = tf.reshape(r,(-1,20,10,1)) x = Concatenate(axis=2)([h,r]) x = Conv2D(16,(5,5),activation='relu')(x) x = Dropout(0.2)(x) x = Conv2D(16,(3,3),activation='relu')(x) x = Dropout(0.2)(x) x = Flatten()(x) x = Dense(dim)(x) x = Multiply()([x,t]) x = Dense(1,activation='sigmoid')(x) model = Model(inputs=[inp,inp_label],outputs=x) model.compile(optimizer='adam',loss=tf.keras.losses.BinaryCrossentropy(label_smoothing=0.05)) return model
4,177
31.897638
108
py
kge_ecotox_regression
kge_ecotox_regression-main/pretrained_embedding_models.py
import sys import os from itertools import product from KGEkeras import DistMult, HolE, TransE, HAKE, ConvE, ComplEx, ConvR, RotatE, pRotatE, ConvKB, CosinE from kerastuner import RandomSearch, HyperParameters, Objective, Hyperband, BayesianOptimization from random import choice from collections import defaultdict from tensorflow.keras.losses import binary_crossentropy,hinge,mean_squared_error from tensorflow.keras import Input from tensorflow.keras import Model import pandas as pd import numpy as np from sklearn.preprocessing import MinMaxScaler from sklearn.model_selection import train_test_split from tensorflow.keras.optimizers import Adam from tensorflow.keras.callbacks import EarlyStopping, Callback, TerminateOnNaN, ReduceLROnPlateau from sklearn.metrics.cluster import completeness_score from tensorflow.keras.optimizers import Adam import json import tensorflow as tf from tensorflow.keras.optimizers.schedules import ExponentialDecay from KGEkeras import loss_function_lookup from lib.utils import generate_negative, oversample_data, load_data from tqdm import tqdm import string import random from random import choices from lib.hptuner import HPTuner import pickle try: from tensorflow_addons.callbacks import TimeStopping except: pass from rdflib import Graph, URIRef, Literal, Namespace from KGEkeras import LiteralConverter from sklearn.decomposition import PCA SECONDS_PER_TRAIL = 600 SECONDS_TO_TERMINATE = 3600 SEARCH_MAX_EPOCHS = 10 MAX_EPOCHS = 200 MIN_EPOCHS = 50 MAX_TRIALS = 20 PATIENCE = 10 EPSILON = 10e-7 models = { #'DistMult':DistMult, #'TransE':TransE, #'HolE':HolE, 'ComplEx':ComplEx, #'HAKE':HAKE, #'pRotatE':pRotatE, #'RotatE':RotatE, #'ConvE':ConvE, #'ConvKB':ConvKB, } class DataGenerator(tf.keras.utils.Sequence): def __init__(self, kg, ns=10, batch_size=32, shuffle=True): self.batch_size = min(batch_size,len(kg)) self.kg = kg self.ns = ns self.num_e = len(set([s for s,_,_ in kg])|set([o for _,_,o in kg])) self.shuffle = shuffle self.indices = list(range(len(kg))) self.on_epoch_end() def __len__(self): return len(self.kg) // self.batch_size def __getitem__(self, index): index = self.index[index * self.batch_size:(index + 1) * self.batch_size] batch = [self.indices[k] for k in index] X, y = self.__get_data(batch) return X, y def on_epoch_end(self): self.index = np.arange(len(self.indices)) if self.shuffle == True: np.random.shuffle(self.index) def __get_data(self, batch): tmp_kg = np.asarray([self.kg[i] for i in batch]) negative_kg = generate_negative(tmp_kg,N=self.num_e,negative=self.ns) X = oversample_data(kgs=[tmp_kg,negative_kg]) return X, None def build_model(hp): params = hp.copy() params['e_dim'] = params['dim'] params['r_dim'] = params['dim'] params['name'] = 'embedding_model' embedding_model = models[params['embedding_model']] embedding_model = embedding_model(**params) triple = Input((3,)) ftriple = Input((3,)) inputs = [triple, ftriple] score = embedding_model(triple) fscore = embedding_model(ftriple) loss_function = loss_function_lookup(params['loss_function']) loss = loss_function(score,fscore,params['margin'] or 1, 1) model = Model(inputs=inputs, outputs=loss) model.add_loss(loss) model.compile(optimizer=Adam(learning_rate=ExponentialDecay(params['learning_rate'],decay_steps=100000,decay_rate=0.96)), loss=None) return model def optimize_model(model, kg, lit=False, name='name', hp=None): if lit: lc = LiteralConverter(kg) literals = lc.fit_transform() kg = lc.g literals = PCA(min(len(literals[0]),100)).fit_transform(literals) else: literals = None kg -= [(s,p,o) for s,p,o in kg if isinstance(o,Literal)] entities = set(kg.subjects()) | set(kg.objects()) relations = set(kg.predicates()) me = {k:i for i,k in enumerate(entities)} mr = {k:i for i,k in enumerate(relations)} kg = list(map(lambda x: (me[x[0]],mr[x[1]],me[x[2]]), kg)) bs = 512 kg = np.asarray(kg) model_name = model N = len(me) M = len(mr) hptuner = HPTuner(runs=MAX_TRIALS, objectiv_direction='min') hptuner.add_value_hp('gamma',0,21) hptuner.add_value_hp('dim',100,401,dtype=int) hptuner.add_value_hp('negative_samples',10,101,dtype=int) hptuner.add_value_hp('margin',1,11,dtype=int) hptuner.add_list_hp('loss_function',['pairwize_hinge','pairwize_logistic','pointwize_hinge','pointwize_logistic'],exhaustive=True) hptuner.add_fixed_hp('embedding_model',model) hptuner.add_fixed_hp('dp',0.2) hptuner.add_fixed_hp('hidden_dp',0.2) hptuner.add_fixed_hp('num_entities',N) hptuner.add_fixed_hp('num_relations',M) if hp: for k,i in hp.items(): hptuner.add_fixed_hp(k,i) hptuner.add_fixed_hp('num_entities',N) hptuner.add_fixed_hp('num_relations',M) hptuner.add_fixed_hp('learning_rate',0.001) hptuner.add_fixed_hp('regularization',0.001) if lit: hptuner.add_fixed_hp('literals',literals) hptuner.add_fixed_hp('literal_activation','tanh') if hp: hptuner.next_hp_config() hptuner.add_result(0.0) with tqdm(total=hptuner.runs, desc='Trials') as pbar: while hptuner.is_active and hp is None: hp = hptuner.next_hp_config() model = build_model(hp) tr_gen = DataGenerator(kg, batch_size=bs, shuffle=True, ns=hp['negative_samples']) hist = model.fit(tr_gen,epochs=SEARCH_MAX_EPOCHS,verbose=2, callbacks=[EarlyStopping('loss'),TerminateOnNaN()]) score = hist.history['loss'][-1]/hist.history['loss'][0] hptuner.add_result(score) tf.keras.backend.clear_session() pbar.update(1) hp = hptuner.best_config() #if hp is None: #with open('./pretrained_hp/%s%s_kg.json' % (model_name,name), 'w') as fp: #json.dump(hp, fp) model = build_model(hp) tr_gen = DataGenerator(kg, batch_size=bs, shuffle=True, ns=hp['negative_samples']) hist = model.fit(tr_gen,epochs=MAX_EPOCHS, verbose=2, callbacks=[EarlyStopping('loss',patience=PATIENCE), TerminateOnNaN()]) if np.isnan(hist.history['loss'][-1]): print(model_name,'nan loss.') return optimize_model(model_name,kg,lit,name,None) for l in model.layers: if isinstance(l,models[model_name]): m = l.name m, W1, W2 = model, model.get_layer(m).entity_embedding.get_weights()[0], model.get_layer(m).relational_embedding.get_weights()[0] m.save_weights('pretrained_models/model/'+name) np.save(name+'_entity_embeddings.npy', W1) np.save(name+'_entity_ids.npy',np.asarray(list(zip(entities,range(len(entities)))))) np.save(name+'_relational_embeddings.npy', W2) np.save(name+'_relation_ids.npy',np.asarray(list(zip(relations,range(len(relations)))))) def main(): d = './data/embeddings/' use_literals = product([False,True],[False,True]) g1_parts = [[0],[0,1],[0,1,2]] g2_parts = [[0],[0,1]] p = list(product(g1_parts,g2_parts)) p += [p[-1]] ul = (False,False) for g1p,g2p in tqdm(p): g1,g2 = Graph(),Graph() for i in g1p: g = Graph() g.load('./data/chemicals_%s.ttl' % str(i),format='ttl') g1 += g for i in g2p: g = Graph() g.load('./data/taxonomy_%s.ttl' % str(i),format='ttl') g2 += g for lit,gp,kg,name in zip([*ul],[g1p,g2p],[g1,g2],['_chemical_','_taxonomy_']): #hp_file = '../KGE-CEP/pretrained_hp/%s%s_kg.json' % (model,name) hp = {'e_dim':100, 'negative_samples':10, 'loss_function':'pairwize_logistic'} model = 'ComplEx' f = d+model+name+str(hash((lit,*gp))) optimize_model(model,kg,lit,name=f,hp=hp) tf.keras.backend.clear_session() if (g1p,g2p) == p[-1]: ul = (True,True) if __name__ == '__main__': main()
8,625
30.140794
134
py
kge_ecotox_regression
kge_ecotox_regression-main/autoencoder.py
from tensorflow.keras.layers import Dense, GaussianNoise, Input, LayerNormalization from tensorflow.keras.models import Model from tensorflow import keras def create_auto_encoder(input_size, dense_layers = (10,), noise=0): autoencoder = keras.Sequential() if noise > 0: autoencoder.add(GaussianNoise(noise)) for l in dense_layers: autoencoder.add(Dense(l,activation='relu')) encoder = autoencoder for l in dense_layers[::-1]: autoencoder.add(Dense(l,activation='relu')) autoencoder.add(Dense(input_size,activation='sigmoid')) return encoder, autoencoder
613
33.111111
83
py
lepard
lepard-main/main.py
import os, torch, json, argparse, shutil from easydict import EasyDict as edict import yaml from datasets.dataloader import get_dataloader, get_datasets from models.pipeline import Pipeline from lib.utils import setup_seed from lib.tester import get_trainer from models.loss import MatchMotionLoss from lib.tictok import Timers from configs.models import architectures from torch import optim setup_seed(0) def join(loader, node): seq = loader.construct_sequence(node) return '_'.join([str(i) for i in seq]) yaml.add_constructor('!join', join) if __name__ == '__main__': # load configs parser = argparse.ArgumentParser() parser.add_argument('config', type=str, help= 'Path to the config file.') args = parser.parse_args() with open(args.config,'r') as f: config = yaml.load(f, Loader=yaml.Loader) config['snapshot_dir'] = 'snapshot/%s/%s' % (config['dataset']+config['folder'], config['exp_dir']) config['tboard_dir'] = 'snapshot/%s/%s/tensorboard' % (config['dataset']+config['folder'], config['exp_dir']) config['save_dir'] = 'snapshot/%s/%s/checkpoints' % (config['dataset']+config['folder'], config['exp_dir']) config = edict(config) os.makedirs(config.snapshot_dir, exist_ok=True) os.makedirs(config.save_dir, exist_ok=True) os.makedirs(config.tboard_dir, exist_ok=True) if config.gpu_mode: config.device = torch.device("cuda:0") else: config.device = torch.device('cpu') # backup the if config.mode == 'train': os.system(f'cp -r models {config.snapshot_dir}') os.system(f'cp -r configs {config.snapshot_dir}') os.system(f'cp -r cpp_wrappers {config.snapshot_dir}') os.system(f'cp -r datasets {config.snapshot_dir}') os.system(f'cp -r kernels {config.snapshot_dir}') os.system(f'cp -r lib {config.snapshot_dir}') shutil.copy2('main.py',config.snapshot_dir) # model initialization config.kpfcn_config.architecture = architectures[config.dataset] config.model = Pipeline(config) # config.model = KPFCNN(config) # create optimizer if config.optimizer == 'SGD': config.optimizer = optim.SGD( config.model.parameters(), lr=config.lr, momentum=config.momentum, weight_decay=config.weight_decay, ) elif config.optimizer == 'ADAM': config.optimizer = optim.Adam( config.model.parameters(), lr=config.lr, betas=(0.9, 0.999), weight_decay=config.weight_decay, ) #create learning rate scheduler if 'overfit' in config.exp_dir : config.scheduler = optim.lr_scheduler.MultiStepLR( config.optimizer, milestones=[config.max_epoch-1], # fix lr during overfitting gamma=0.1, last_epoch=-1) else: config.scheduler = optim.lr_scheduler.ExponentialLR( config.optimizer, gamma=config.scheduler_gamma, ) config.timers = Timers() # create dataset and dataloader train_set, val_set, test_set = get_datasets(config) config.train_loader, neighborhood_limits = get_dataloader(train_set,config,shuffle=True) config.val_loader, _ = get_dataloader(val_set, config, shuffle=False, neighborhood_limits=neighborhood_limits) config.test_loader, _ = get_dataloader(test_set, config, shuffle=False, neighborhood_limits=neighborhood_limits) # config.desc_loss = MetricLoss(config) config.desc_loss = MatchMotionLoss (config['train_loss']) trainer = get_trainer(config) if(config.mode=='train'): trainer.train() else: trainer.test()
3,723
32.54955
116
py
lepard
lepard-main/models/matching.py
import torch import torch.nn as nn import torch.nn.functional as F from models.position_encoding import VolumetricPositionEncoding as VolPE def log_optimal_transport(scores, alpha, iters, src_mask, tgt_mask ): b, m, n = scores.shape if src_mask is None: ms = m ns = n else : ms = src_mask.sum(dim=1, keepdim=True) ns = tgt_mask.sum(dim=1, keepdim=True) bins0 = alpha.expand(b, m, 1) bins1 = alpha.expand(b, 1, n) alpha = alpha.expand(b, 1, 1) Z = torch.cat([torch.cat([scores, bins0], -1), torch.cat([bins1, alpha], -1)], 1) norm = - (ms + ns).log() # [b, 1] log_mu = torch.cat([norm .repeat(1, m), ns.log() + norm], dim=1) log_nu = torch.cat([norm.repeat(1, n), ms.log() + norm], dim=1) u, v = torch.zeros_like(log_mu), torch.zeros_like(log_nu) for _ in range(iters): u = log_mu - torch.logsumexp( Z + v.unsqueeze(1), dim=2) v = log_nu - torch.logsumexp(Z + u.unsqueeze(2), dim=1) Z= Z + u.unsqueeze(2) + v.unsqueeze(1) Z = Z - norm.view(-1,1,1) return Z class Matching(nn.Module): def __init__(self, config): super().__init__() self.match_type = config['match_type'] self.confidence_threshold = config['confidence_threshold'] d_model = config['feature_dim'] self.src_proj = nn.Linear(d_model, d_model, bias=False) self.tgt_proj = nn.Linear(d_model, d_model, bias=False) self.entangled= config['entangled'] if self.match_type == "dual_softmax": self.temperature = config['dsmax_temperature'] elif self.match_type == 'sinkhorn': #sinkhorn algorithm self.skh_init_bin_score = config['skh_init_bin_score'] self.skh_iters = config['skh_iters'] self.skh_prefilter = config['skh_prefilter'] self.bin_score = nn.Parameter( torch.tensor( self.skh_init_bin_score, requires_grad=True)) else: raise NotImplementedError() @staticmethod @torch.no_grad() def get_match( conf_matrix, thr, mutual=True): mask = conf_matrix > thr #mutual nearest if mutual: mask = mask \ * (conf_matrix == conf_matrix.max(dim=2, keepdim=True)[0]) \ * (conf_matrix == conf_matrix.max(dim=1, keepdim=True)[0]) #find all valid coarse matches index = (mask==True).nonzero() b_ind, src_ind, tgt_ind = index[:,0], index[:,1], index[:,2] mconf = conf_matrix[b_ind, src_ind, tgt_ind] return index, mconf, mask @staticmethod @torch.no_grad() def get_topk_match( conf_matrix, thr, mutual=True): mask = conf_matrix > thr #mutual nearest if mutual: mask = mask \ * (conf_matrix == conf_matrix.max(dim=2, keepdim=True)[0]) \ * (conf_matrix == conf_matrix.max(dim=1, keepdim=True)[0]) #find all valid coarse matches index = (mask==True).nonzero() b_ind, src_ind, tgt_ind = index[:,0], index[:,1], index[:,2] mconf = conf_matrix[b_ind, src_ind, tgt_ind] return index, mconf, mask def forward(self, src_feats, tgt_feats, src_pe, tgt_pe, src_mask, tgt_mask, data, pe_type="rotary"): ''' @param src_feats: [B, S, C] @param tgt_feats: [B, T, C] @param src_mask: [B, S] @param tgt_mask: [B, T] @return: ''' src_feats = self.src_proj(src_feats) tgt_feats = self.src_proj(tgt_feats) data["src_feats_nopos"] = src_feats data["tgt_feats_nopos"] = tgt_feats if not self.entangled : src_feats = VolPE.embed_pos(pe_type, src_feats, src_pe) tgt_feats = VolPE.embed_pos(pe_type, tgt_feats, tgt_pe) data["src_feats"] = src_feats data["tgt_feats"] = tgt_feats src_feats, tgt_feats = map(lambda feat: feat / feat.shape[-1] ** .5, [src_feats, tgt_feats]) if self.match_type == "dual_softmax": # dual softmax matching sim_matrix_1 = torch.einsum("bsc,btc->bst", src_feats, tgt_feats) / self.temperature if src_mask is not None: sim_matrix_2 = sim_matrix_1.clone() sim_matrix_1.masked_fill_(~src_mask[:, :, None], float('-inf')) sim_matrix_2.masked_fill_(~tgt_mask[:, None, :], float('-inf')) conf_matrix = F.softmax(sim_matrix_1, 1) * F.softmax(sim_matrix_2, 2) else : conf_matrix = F.softmax(sim_matrix_1, 1) * F.softmax(sim_matrix_1, 2) elif self.match_type == "sinkhorn" : #optimal transport sinkhoron sim_matrix = torch.einsum("bsc,btc->bst", src_feats, tgt_feats) if src_mask is not None: sim_matrix.masked_fill_( ~(src_mask[..., None] * tgt_mask[:, None]).bool(), float('-inf')) log_assign_matrix = log_optimal_transport( sim_matrix, self.bin_score, self.skh_iters, src_mask, tgt_mask) assign_matrix = log_assign_matrix.exp() conf_matrix = assign_matrix[:, :-1, :-1].contiguous() coarse_match, _, _ = self.get_match(conf_matrix, self.confidence_threshold) return conf_matrix, coarse_match
5,412
29.931429
118
py
lepard
lepard-main/models/loss.py
import torch import torch.nn as nn import numpy as np import open3d as o3d from lib.benchmark_utils import to_o3d_pcd from lib.visualization import * import nibabel.quaternions as nq from sklearn.metrics import precision_recall_fscore_support from datasets.utils import blend_scene_flow, multual_nn_correspondence, knn_point_np from models.matching import Matching as CM def ransac_pose_estimation(src_pcd, tgt_pcd, corrs, distance_threshold=0.05, ransac_n=3): src_pcd = to_o3d_pcd(src_pcd) tgt_pcd = to_o3d_pcd(tgt_pcd) corrs = o3d.utility.Vector2iVector(np.array(corrs).T) result_ransac = o3d.registration.registration_ransac_based_on_correspondence( source=src_pcd, target=tgt_pcd, corres=corrs, max_correspondence_distance=distance_threshold, estimation_method=o3d.registration.TransformationEstimationPointToPoint(False), ransac_n=ransac_n, criteria=o3d.registration.RANSACConvergenceCriteria(50000, 1000)) return result_ransac.transformation def computeTransformationErr(trans, info): """ Computer the transformation error as an approximation of the RMSE of corresponding points. More informaiton at http://redwood-data.org/indoor/registration.html Args: trans (numpy array): transformation matrices [n,4,4] info (numpy array): covariance matrices of the gt transformation paramaters [n,4,4] Returns: p (float): transformation error """ t = trans[:3, 3] r = trans[:3, :3] q = nq.mat2quat(r) er = np.concatenate([t, q[1:]], axis=0) p = er.reshape(1, 6) @ info @ er.reshape(6, 1) / info[0, 0] return p.item() class MatchMotionLoss(nn.Module): def __init__(self, config): super().__init__() self.focal_alpha = config['focal_alpha'] self.focal_gamma = config['focal_gamma'] self.pos_w = config['pos_weight'] self.neg_w = config['neg_weight'] self.mot_w = config['motion_weight'] self.mat_w = config['match_weight'] self.motion_loss_type = config['motion_loss_type'] self.match_type = config['match_type'] self.positioning_type = config['positioning_type'] self.registration_threshold = config['registration_threshold'] self.confidence_threshold_metric = config['confidence_threshold_metric'] self.inlier_thr = config['inlier_thr'] self.fmr_thr = config['fmr_thr'] self.mutual_nearest = config['mutual_nearest'] self.dataset = config['dataset'] def forward(self, data): loss_info = {} loss = self.ge_coarse_loss(data, loss_info) loss_info.update({ 'loss': loss }) return loss_info def ge_coarse_loss(self, data, loss_info, eval_metric=False): if self.dataset == "4dmatch": s2t_flow = torch.zeros_like(data['s_pcd']) for i, cflow in enumerate(data['coarse_flow']): s2t_flow[i][: len(cflow)] = cflow loss = 0. src_mask = data['src_mask'] tgt_mask = data['tgt_mask'] conf_matrix_pred = data['conf_matrix_pred'] match_gt = data['coarse_matches'] R_s2t_gt = data['batched_rot'] t_s2t_gt = data['batched_trn'] #get the overlap mask, for dense motion loss s_overlap_mask = torch.zeros_like(src_mask).bool() for bi, corr in enumerate (match_gt): s_overlap_mask[bi][ corr[0] ] = True # compute focal loss c_weight = (src_mask[:, :, None] * tgt_mask[:, None, :]).float() conf_matrix_gt = self.match_2_conf_matrix(match_gt, conf_matrix_pred) data['conf_matrix_gt'] = conf_matrix_gt focal_coarse = self.compute_correspondence_loss(conf_matrix_pred, conf_matrix_gt, weight=c_weight) recall, precision = self.compute_match_recall( conf_matrix_gt, data['coarse_match_pred']) loss_info.update( { "focal_coarse": focal_coarse, "recall_coarse": recall, "precision_coarse": precision } ) loss = loss + self.mat_w * focal_coarse if recall > 0.01 and self.mot_w > 0: R_s2t_pred = data["R_s2t_pred"] t_s2t_pred = data["t_s2t_pred"] #compute predicted flow. Note, if 4dmatch, the R_pred,t_pred try to find the best rigid fit of deformation src_pcd_wrapped_pred = (torch.matmul(R_s2t_pred, data['s_pcd'].transpose(1, 2)) + t_s2t_pred).transpose(1, 2) sflow_pred = src_pcd_wrapped_pred - data['s_pcd'] if self.dataset == '4dmatch': spcd_deformed = data['s_pcd'] + s2t_flow src_pcd_wrapped_gt = (torch.matmul(R_s2t_gt, spcd_deformed.transpose(1, 2)) + t_s2t_gt).transpose(1, 2) else : # 3dmatch src_pcd_wrapped_gt = (torch.matmul(R_s2t_gt, data['s_pcd'].transpose(1, 2)) + t_s2t_gt).transpose(1, 2) sflow_gt = src_pcd_wrapped_gt - data['s_pcd'] e1 = torch.sum(torch.abs(sflow_pred - sflow_gt), 2) e1 = e1[s_overlap_mask] # [data['src_mask']] l1_loss = torch.mean(e1) loss = loss + self.mot_w * l1_loss # # if eval_metric : # # match_pred, _, _ = CM.get_match(data['conf_matrix_pred'], thr=self.confidence_threshold_metric, mutual=self.mutual_nearest) # # '''Inlier Ratio (IR)''' # ir = self.compute_inlier_ratio(match_pred, data, self.inlier_thr, # s2t_flow=s2t_flow if self.dataset == "4dmatch" else None) # loss_info.update({"Inlier Ratio": ir.mean()}) # # if self.dataset == '3dmatch': # # '''Feature Matching Recall (FMR)''' # fmr = (ir > self.fmr_thr).float().sum() / len(ir) # loss_info.update({"Feature Matching Recall": fmr}) # # '''Registration Recall (RR)''' # rot_, trn_ = self.ransac_regist_coarse(data['s_pcd'], data['t_pcd'], src_mask, tgt_mask , match_pred) # rot, trn = rot_.to(data['s_pcd']) , trn_.to(data['s_pcd']) # rr = self.compute_registration_recall(rot, trn, data, self.registration_threshold) # loss_info.update({'Registration_Recall': rr}) if self.positioning_type == "procrustes": for layer_ind in data["position_layers"]: # compute focal loss rpe_conf_matrix = data["position_layers"][layer_ind]["conf_matrix"] focal_rpe = self.compute_correspondence_loss(rpe_conf_matrix, conf_matrix_gt, weight=c_weight) recall, precision = self.compute_match_recall(conf_matrix_gt, data["position_layers"][layer_ind]['match_pred']) # loss_info.update({'focal_layer_%d' % layer_ind: focal_rpe, 'recall_layer_%d' % layer_ind: recall, # 'precision_layer_%d' % layer_ind: precision}) loss = loss + self.mat_w * focal_rpe if recall >0.01 and self.mot_w > 0: R_s2t_pred = data["position_layers"][layer_ind]["R_s2t_pred"] t_s2t_pred = data["position_layers"][layer_ind]["t_s2t_pred"] src_pcd_wrapped_pred = (torch.matmul(R_s2t_pred, data['s_pcd'].transpose(1, 2)) + t_s2t_pred).transpose(1, 2) sflow_pred = src_pcd_wrapped_pred - data['s_pcd'] if self.dataset == '4dmatch': spcd_deformed = data['s_pcd'] + s2t_flow src_pcd_wrapped_gt = ( torch.matmul(R_s2t_gt, spcd_deformed.transpose(1, 2)) + t_s2t_gt).transpose(1, 2) else: # 3dmatch src_pcd_wrapped_gt = ( torch.matmul(R_s2t_gt, data['s_pcd'].transpose(1, 2)) + t_s2t_gt).transpose(1, 2) sflow_gt = src_pcd_wrapped_gt - data['s_pcd'] e1 = torch.sum(torch.abs(sflow_pred - sflow_gt), 2) #[data['src_mask']] e1 = e1[s_overlap_mask] # [data['src_mask']] l1_loss = torch.mean(e1) loss = loss + self.mot_w * l1_loss return loss @staticmethod def compute_nrfmr(match_pred, data, recall_thr=0.04): s_pcd, t_pcd = data['s_pcd'], data['t_pcd'] s_pcd_raw = data['src_pcd_list'] sflow_list = data['sflow_list'] metric_index_list = data['metric_index_list'] batched_rot = data['batched_rot'] # B,3,3 batched_trn = data['batched_trn'] nrfmr = 0. for i in range(len(s_pcd_raw)): # use the match prediction as the motion anchor match_pred_i = match_pred[match_pred[:, 0] == i] s_id, t_id = match_pred_i[:, 1], match_pred_i[:, 2] s_pcd_matched = s_pcd[i][s_id] t_pcd_matched = t_pcd[i][t_id] motion_pred = t_pcd_matched - s_pcd_matched if len(s_pcd_matched) >= 3 : # get the wrapped metric points metric_index = metric_index_list[i] sflow = sflow_list[i] s_pcd_raw_i = s_pcd_raw[i] metric_pcd = s_pcd_raw_i[metric_index] metric_sflow = sflow[metric_index] metric_pcd_deformed = metric_pcd + metric_sflow metric_pcd_wrapped_gt = (torch.matmul(batched_rot[i], metric_pcd_deformed.T) + batched_trn[i]).T # blend the motion for metric points try: metric_motion_pred, valid_mask = MatchMotionLoss.blend_anchor_motion( metric_pcd.cpu().numpy(), s_pcd_matched.cpu().numpy(), motion_pred.cpu().numpy(), knn=3, search_radius=0.1) metric_pcd_wrapped_pred = metric_pcd + torch.from_numpy(metric_motion_pred).to(metric_pcd) dist = torch.sqrt(torch.sum((metric_pcd_wrapped_pred - metric_pcd_wrapped_gt) ** 2, dim=1)) r = (dist < recall_thr).float().sum() / len(dist) except : r = 0 nrfmr = nrfmr + r debug = False if debug: import mayavi.mlab as mlab c_red = (224. / 255., 0 / 255., 125 / 255.) c_pink = (224. / 255., 75. / 255., 232. / 255.) c_blue = (0. / 255., 0. / 255., 255. / 255.) scale_factor = 0.013 metric_pcd_wrapped_gt = metric_pcd_wrapped_gt.cpu() metric_pcd_wrapped_pred = metric_pcd_wrapped_pred.cpu() err = metric_pcd_wrapped_pred - metric_pcd_wrapped_gt mlab.points3d(metric_pcd_wrapped_gt[:, 0], metric_pcd_wrapped_gt[:, 1], metric_pcd_wrapped_gt[:, 2], scale_factor=scale_factor, color=c_pink) mlab.points3d(metric_pcd_wrapped_pred[:, 0], metric_pcd_wrapped_pred[:, 1], metric_pcd_wrapped_pred[:, 2], scale_factor=scale_factor, color=c_blue) mlab.quiver3d(metric_pcd_wrapped_gt[:, 0], metric_pcd_wrapped_gt[:, 1], metric_pcd_wrapped_gt[:, 2], err[:, 0], err[:, 1], err[:, 2], scale_factor=1, mode='2ddash', line_width=1.) mlab.show() nrfmr = nrfmr / len(s_pcd_raw) return nrfmr @staticmethod def blend_anchor_motion(query_loc, reference_loc, reference_flow, knn=3, search_radius=0.1): '''approximate flow on query points this function assume query points are sub- or un-sampled from reference locations @param query_loc:[m,3] @param reference_loc:[n,3] @param reference_flow:[n,3] @param knn: @return: blended_flow:[m,3] ''' dists, idx = knn_point_np(knn, reference_loc, query_loc) dists[dists < 1e-10] = 1e-10 mask = dists > search_radius dists[mask] = 1e+10 weight = 1.0 / dists weight = weight / np.sum(weight, -1, keepdims=True) # [B,N,3] blended_flow = np.sum(reference_flow[idx] * weight.reshape([-1, knn, 1]), axis=1, keepdims=False) mask = mask.sum(axis=1) < 3 return blended_flow, mask def compute_correspondence_loss(self, conf, conf_gt, weight=None): ''' @param conf: [B, L, S] @param conf_gt: [B, L, S] @param weight: [B, L, S] @return: ''' pos_mask = conf_gt == 1 neg_mask = conf_gt == 0 pos_w, neg_w = self.pos_w, self.neg_w #corner case assign a wrong gt if not pos_mask.any(): pos_mask[0, 0, 0] = True if weight is not None: weight[0, 0, 0] = 0. pos_w = 0. if not neg_mask.any(): neg_mask[0, 0, 0] = True if weight is not None: weight[0, 0, 0] = 0. neg_w = 0. # focal loss conf = torch.clamp(conf, 1e-6, 1 - 1e-6) alpha = self.focal_alpha gamma = self.focal_gamma if self.match_type == "dual_softmax": pos_conf = conf[pos_mask] loss_pos = - alpha * torch.pow(1 - pos_conf, gamma) * pos_conf.log() if weight is not None: loss_pos = loss_pos * weight[pos_mask] loss = pos_w * loss_pos.mean() return loss elif self.match_type == "sinkhorn": # no supervision on dustbin row & column. loss_pos = - alpha * torch.pow(1 - conf[pos_mask], gamma) * (conf[pos_mask]).log() loss_neg = - alpha * torch.pow(conf[neg_mask], gamma) * (1 - conf[neg_mask]).log() loss = pos_w * loss_pos.mean() + neg_w * loss_neg.mean() return loss def match_2_conf_matrix(self, matches_gt, matrix_pred): matrix_gt = torch.zeros_like(matrix_pred) for b, match in enumerate (matches_gt) : matrix_gt [ b][ match[0], match[1] ] = 1 return matrix_gt @staticmethod def compute_match_recall(conf_matrix_gt, match_pred) : #, s_pcd, t_pcd, search_radius=0.3): ''' @param conf_matrix_gt: @param match_pred: @return: ''' pred_matrix = torch.zeros_like(conf_matrix_gt) b_ind, src_ind, tgt_ind = match_pred[:, 0], match_pred[:, 1], match_pred[:, 2] pred_matrix[b_ind, src_ind, tgt_ind] = 1. true_positive = (pred_matrix == conf_matrix_gt) * conf_matrix_gt recall = true_positive.sum() / conf_matrix_gt.sum() precision = true_positive.sum() / max(len(match_pred), 1) return recall, precision @staticmethod def ransac_regist_coarse(batched_src_pcd, batched_tgt_pcd, src_mask, tgt_mask, match_pred ): s_len = src_mask.sum(dim=1).int() t_len = tgt_mask.sum(dim=1).int() bsize = len(batched_src_pcd) batched_src_pcd = MatchMotionLoss.tensor2numpy( batched_src_pcd) batched_tgt_pcd = MatchMotionLoss.tensor2numpy( batched_tgt_pcd) match_pred = MatchMotionLoss.tensor2numpy(match_pred) rot = [] trn = [] for i in range(bsize): s_pcd = batched_src_pcd[i][:s_len[i]] t_pcd = batched_tgt_pcd[i][:t_len[i]] pair_i = match_pred[:, 0] == i n_pts = pair_i.sum() if n_pts < 3 : rot.append(torch.eye(3)) trn.append(torch.zeros((3,1))) continue ind = match_pred[pair_i] s_ind, t_ind = ind[:, 1], ind[:, 2] pose = ransac_pose_estimation(s_pcd, t_pcd, [s_ind, t_ind], distance_threshold=0.05) pose = pose.copy() rot.append(torch.from_numpy(pose[:3,:3])) trn.append(torch.from_numpy(pose[:3,3:])) return torch.stack(rot, dim=0 ), torch.stack(trn , dim=0)#ndarray @staticmethod def compute_inlier_ratio(match_pred, data, inlier_thr, s2t_flow=None): s_pcd, t_pcd = data['s_pcd'], data['t_pcd'] #B,N,3 batched_rot = data['batched_rot'] #B,3,3 batched_trn = data['batched_trn'] if s2t_flow is not None: # 4dmatch s_pcd_deformed = s_pcd + s2t_flow s_pcd_wrapped = (torch.matmul(batched_rot, s_pcd_deformed.transpose(1, 2)) + batched_trn).transpose(1,2) else: # 3dmatch s_pcd_wrapped = (torch.matmul(batched_rot, s_pcd.transpose(1, 2)) + batched_trn).transpose(1,2) s_pcd_matched = s_pcd_wrapped [match_pred[:,0], match_pred[:,1]] t_pcd_matched = t_pcd [match_pred[:,0], match_pred[:,2]] inlier = torch.sum( (s_pcd_matched - t_pcd_matched)**2 , dim= 1) < inlier_thr**2 bsize = len(s_pcd) IR=[] for i in range(bsize): pair_i = match_pred[:, 0] == i n_match = pair_i.sum() inlier_i = inlier[pair_i] n_inlier = inlier_i.sum().float() if n_match <3: IR.append( n_match.float()*0) else : IR.append(n_inlier/n_match) return torch.stack(IR, dim=0) @staticmethod def compute_registration_recall(R_est, t_est, data, thr=0.2): bs = len(R_est) success = 0. if data['gt_cov'] is not None: err2 = thr ** 2 gt = np.zeros( (bs, 4, 4)) gt[:, -1,-1] = 1 gt[:, :3, :3] = data['batched_rot'].cpu().numpy() gt[:, :3, 3:] = data['batched_trn'].cpu().numpy() pred = np.zeros((bs, 4, 4)) pred[:, -1, -1] = 1 pred[:, :3, :3] = R_est.detach().cpu().numpy() pred[:, :3, 3:] = t_est.detach().cpu().numpy() for i in range(bs): p = computeTransformationErr( np.linalg.inv(gt[i]) @ pred[i], data['gt_cov'][i]) if p <= err2: success += 1 rr = success / bs return rr else : return 0. @staticmethod def tensor2numpy(tensor): if tensor.requires_grad: tensor=tensor.detach() return tensor.cpu().numpy()
18,271
38.042735
137
py
lepard
lepard-main/models/position_encoding.py
import math import torch from torch import nn class VolumetricPositionEncoding(nn.Module): def __init__(self, config): super().__init__() self.feature_dim = config.feature_dim self.vol_bnds = config.vol_bnds self.voxel_size = config.voxel_size self.vol_origin = self.vol_bnds[0] self.pe_type = config.pe_type def voxelize(self, xyz): ''' @param xyz: B,N,3 @return: B,N,3 ''' if type ( self.vol_origin ) == list : self.vol_origin = torch.FloatTensor(self.vol_origin ).view(1, 1, -1).to( xyz.device ) return (xyz - self.vol_origin) / self.voxel_size @staticmethod def embed_rotary(x, cos, sin): ''' @param x: [B,N,d] @param cos: [B,N,d] [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1] @param sin: [B,N,d] [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1] @return: ''' x2 = torch.stack([-x[..., 1::2], x[..., ::2]], dim=-1).reshape_as(x).contiguous() x = x * cos + x2 * sin return x @staticmethod def embed_pos(pe_type, x, pe): """ combine feature and position code """ if pe_type == 'rotary': return VolumetricPositionEncoding.embed_rotary(x, pe[..., 0], pe[..., 1]) elif pe_type == 'sinusoidal': return x + pe else: raise KeyError() def forward(self, XYZ): ''' @param XYZ: [B,N,3] @return: ''' bsize, npoint, _ = XYZ.shape vox = self.voxelize( XYZ) x_position, y_position, z_position = vox[..., 0:1], vox[...,1:2], vox[...,2:3] div_term = torch.exp( torch.arange(0, self.feature_dim // 3, 2, dtype=torch.float, device=XYZ.device) * (-math.log(10000.0) / (self.feature_dim // 3))) div_term = div_term.view( 1,1, -1) # [1, 1, d//6] sinx = torch.sin(x_position * div_term) # [B, N, d//6] cosx = torch.cos(x_position * div_term) siny = torch.sin(y_position * div_term) cosy = torch.cos(y_position * div_term) sinz = torch.sin(z_position * div_term) cosz = torch.cos(z_position * div_term) if self.pe_type == 'sinusoidal' : position_code = torch.cat( [ sinx, cosx, siny, cosy, sinz, cosz] , dim=-1 ) elif self.pe_type == "rotary" : # sin/cos [θ0,θ1,θ2......θd/6-1] -> sin/cos [θ0,θ0,θ1,θ1,θ2,θ2......θd/6-1,θd/6-1] sinx, cosx, siny, cosy, sinz, cosz = map( lambda feat:torch.stack([feat, feat], dim=-1).view(bsize, npoint, -1), [ sinx, cosx, siny, cosy, sinz, cosz] ) sin_pos = torch.cat([sinx,siny,sinz], dim=-1) cos_pos = torch.cat([cosx,cosy,cosz], dim=-1) position_code = torch.stack( [cos_pos, sin_pos] , dim=-1) else: raise KeyError() if position_code.requires_grad: position_code = position_code.detach() return position_code
2,989
33.367816
160
py
lepard
lepard-main/models/backbone.py
from models.blocks import * import torch.nn.functional as F import numpy as np class KPFCN(nn.Module): def __init__(self, config): super(KPFCN, self).__init__() ############ # Parameters ############ layer = 0 r = config.first_subsampling_dl * config.conv_radius in_dim = config.in_feats_dim out_dim = config.first_feats_dim ##################### # List Encoder blocks ##################### self.encoder_blocks = nn.ModuleList() self.encoder_skip_dims = [] self.encoder_skips = [] # Loop over consecutive blocks for block_i, block in enumerate(config.architecture): # Check equivariance if ('equivariant' in block) and (not out_dim % 3 == 0): raise ValueError('Equivariant block but features dimension is not a factor of 3') # Detect change to next layer for skip connection if np.any([tmp in block for tmp in ['pool', 'strided', 'upsample', 'global']]): self.encoder_skips.append(block_i) self.encoder_skip_dims.append(in_dim) # Detect upsampling block to stop if 'upsample' in block: break # Apply the good block function defining tf ops self.encoder_blocks.append(block_decider(block, r, in_dim, out_dim, layer, config)) # Update dimension of input from output if 'simple' in block: in_dim = out_dim // 2 else: in_dim = out_dim # Detect change to a subsampled layer if 'pool' in block or 'strided' in block: # Update radius and feature dimension for next layer layer += 1 r *= 2 out_dim *= 2 ##################### # bottleneck output & input layer self.coarse_out = nn.Conv1d(in_dim//2, config.coarse_feature_dim, kernel_size=1, bias=True) coarse_in_dim = config.coarse_feature_dim self.coarse_in = nn.Conv1d(coarse_in_dim, in_dim//2, kernel_size=1, bias=True) ##################### # List Decoder blocks ##################### # Save all block operations in a list of modules self.decoder_blocks = nn.ModuleList() self.decoder_concats = [] # Find first upsampling block start_i = 0 for block_i, block in enumerate(config.architecture): if 'upsample' in block: start_i = block_i break # Loop over consecutive blocks for block_i, block in enumerate(config.architecture[start_i:]): # Add dimension of skip connection concat if block_i > 0 and 'upsample' in config.architecture[start_i + block_i - 1]: in_dim += self.encoder_skip_dims[layer] self.decoder_concats.append(block_i) # Apply the good block function defining tf ops self.decoder_blocks.append(block_decider(block, r, in_dim, out_dim, layer, config)) # Update dimension of input from output in_dim = out_dim # Detect change to a subsampled layer if 'upsample' in block: # Update radius and feature dimension for next layer layer -= 1 r *= 0.5 out_dim = out_dim // 2 ##################### # fine output layer ##################### fine_feature_dim = config.fine_feature_dim self.fine_out = nn.Conv1d(out_dim, fine_feature_dim, kernel_size=1, bias=True) def forward(self, batch, phase = 'encode'): # Get input features if phase == 'coarse' : x = batch['features'].clone().detach() # 1. joint encoder part self.skip_x = [] for block_i, block_op in enumerate(self.encoder_blocks): if block_i in self.encoder_skips: self.skip_x.append(x) x = block_op(x, batch) # [N,C] for block_i, block_op in enumerate(self.decoder_blocks): if block_i in self.decoder_concats: x = torch.cat([x, self.skip_x.pop()], dim=1) x = block_op(x, batch) if block_i == 1 : coarse_feats = x.transpose(0,1).unsqueeze(0) #[B, C, N] coarse_feats = self.coarse_out(coarse_feats) #[B, C, N] coarse_feats = coarse_feats.transpose(1,2).squeeze(0) return coarse_feats #[N,C2] # # elif phase == "fine": # # coarse_feats = batch['coarse_feats'] # coarse_feats = coarse_feats.transpose(0,1).unsqueeze(0) # coarse_feats = self.coarse_in(coarse_feats) # x = coarse_feats.transpose(1,2).squeeze(0) # # # for block_i, block_op in enumerate(self.decoder_blocks): # if block_i > 1 : # if block_i in self.decoder_concats: # x = torch.cat([x, self.skip_x.pop()], dim=1) # x = block_op(x, batch) # # fine_feats = x.transpose(0, 1).unsqueeze(0) # [1, C, N] # fine_feats = self.fine_out(fine_feats) # [1, C, N] # fine_feats = fine_feats.transpose(1, 2).squeeze(0) # # return fine_feats
6,033
36.018405
100
py
lepard
lepard-main/models/transformer.py
import copy import math import torch from torch import nn from torch.nn import Module, Dropout from models.position_encoding import VolumetricPositionEncoding as VolPE from models.matching import Matching from models.procrustes import SoftProcrustesLayer import numpy as np import random from scipy.spatial.transform import Rotation class GeometryAttentionLayer(nn.Module): def __init__(self, config): super(GeometryAttentionLayer, self).__init__() d_model = config['feature_dim'] nhead = config['n_head'] self.dim = d_model // nhead self.nhead = nhead self.pe_type = config['pe_type'] # multi-head attention self.q_proj = nn.Linear(d_model, d_model, bias=False) self.k_proj = nn.Linear(d_model, d_model, bias=False) self.v_proj = nn.Linear(d_model, d_model, bias=False) # self.attention = Attention() #LinearAttention() if attention == 'linear' else FullAttention() self.merge = nn.Linear(d_model, d_model, bias=False) # feed-forward network self.mlp = nn.Sequential( nn.Linear(d_model*2, d_model*2, bias=False), nn.ReLU(True), nn.Linear(d_model*2, d_model, bias=False), ) # norm and dropout self.norm1 = nn.LayerNorm(d_model) self.norm2 = nn.LayerNorm(d_model) def forward(self, x, source, x_pe, source_pe, x_mask=None, source_mask=None): bs = x.size(0) q, k, v = x, source, source qp, kvp = x_pe, source_pe q_mask, kv_mask = x_mask, source_mask if self.pe_type == 'sinusoidal': #w(x+p), attention is all you need : https://arxiv.org/abs/1706.03762 if qp is not None: # disentangeld q = q + qp k = k + kvp qw = self.q_proj(q).view(bs, -1, self.nhead, self.dim) # [N, L, (H, D)] kw = self.k_proj(k).view(bs, -1, self.nhead, self.dim) # [N, S, (H, D)] vw = self.v_proj(v).view(bs, -1, self.nhead, self.dim) elif self.pe_type == 'rotary': #Rwx roformer : https://arxiv.org/abs/2104.09864 qw = self.q_proj(q) kw = self.k_proj(k) vw = self.v_proj(v) if qp is not None: # disentangeld q_cos, q_sin = qp[...,0] ,qp[...,1] k_cos, k_sin = kvp[...,0],kvp[...,1] qw = VolPE.embed_rotary(qw, q_cos, q_sin) kw = VolPE.embed_rotary(kw, k_cos, k_sin) qw = qw.view(bs, -1, self.nhead, self.dim) kw = kw.view(bs, -1, self.nhead, self.dim) vw = vw.view(bs, -1, self.nhead, self.dim) else: raise KeyError() # attention a = torch.einsum("nlhd,nshd->nlsh", qw, kw) if kv_mask is not None: a.masked_fill_( q_mask[:, :, None, None] * (~kv_mask[:, None, :, None]), float('-inf')) a = a / qw.size(3) **0.5 a = torch.softmax(a, dim=2) o = torch.einsum("nlsh,nshd->nlhd", a, vw).contiguous() # [N, L, (H, D)] message = self.merge(o.view(bs, -1, self.nhead*self.dim)) # [N, L, C] message = self.norm1(message) # feed-forward network message = self.mlp(torch.cat([x, message], dim=2)) message = self.norm2(message) e = x + message return e class RepositioningTransformer(nn.Module): def __init__(self, config): super(RepositioningTransformer, self).__init__() self.d_model = config['feature_dim'] self.nhead = config['n_head'] self.layer_types = config['layer_types'] self.positioning_type = config['positioning_type'] self.pe_type =config['pe_type'] self.entangled= config['entangled'] self.positional_encoding = VolPE(config) encoder_layer = GeometryAttentionLayer (config) self.layers = nn.ModuleList() for l_type in self.layer_types: if l_type in ['self','cross']: self.layers.append( copy.deepcopy(encoder_layer)) elif l_type == "positioning": if self.positioning_type == 'procrustes': positioning_layer = nn.ModuleList() positioning_layer.append( Matching(config['feature_matching'])) positioning_layer.append( SoftProcrustesLayer(config['procrustes']) ) self.layers.append(positioning_layer) elif self.positioning_type in ['oracle', 'randSO3']: self.layers.append( None) else : raise KeyError(self.positioning_type + " undefined positional encoding type") else: raise KeyError() self._reset_parameters() def forward(self, src_feat, tgt_feat, s_pcd, t_pcd, src_mask, tgt_mask, data, T = None, timers = None): self.timers = timers assert self.d_model == src_feat.size(2), "the feature number of src and transformer must be equal" if T is not None: R, t = T src_pcd_wrapped = (torch.matmul(R, s_pcd.transpose(1, 2)) + t).transpose(1, 2) tgt_pcd_wrapped = t_pcd else: src_pcd_wrapped = s_pcd tgt_pcd_wrapped = t_pcd src_pe = self.positional_encoding( src_pcd_wrapped) tgt_pe = self.positional_encoding( tgt_pcd_wrapped) if not self.entangled: position_layer = 0 data.update({"position_layers":{}}) for layer, name in zip(self.layers, self.layer_types) : if name == 'self': if self.timers: self.timers.tic('self atten') src_feat = layer(src_feat, src_feat, src_pe, src_pe, src_mask, src_mask,) tgt_feat = layer(tgt_feat, tgt_feat, tgt_pe, tgt_pe, tgt_mask, tgt_mask) if self.timers: self.timers.toc('self atten') elif name == 'cross': if self.timers: self.timers.tic('cross atten') src_feat = layer(src_feat, tgt_feat, src_pe, tgt_pe, src_mask, tgt_mask) tgt_feat = layer(tgt_feat, src_feat, tgt_pe, src_pe, tgt_mask, src_mask) if self.timers: self.timers.toc('cross atten') elif name =='positioning': if self.positioning_type == 'procrustes': conf_matrix, match_pred = layer[0](src_feat, tgt_feat, src_pe, tgt_pe, src_mask, tgt_mask, data, pe_type=self.pe_type) position_layer += 1 data["position_layers"][position_layer] = {"conf_matrix": conf_matrix, "match_pred": match_pred} if self.timers: self.timers.tic('procrustes_layer') R, t, R_forwd, t_forwd, condition, solution_mask = layer[1] (conf_matrix, s_pcd, t_pcd, src_mask, tgt_mask) if self.timers: self.timers.toc('procrustes_layer') data["position_layers"][position_layer].update({ "R_s2t_pred": R,"t_s2t_pred": t, "solution_mask": solution_mask, "condition": condition}) src_pcd_wrapped = (torch.matmul(R_forwd, s_pcd.transpose(1, 2)) + t_forwd).transpose(1, 2) tgt_pcd_wrapped = t_pcd src_pe = self.positional_encoding(src_pcd_wrapped) tgt_pe = self.positional_encoding(tgt_pcd_wrapped) elif self.positioning_type == 'randSO3': src_pcd_wrapped = self.rand_rot_pcd( s_pcd, src_mask) tgt_pcd_wrapped = t_pcd src_pe = self.positional_encoding(src_pcd_wrapped) tgt_pe = self.positional_encoding(tgt_pcd_wrapped) elif self.positioning_type == 'oracle': #Note R,t ground truth is only available for computing oracle position encoding rot_gt = data['batched_rot'] trn_gt = data['batched_trn'] src_pcd_wrapped = (torch.matmul(rot_gt, s_pcd.transpose(1, 2)) + trn_gt).transpose(1, 2) tgt_pcd_wrapped = t_pcd src_pe = self.positional_encoding(src_pcd_wrapped) tgt_pe = self.positional_encoding(tgt_pcd_wrapped) else: raise KeyError(self.positioning_type + " undefined positional encoding type") else : raise KeyError return src_feat, tgt_feat, src_pe, tgt_pe else : # pos. fea. entangeled position_layer = 0 data.update({"position_layers":{}}) src_feat = VolPE.embed_pos(self.pe_type, src_feat, src_pe) tgt_feat = VolPE.embed_pos(self.pe_type, tgt_feat, tgt_pe) for layer, name in zip(self.layers, self.layer_types): if name == 'self': if self.timers: self.timers.tic('self atten') src_feat = layer(src_feat, src_feat, None, None, src_mask, src_mask, ) tgt_feat = layer(tgt_feat, tgt_feat, None, None, tgt_mask, tgt_mask) if self.timers: self.timers.toc('self atten') elif name == 'cross': if self.timers: self.timers.tic('cross atten') src_feat = layer(src_feat, tgt_feat, None, None, src_mask, tgt_mask) tgt_feat = layer(tgt_feat, src_feat, None, None, tgt_mask, src_mask) if self.timers: self.timers.toc('cross atten') elif name == 'positioning': pass return src_feat, tgt_feat, src_pe, tgt_pe def rand_rot_pcd (self, pcd, mask): ''' @param pcd: B, N, 3 @param mask: B, N @return: ''' pcd[~mask]=0. N = mask.shape[1] n_points = mask.sum(dim=1, keepdim=True).view(-1,1,1) bs = pcd.shape[0] euler_ab = np.random.rand(bs, 3) * np.pi * 2 # anglez, angley, anglex rand_rot = torch.from_numpy( Rotation.from_euler('zyx', euler_ab).as_matrix() ).to(pcd) pcd_u = pcd.mean(dim=1, keepdim=True) * N / n_points pcd_centered = pcd - pcd_u pcd_rand_rot = torch.matmul( rand_rot, pcd_centered.transpose(1,2) ).transpose(1,2) + pcd_u return pcd_rand_rot def _reset_parameters(self): for p in self.parameters(): if p.dim() > 1: nn.init.xavier_uniform_(p)
10,666
36.559859
142
py
lepard
lepard-main/models/procrustes.py
import torch import torch.nn as nn def topk(data, num_topk): sort, idx = data.sort(descending=True) return sort[:num_topk], idx[:num_topk] class SoftProcrustesLayer(nn.Module): def __init__(self, config): super(SoftProcrustesLayer, self).__init__() self.sample_rate = config.sample_rate self.max_condition_num= config.max_condition_num @staticmethod def batch_weighted_procrustes( X, Y, w, eps=0.0001): ''' @param X: source frame [B, N,3] @param Y: target frame [B, N,3] @param w: weights [B, N,1] @param eps: @return: ''' # https://ieeexplore.ieee.org/document/88573 bsize = X.shape[0] device = X.device W1 = torch.abs(w).sum(dim=1, keepdim=True) w_norm = w / (W1 + eps) mean_X = (w_norm * X).sum(dim=1, keepdim=True) mean_Y = (w_norm * Y).sum(dim=1, keepdim=True) Sxy = torch.matmul( (Y - mean_Y).transpose(1,2), w_norm * (X - mean_X) ) Sxy = Sxy.cpu().double() U, D, V = Sxy.svd() # small SVD runs faster on cpu condition = D.max(dim=1)[0] / D.min(dim=1)[0] S = torch.eye(3)[None].repeat(bsize,1,1).double() UV_det = U.det() * V.det() S[:, 2:3, 2:3] = UV_det.view(-1, 1,1) svT = torch.matmul( S, V.transpose(1,2) ) R = torch.matmul( U, svT).float().to(device) t = mean_Y.transpose(1,2) - torch.matmul( R, mean_X.transpose(1,2) ) return R, t, condition def forward(self, conf_matrix, src_pcd, tgt_pcd, src_mask, tgt_mask): ''' @param conf_matrix: @param src_pcd: @param tgt_pcd: @param src_mask: @param tgt_mask: @return: ''' bsize, N, M = conf_matrix.shape # subsample correspondence src_len = src_mask.sum(dim=1) tgt_len = tgt_mask.sum(dim=1) entry_max, _ = torch.stack([src_len,tgt_len], dim=0).max(dim=0) entry_max = (entry_max * self.sample_rate).int() sample_n_points = entry_max.float().mean().int() #entry_max.max() conf, idx = conf_matrix.view(bsize, -1).sort(descending=True,dim=1) w = conf [:, :sample_n_points] idx= idx[:, :sample_n_points] idx_src = idx//M #torch.div(idx, M, rounding_mode='trunc') idx_tgt = idx%M b_index = torch.arange(bsize).view(-1, 1).repeat((1, sample_n_points)).view(-1) src_pcd_sampled = src_pcd[b_index, idx_src.view(-1)].view(bsize, sample_n_points, -1) tgt_pcd_sampled = tgt_pcd[b_index, idx_tgt.view(-1)].view(bsize, sample_n_points, -1) w_mask = torch.arange(sample_n_points).view(1,-1).repeat(bsize,1).to(w) w_mask = w_mask < entry_max[:,None] w[~w_mask] = 0. # solve try : R, t, condition = self.batch_weighted_procrustes(src_pcd_sampled, tgt_pcd_sampled, w[...,None]) except: # fail to get valid solution, this usually happens at the early stage of training R = torch.eye(3)[None].repeat(bsize,1,1).type_as(conf_matrix) t = torch.zeros(3, 1)[None].repeat(bsize,1,1).type_as(conf_matrix) condition = torch.zeros(bsize).type_as(conf_matrix) #filter unreliable solution with condition nnumber solution_mask = condition < self.max_condition_num R_forwd = R.clone() t_forwd = t.clone() R_forwd[~solution_mask] = torch.eye(3).type_as(R) t_forwd[~solution_mask] = torch.zeros(3, 1).type_as(R) return R, t, R_forwd, t_forwd, condition, solution_mask
3,591
37.623656
107
py
lepard
lepard-main/models/pipeline.py
from models.blocks import * from models.backbone import KPFCN from models.transformer import RepositioningTransformer from models.matching import Matching from models.procrustes import SoftProcrustesLayer class Pipeline(nn.Module): def __init__(self, config): super(Pipeline, self).__init__() self.config = config self.backbone = KPFCN(config['kpfcn_config']) self.pe_type = config['coarse_transformer']['pe_type'] self.positioning_type = config['coarse_transformer']['positioning_type'] self.coarse_transformer = RepositioningTransformer(config['coarse_transformer']) self.coarse_matching = Matching(config['coarse_matching']) self.soft_procrustes = SoftProcrustesLayer(config['coarse_transformer']['procrustes']) def forward(self, data, timers=None): self.timers = timers if self.timers: self.timers.tic('kpfcn backbone encode') coarse_feats = self.backbone(data, phase="coarse") if self.timers: self.timers.toc('kpfcn backbone encode') if self.timers: self.timers.tic('coarse_preprocess') src_feats, tgt_feats, s_pcd, t_pcd, src_mask, tgt_mask = self.split_feats (coarse_feats, data) data.update({ 's_pcd': s_pcd, 't_pcd': t_pcd }) if self.timers: self.timers.toc('coarse_preprocess') if self.timers: self.timers.tic('coarse feature transformer') src_feats, tgt_feats, src_pe, tgt_pe = self.coarse_transformer(src_feats, tgt_feats, s_pcd, t_pcd, src_mask, tgt_mask, data, timers=timers) if self.timers: self.timers.toc('coarse feature transformer') if self.timers: self.timers.tic('match feature coarse') conf_matrix_pred, coarse_match_pred = self.coarse_matching(src_feats, tgt_feats, src_pe, tgt_pe, src_mask, tgt_mask, data, pe_type = self.pe_type) data.update({'conf_matrix_pred': conf_matrix_pred, 'coarse_match_pred': coarse_match_pred }) if self.timers: self.timers.toc('match feature coarse') if self.timers: self.timers.tic('procrustes_layer') R, t, _, _, _, _ = self.soft_procrustes(conf_matrix_pred, s_pcd, t_pcd, src_mask, tgt_mask) data.update({"R_s2t_pred": R, "t_s2t_pred": t}) if self.timers: self.timers.toc('procrustes_layer') return data def split_feats(self, geo_feats, data): pcd = data['points'][self.config['kpfcn_config']['coarse_level']] src_mask = data['src_mask'] tgt_mask = data['tgt_mask'] src_ind_coarse_split = data[ 'src_ind_coarse_split'] tgt_ind_coarse_split = data['tgt_ind_coarse_split'] src_ind_coarse = data['src_ind_coarse'] tgt_ind_coarse = data['tgt_ind_coarse'] b_size, src_pts_max = src_mask.shape tgt_pts_max = tgt_mask.shape[1] src_feats = torch.zeros([b_size * src_pts_max, geo_feats.shape[-1]]).type_as(geo_feats) tgt_feats = torch.zeros([b_size * tgt_pts_max, geo_feats.shape[-1]]).type_as(geo_feats) src_pcd = torch.zeros([b_size * src_pts_max, 3]).type_as(pcd) tgt_pcd = torch.zeros([b_size * tgt_pts_max, 3]).type_as(pcd) src_feats[src_ind_coarse_split] = geo_feats[src_ind_coarse] tgt_feats[tgt_ind_coarse_split] = geo_feats[tgt_ind_coarse] src_pcd[src_ind_coarse_split] = pcd[src_ind_coarse] tgt_pcd[tgt_ind_coarse_split] = pcd[tgt_ind_coarse] return src_feats.view( b_size , src_pts_max , -1), \ tgt_feats.view( b_size , tgt_pts_max , -1), \ src_pcd.view( b_size , src_pts_max , -1), \ tgt_pcd.view( b_size , tgt_pts_max , -1), \ src_mask, \ tgt_mask
3,685
43.95122
154
py
lepard
lepard-main/models/blocks.py
import time import math import torch import torch.nn as nn from torch.nn.parameter import Parameter from torch.nn.init import kaiming_uniform_ from kernels.kernel_points import load_kernels # from lib.ply import write_ply def gather(x, idx, method=2): """ implementation of a custom gather operation for faster backwards. :param x: input with shape [N, D_1, ... D_d] :param idx: indexing with shape [n_1, ..., n_m] :param method: Choice of the method :return: x[idx] with shape [n_1, ..., n_m, D_1, ... D_d] """ if method == 0: return x[idx] elif method == 1: x = x.unsqueeze(1) x = x.expand((-1, idx.shape[-1], -1)) idx = idx.unsqueeze(2) idx = idx.expand((-1, -1, x.shape[-1])) return x.gather(0, idx) elif method == 2: for i, ni in enumerate(idx.size()[1:]): x = x.unsqueeze(i+1) new_s = list(x.size()) new_s[i+1] = ni x = x.expand(new_s) n = len(idx.size()) for i, di in enumerate(x.size()[n:]): idx = idx.unsqueeze(i+n) new_s = list(idx.size()) new_s[i+n] = di idx = idx.expand(new_s) return x.gather(0, idx) else: raise ValueError('Unkown method') def radius_gaussian(sq_r, sig, eps=1e-9): """ Compute a radius gaussian (gaussian of distance) :param sq_r: input radiuses [dn, ..., d1, d0] :param sig: extents of gaussians [d1, d0] or [d0] or float :return: gaussian of sq_r [dn, ..., d1, d0] """ return torch.exp(-sq_r / (2 * sig**2 + eps)) def closest_pool(x, inds): """ Pools features from the closest neighbors. WARNING: this function assumes the neighbors are ordered. :param x: [n1, d] features matrix :param inds: [n2, max_num] Only the first column is used for pooling :return: [n2, d] pooled features matrix """ # Add a last row with minimum features for shadow pools x = torch.cat((x, torch.zeros_like(x[:1, :])), 0) # Get features for each pooling location [n2, d] return gather(x, inds[:, 0]) def max_pool(x, inds): """ Pools features with the maximum values. :param x: [n1, d] features matrix :param inds: [n2, max_num] pooling indices :return: [n2, d] pooled features matrix """ # Add a last row with minimum features for shadow pools x = torch.cat((x, torch.zeros_like(x[:1, :])), 0) # Get all features for each pooling location [n2, max_num, d] pool_features = gather(x, inds) # Pool the maximum [n2, d] max_features, _ = torch.max(pool_features, 1) return max_features def global_average(x, batch_lengths): """ Block performing a global average over batch pooling :param x: [N, D] input features :param batch_lengths: [B] list of batch lengths :return: [B, D] averaged features """ # Loop over the clouds of the batch averaged_features = [] i0 = 0 for b_i, length in enumerate(batch_lengths): # Average features for each batch cloud averaged_features.append(torch.mean(x[i0:i0 + length], dim=0)) # Increment for next cloud i0 += length # Average features in each batch return torch.stack(averaged_features) # ---------------------------------------------------------------------------------------------------------------------- # # KPConv class # \******************/ # class KPConv(nn.Module): def __init__(self, kernel_size, p_dim, in_channels, out_channels, KP_extent, radius, fixed_kernel_points='center', KP_influence='linear', aggregation_mode='sum', deformable=False, modulated=False): """ Initialize parameters for KPConvDeformable. :param kernel_size: Number of kernel points. :param p_dim: dimension of the point space. :param in_channels: dimension of input features. :param out_channels: dimension of output features. :param KP_extent: influence radius of each kernel point. :param radius: radius used for kernel point init. Even for deformable, use the config.conv_radius :param fixed_kernel_points: fix position of certain kernel points ('none', 'center' or 'verticals'). :param KP_influence: influence function of the kernel points ('constant', 'linear', 'gaussian'). :param aggregation_mode: choose to sum influences, or only keep the closest ('closest', 'sum'). :param deformable: choose deformable or not :param modulated: choose if kernel weights are modulated in addition to deformed """ super(KPConv, self).__init__() # Save parameters self.K = kernel_size self.p_dim = p_dim self.in_channels = in_channels self.out_channels = out_channels self.radius = radius self.KP_extent = KP_extent self.fixed_kernel_points = fixed_kernel_points self.KP_influence = KP_influence self.aggregation_mode = aggregation_mode self.deformable = deformable self.modulated = modulated # Running variable containing deformed KP distance to input points. (used in regularization loss) self.min_d2 = None self.deformed_KP = None self.offset_features = None # Initialize weights self.weights = Parameter(torch.zeros((self.K, in_channels, out_channels), dtype=torch.float32), requires_grad=True) # Initiate weights for offsets if deformable: if modulated: self.offset_dim = (self.p_dim + 1) * self.K else: self.offset_dim = self.p_dim * self.K self.offset_conv = KPConv(self.K, self.p_dim, self.in_channels, self.offset_dim, KP_extent, radius, fixed_kernel_points=fixed_kernel_points, KP_influence=KP_influence, aggregation_mode=aggregation_mode) self.offset_bias = Parameter(torch.zeros(self.offset_dim, dtype=torch.float32), requires_grad=True) else: self.offset_dim = None self.offset_conv = None self.offset_bias = None # Reset parameters self.reset_parameters() # Initialize kernel points self.kernel_points = self.init_KP() return def reset_parameters(self): kaiming_uniform_(self.weights, a=math.sqrt(5)) if self.deformable: nn.init.zeros_(self.offset_bias) return def init_KP(self): """ Initialize the kernel point positions in a sphere :return: the tensor of kernel points """ # Create one kernel disposition (as numpy array). Choose the KP distance to center thanks to the KP extent K_points_numpy = load_kernels(self.radius, self.K, dimension=self.p_dim, fixed=self.fixed_kernel_points) return Parameter(torch.tensor(K_points_numpy, dtype=torch.float32), requires_grad=False) def forward(self, q_pts, s_pts, neighb_inds, x): ################### # Offset generation ################### if self.deformable: # Get offsets with a KPConv that only takes part of the features self.offset_features = self.offset_conv(q_pts, s_pts, neighb_inds, x) + self.offset_bias if self.modulated: # Get offset (in normalized scale) from features unscaled_offsets = self.offset_features[:, :self.p_dim * self.K] unscaled_offsets = unscaled_offsets.view(-1, self.K, self.p_dim) # Get modulations modulations = 2 * torch.sigmoid(self.offset_features[:, self.p_dim * self.K:]) else: # Get offset (in normalized scale) from features unscaled_offsets = self.offset_features.view(-1, self.K, self.p_dim) # No modulations modulations = None # Rescale offset for this layer offsets = unscaled_offsets * self.KP_extent else: offsets = None modulations = None ###################### # Deformed convolution ###################### # Add a fake point in the last row for shadow neighbors s_pts = torch.cat((s_pts, torch.zeros_like(s_pts[:1, :]) + 1e6), 0) # Get neighbor points [n_points, n_neighbors, dim] neighbors = s_pts[neighb_inds, :] # Center every neighborhood neighbors = neighbors - q_pts.unsqueeze(1) # Apply offsets to kernel points [n_points, n_kpoints, dim] if self.deformable: self.deformed_KP = offsets + self.kernel_points deformed_K_points = self.deformed_KP.unsqueeze(1) else: deformed_K_points = self.kernel_points # Get all difference matrices [n_points, n_neighbors, n_kpoints, dim] neighbors.unsqueeze_(2) differences = neighbors - deformed_K_points # Get the square distances [n_points, n_neighbors, n_kpoints] sq_distances = torch.sum(differences ** 2, dim=3) # Optimization by ignoring points outside a deformed KP range if self.deformable: # Save distances for loss self.min_d2, _ = torch.min(sq_distances, dim=1) # Boolean of the neighbors in range of a kernel point [n_points, n_neighbors] in_range = torch.any(sq_distances < self.KP_extent ** 2, dim=2).type(torch.int32) # New value of max neighbors new_max_neighb = torch.max(torch.sum(in_range, dim=1)) # For each row of neighbors, indices of the ones that are in range [n_points, new_max_neighb] neighb_row_bool, neighb_row_inds = torch.topk(in_range, new_max_neighb.item(), dim=1) # Gather new neighbor indices [n_points, new_max_neighb] new_neighb_inds = neighb_inds.gather(1, neighb_row_inds, sparse_grad=False) # Gather new distances to KP [n_points, new_max_neighb, n_kpoints] neighb_row_inds.unsqueeze_(2) neighb_row_inds = neighb_row_inds.expand(-1, -1, self.K) sq_distances = sq_distances.gather(1, neighb_row_inds, sparse_grad=False) # New shadow neighbors have to point to the last shadow point new_neighb_inds *= neighb_row_bool new_neighb_inds -= (neighb_row_bool.type(torch.int64) - 1) * int(s_pts.shape[0] - 1) else: new_neighb_inds = neighb_inds # Get Kernel point influences [n_points, n_kpoints, n_neighbors] if self.KP_influence == 'constant': # Every point get an influence of 1. all_weights = torch.ones_like(sq_distances) all_weights = torch.transpose(all_weights, 1, 2) elif self.KP_influence == 'linear': # Influence decrease linearly with the distance, and get to zero when d = KP_extent. all_weights = torch.clamp(1 - torch.sqrt(sq_distances) / self.KP_extent, min=0.0) all_weights = torch.transpose(all_weights, 1, 2) elif self.KP_influence == 'gaussian': # Influence in gaussian of the distance. sigma = self.KP_extent * 0.3 all_weights = radius_gaussian(sq_distances, sigma) all_weights = torch.transpose(all_weights, 1, 2) else: raise ValueError('Unknown influence function type (config.KP_influence)') # In case of closest mode, only the closest KP can influence each point if self.aggregation_mode == 'closest': neighbors_1nn = torch.argmin(sq_distances, dim=2) all_weights *= torch.transpose(nn.functional.one_hot(neighbors_1nn, self.K), 1, 2) elif self.aggregation_mode != 'sum': raise ValueError("Unknown convolution mode. Should be 'closest' or 'sum'") # Add a zero feature for shadow neighbors x = torch.cat((x, torch.zeros_like(x[:1, :])), 0) # Get the features of each neighborhood [n_points, n_neighbors, in_fdim] neighb_x = gather(x, new_neighb_inds) # Apply distance weights [n_points, n_kpoints, in_fdim] weighted_features = torch.matmul(all_weights, neighb_x) # Apply modulations if self.deformable and self.modulated: weighted_features *= modulations.unsqueeze(2) # Apply network weights [n_kpoints, n_points, out_fdim] weighted_features = weighted_features.permute((1, 0, 2)) kernel_outputs = torch.matmul(weighted_features, self.weights) # Convolution sum [n_points, out_fdim] # return torch.sum(kernel_outputs, dim=0) output_features = torch.sum(kernel_outputs, dim=0, keepdim=False) # normalization term. neighbor_features_sum = torch.sum(neighb_x, dim=-1) neighbor_num = torch.sum(torch.gt(neighbor_features_sum, 0.0), dim=-1) neighbor_num = torch.max(neighbor_num, torch.ones_like(neighbor_num)) output_features = output_features / neighbor_num.unsqueeze(1) return output_features def __repr__(self): return 'KPConv(radius: {:.2f}, extent: {:.2f}, in_feat: {:d}, out_feat: {:d})'.format(self.radius, self.KP_extent, self.in_channels, self.out_channels) # ---------------------------------------------------------------------------------------------------------------------- # # Complex blocks # \********************/ # def block_decider(block_name, radius, in_dim, out_dim, layer_ind, config): if block_name == 'unary': return UnaryBlock(in_dim, out_dim, config.use_batch_norm, config.batch_norm_momentum) elif block_name in ['simple', 'simple_deformable', 'simple_invariant', 'simple_equivariant', 'simple_strided', 'simple_deformable_strided', 'simple_invariant_strided', 'simple_equivariant_strided']: return SimpleBlock(block_name, in_dim, out_dim, radius, layer_ind, config) elif block_name in ['resnetb', 'resnetb_invariant', 'resnetb_equivariant', 'resnetb_deformable', 'resnetb_strided', 'resnetb_deformable_strided', 'resnetb_equivariant_strided', 'resnetb_invariant_strided']: return ResnetBottleneckBlock(block_name, in_dim, out_dim, radius, layer_ind, config) elif block_name == 'max_pool' or block_name == 'max_pool_wide': return MaxPoolBlock(layer_ind) elif block_name == 'global_average': return GlobalAverageBlock() elif block_name == 'nearest_upsample': return NearestUpsampleBlock(layer_ind) else: raise ValueError('Unknown block name in the architecture definition : ' + block_name) class BatchNormBlock(nn.Module): def __init__(self, in_dim, use_bn, bn_momentum): """ Initialize a batch normalization block. If network does not use batch normalization, replace with biases. :param in_dim: dimension input features :param use_bn: boolean indicating if we use Batch Norm :param bn_momentum: Batch norm momentum """ super(BatchNormBlock, self).__init__() self.bn_momentum = bn_momentum self.use_bn = use_bn self.in_dim = in_dim if self.use_bn: #self.batch_norm = nn.BatchNorm1d(in_dim, momentum=bn_momentum) self.batch_norm = nn.InstanceNorm1d(in_dim, momentum=bn_momentum) else: self.bias = Parameter(torch.zeros(in_dim, dtype=torch.float32), requires_grad=True) return def reset_parameters(self): nn.init.zeros_(self.bias) def forward(self, x): if self.use_bn: x = x.unsqueeze(2) x = x.transpose(0, 2) x = self.batch_norm(x) x = x.transpose(0, 2) return x.squeeze() else: return x + self.bias def __repr__(self): return 'BatchNormBlock(in_feat: {:d}, momentum: {:.3f}, only_bias: {:s})'.format(self.in_dim, self.bn_momentum, str(not self.use_bn)) class UnaryBlock(nn.Module): def __init__(self, in_dim, out_dim, use_bn, bn_momentum, no_relu=False): """ Initialize a standard unary block with its ReLU and BatchNorm. :param in_dim: dimension input features :param out_dim: dimension input features :param use_bn: boolean indicating if we use Batch Norm :param bn_momentum: Batch norm momentum """ super(UnaryBlock, self).__init__() self.bn_momentum = bn_momentum self.use_bn = use_bn self.no_relu = no_relu self.in_dim = in_dim self.out_dim = out_dim self.mlp = nn.Linear(in_dim, out_dim, bias=False) self.batch_norm = BatchNormBlock(out_dim, self.use_bn, self.bn_momentum) if not no_relu: self.leaky_relu = nn.LeakyReLU(0.1) return def forward(self, x, batch=None): x = self.mlp(x) x = self.batch_norm(x) if not self.no_relu: x = self.leaky_relu(x) return x def __repr__(self): return 'UnaryBlock(in_feat: {:d}, out_feat: {:d}, BN: {:s}, ReLU: {:s})'.format(self.in_dim, self.out_dim, str(self.use_bn), str(not self.no_relu)) class LastUnaryBlock(nn.Module): def __init__(self, in_dim, out_dim, use_bn, bn_momentum, no_relu=False): """ Initialize a standard last_unary block without BN, ReLU. :param in_dim: dimension input features :param out_dim: dimension input features :param use_bn: boolean indicating if we use Batch Norm :param bn_momentum: Batch norm momentum """ super(LastUnaryBlock, self).__init__() self.in_dim = in_dim self.out_dim = out_dim self.mlp = nn.Linear(in_dim, out_dim, bias=False) return def forward(self, x, batch=None): x = self.mlp(x) return x def __repr__(self): return 'LastUnaryBlock(in_feat: {:d}, out_feat: {:d})'.format(self.in_dim, self.out_dim) class SimpleBlock(nn.Module): def __init__(self, block_name, in_dim, out_dim, radius, layer_ind, config): """ Initialize a simple convolution block with its ReLU and BatchNorm. :param in_dim: dimension input features :param out_dim: dimension input features :param radius: current radius of convolution :param config: parameters """ super(SimpleBlock, self).__init__() # get KP_extent from current radius current_extent = radius * config.KP_extent / config.conv_radius # Get other parameters self.bn_momentum = config.batch_norm_momentum self.use_bn = config.use_batch_norm self.layer_ind = layer_ind self.block_name = block_name self.in_dim = in_dim self.out_dim = out_dim # Define the KPConv class self.KPConv = KPConv(config.num_kernel_points, config.in_points_dim, in_dim, out_dim // 2, current_extent, radius, fixed_kernel_points=config.fixed_kernel_points, KP_influence=config.KP_influence, aggregation_mode=config.aggregation_mode, deformable='deform' in block_name, modulated=config.modulated) # Other opperations self.batch_norm = BatchNormBlock(out_dim // 2, self.use_bn, self.bn_momentum) self.leaky_relu = nn.LeakyReLU(0.1) return def forward(self, x, batch): if 'strided' in self.block_name: q_pts = batch['points'][self.layer_ind + 1] s_pts = batch['points'][self.layer_ind] neighb_inds = batch['pools'][self.layer_ind] else: q_pts = batch['points'][self.layer_ind] s_pts = batch['points'][self.layer_ind] neighb_inds = batch['neighbors'][self.layer_ind] x = self.KPConv(q_pts, s_pts, neighb_inds, x) return self.leaky_relu(self.batch_norm(x)) class ResnetBottleneckBlock(nn.Module): def __init__(self, block_name, in_dim, out_dim, radius, layer_ind, config): """ Initialize a resnet bottleneck block. :param in_dim: dimension input features :param out_dim: dimension input features :param radius: current radius of convolution :param config: parameters """ super(ResnetBottleneckBlock, self).__init__() # get KP_extent from current radius current_extent = radius * config.KP_extent / config.conv_radius # Get other parameters self.bn_momentum = config.batch_norm_momentum self.use_bn = config.use_batch_norm self.block_name = block_name self.layer_ind = layer_ind self.in_dim = in_dim self.out_dim = out_dim # First downscaling mlp if in_dim != out_dim // 4: self.unary1 = UnaryBlock(in_dim, out_dim // 4, self.use_bn, self.bn_momentum) else: self.unary1 = nn.Identity() # KPConv block self.KPConv = KPConv(config.num_kernel_points, config.in_points_dim, out_dim // 4, out_dim // 4, current_extent, radius, fixed_kernel_points=config.fixed_kernel_points, KP_influence=config.KP_influence, aggregation_mode=config.aggregation_mode, deformable='deform' in block_name, modulated=config.modulated) self.batch_norm_conv = BatchNormBlock(out_dim // 4, self.use_bn, self.bn_momentum) # Second upscaling mlp self.unary2 = UnaryBlock(out_dim // 4, out_dim, self.use_bn, self.bn_momentum, no_relu=True) # Shortcut optional mpl if in_dim != out_dim: self.unary_shortcut = UnaryBlock(in_dim, out_dim, self.use_bn, self.bn_momentum, no_relu=True) else: self.unary_shortcut = nn.Identity() # Other operations self.leaky_relu = nn.LeakyReLU(0.1) return def forward(self, features, batch): if 'strided' in self.block_name: q_pts = batch['points'][self.layer_ind + 1] s_pts = batch['points'][self.layer_ind] neighb_inds = batch['pools'][self.layer_ind] else: q_pts = batch['points'][self.layer_ind] s_pts = batch['points'][self.layer_ind] neighb_inds = batch['neighbors'][self.layer_ind] # First downscaling mlp x = self.unary1(features) # Convolution x = self.KPConv(q_pts, s_pts, neighb_inds, x) x = self.leaky_relu(self.batch_norm_conv(x)) # Second upscaling mlp x = self.unary2(x) # Shortcut if 'strided' in self.block_name: shortcut = max_pool(features, neighb_inds) else: shortcut = features shortcut = self.unary_shortcut(shortcut) return self.leaky_relu(x + shortcut) class GlobalAverageBlock(nn.Module): def __init__(self): """ Initialize a global average block with its ReLU and BatchNorm. """ super(GlobalAverageBlock, self).__init__() return def forward(self, x, batch): return global_average(x, batch['stack_lengths'][-1]) class NearestUpsampleBlock(nn.Module): def __init__(self, layer_ind): """ Initialize a nearest upsampling block with its ReLU and BatchNorm. """ super(NearestUpsampleBlock, self).__init__() self.layer_ind = layer_ind return def forward(self, x, batch): return closest_pool(x, batch['upsamples'][self.layer_ind - 1]) def __repr__(self): return 'NearestUpsampleBlock(layer: {:d} -> {:d})'.format(self.layer_ind, self.layer_ind - 1) class MaxPoolBlock(nn.Module): def __init__(self, layer_ind): """ Initialize a max pooling block with its ReLU and BatchNorm. """ super(MaxPoolBlock, self).__init__() self.layer_ind = layer_ind return def forward(self, x, batch): return max_pool(x, batch['pools'][self.layer_ind + 1])
26,090
35.956091
122
py
lepard
lepard-main/datasets/_4dmatch.py
import os, sys, glob, torch # sys.path.append("../") [sys.path.append(i) for i in ['.', '..']] import numpy as np import torch import random from scipy.spatial.transform import Rotation from torch.utils.data import Dataset from lib.benchmark_utils import to_o3d_pcd, to_tsfm, KDTree_corr from lib.utils import load_obj HMN_intrin = np.array( [443, 256, 443, 250 ]) cam_intrin = np.array( [443, 256, 443, 250 ]) from lib.benchmark_utils import to_o3d_pcd, to_tsfm, get_correspondences class _4DMatch(Dataset): def __init__(self, config, split, data_augmentation=True): super(_4DMatch, self).__init__() assert split in ['train','val','test'] if 'overfit' in config.exp_dir: d_slice = config.batch_size else : d_slice = None self.entries = self.read_entries( config.split[split] , config.data_root, d_slice=d_slice ) self.base_dir = config.data_root self.data_augmentation = data_augmentation self.config = config self.rot_factor = 1. self.augment_noise = config.augment_noise self.max_points = 30000 self.overlap_radius = 0.0375 self.cache = {} self.cache_size = 30000 def read_entries (self, split, data_root, d_slice=None, shuffle= False): entries = glob.glob(os.path.join(data_root, split, "*/*.npz")) if shuffle: random.shuffle(entries) if d_slice: return entries[:d_slice] return entries def __len__(self): return len(self.entries ) def __getitem__(self, index, debug=False): if index in self.cache: entry = self.cache[index] else : entry = np.load(self.entries[index]) if len(self.cache) < self.cache_size: self.cache[index] = entry # get transformation rot = entry['rot'] trans = entry['trans'] s2t_flow = entry['s2t_flow'] src_pcd = entry['s_pc'] tgt_pcd = entry['t_pc'] correspondences = entry['correspondences'] # obtained with search radius 0.015 m src_pcd_deformed = src_pcd + s2t_flow if "metric_index" in entry: metric_index = entry['metric_index'].squeeze() else: metric_index = None # if we get too many points, we do some downsampling if (src_pcd.shape[0] > self.max_points): idx = np.random.permutation(src_pcd.shape[0])[:self.max_points] src_pcd = src_pcd[idx] if (tgt_pcd.shape[0] > self.max_points): idx = np.random.permutation(tgt_pcd.shape[0])[:self.max_points] tgt_pcd = tgt_pcd[idx] if debug: import mayavi.mlab as mlab c_red = (224. / 255., 0 / 255., 125 / 255.) c_pink = (224. / 255., 75. / 255., 232. / 255.) c_blue = (0. / 255., 0. / 255., 255. / 255.) scale_factor = 0.013 src_wrapped = (np.matmul( rot, src_pcd_deformed.T ) + trans ).T mlab.points3d(src_wrapped[:, 0], src_wrapped[:, 1], src_wrapped[:, 2], scale_factor=scale_factor, color=c_pink) mlab.points3d(src_pcd[ :, 0] , src_pcd[ :, 1], src_pcd[:, 2], scale_factor=scale_factor , color=c_red) mlab.points3d(tgt_pcd[ :, 0] , tgt_pcd[ :, 1], tgt_pcd[:, 2], scale_factor=scale_factor , color=c_blue) mlab.show() # add gaussian noise if self.data_augmentation: # rotate the point cloud euler_ab = np.random.rand(3) * np.pi * 2 / self.rot_factor # anglez, angley, anglex rot_ab = Rotation.from_euler('zyx', euler_ab).as_matrix() if (np.random.rand(1)[0] > 0.5): src_pcd = np.matmul(rot_ab, src_pcd.T).T src_pcd_deformed = np.matmul(rot_ab, src_pcd_deformed.T).T rot = np.matmul(rot, rot_ab.T) else: tgt_pcd = np.matmul(rot_ab, tgt_pcd.T).T rot = np.matmul(rot_ab, rot) trans = np.matmul(rot_ab, trans) src_pcd += (np.random.rand(src_pcd.shape[0], 3) - 0.5) * self.augment_noise tgt_pcd += (np.random.rand(tgt_pcd.shape[0], 3) - 0.5) * self.augment_noise s2t_flow = src_pcd_deformed - src_pcd if debug: # wrapp_src = (np.matmul(rot, src_pcd.T)+ trans).T src_wrapped = (np.matmul( rot, src_pcd_deformed.T ) + trans ).T mlab.points3d(src_wrapped[:, 0], src_wrapped[:, 1], src_wrapped[:, 2], scale_factor=scale_factor, color=c_red) mlab.points3d(tgt_pcd[:, 0], tgt_pcd[:, 1], tgt_pcd[:, 2], scale_factor=scale_factor, color=c_blue) mlab.show() if (trans.ndim == 1): trans = trans[:, None] src_feats = np.ones_like(src_pcd[:, :1]).astype(np.float32) tgt_feats = np.ones_like(tgt_pcd[:, :1]).astype(np.float32) rot = rot.astype(np.float32) trans = trans.astype(np.float32) #R * ( Ps + flow ) + t = Pt return src_pcd, tgt_pcd, src_feats, tgt_feats, correspondences, rot, trans, s2t_flow, metric_index if __name__ == '__main__': from lib.utils import load_config from easydict import EasyDict as edict from lib.tictok import Timers import yaml def join(loader, node): seq = loader.construct_sequence(node) return '_'.join([str(i) for i in seq]) yaml.add_constructor('!join', join) config = "/home/liyang/workspace/Regformer/configs/train/4dmatch.yaml" with open(config,'r') as f: config = yaml.load(f, Loader=yaml.Loader) config = edict(config) config.timers=Timers() D = _4DMatch(config, "test") for i in range (len(D)): try: if i%1000 == 0 : print (i,"/",len(D)) D.__getitem__(i, debug=True) except: # print(i, "/", len(D)) pass
5,939
32.75
123
py
lepard
lepard-main/datasets/dataloader.py
import numpy as np from functools import partial import torch import cpp_wrappers.cpp_subsampling.grid_subsampling as cpp_subsampling import cpp_wrappers.cpp_neighbors.radius_neighbors as cpp_neighbors from datasets._3dmatch import _3DMatch from datasets._4dmatch import _4DMatch from datasets.utils import blend_scene_flow, multual_nn_correspondence from lib.visualization import * from torch.utils.data import DataLoader def batch_grid_subsampling_kpconv(points, batches_len, features=None, labels=None, sampleDl=0.1, max_p=0, verbose=0, random_grid_orient=True): """ CPP wrapper for a grid subsampling (method = barycenter for points and features) """ if (features is None) and (labels is None): s_points, s_len = cpp_subsampling.subsample_batch(points, batches_len, sampleDl=sampleDl, max_p=max_p, verbose=verbose) return torch.from_numpy(s_points), torch.from_numpy(s_len) elif (labels is None): s_points, s_len, s_features = cpp_subsampling.subsample_batch(points, batches_len, features=features, sampleDl=sampleDl, max_p=max_p, verbose=verbose) return torch.from_numpy(s_points), torch.from_numpy(s_len), torch.from_numpy(s_features) elif (features is None): s_points, s_len, s_labels = cpp_subsampling.subsample_batch(points, batches_len, classes=labels, sampleDl=sampleDl, max_p=max_p, verbose=verbose) return torch.from_numpy(s_points), torch.from_numpy(s_len), torch.from_numpy(s_labels) else: s_points, s_len, s_features, s_labels = cpp_subsampling.subsample_batch(points, batches_len, features=features, classes=labels, sampleDl=sampleDl, max_p=max_p, verbose=verbose) return torch.from_numpy(s_points), torch.from_numpy(s_len), torch.from_numpy(s_features), torch.from_numpy(s_labels) def batch_neighbors_kpconv(queries, supports, q_batches, s_batches, radius, max_neighbors): """ Computes neighbors for a batch of queries and supports, apply radius search :param queries: (N1, 3) the query points :param supports: (N2, 3) the support points :param q_batches: (B) the list of lengths of batch elements in queries :param s_batches: (B)the list of lengths of batch elements in supports :param radius: float32 :return: neighbors indices """ neighbors = cpp_neighbors.batch_query(queries, supports, q_batches, s_batches, radius=radius) if max_neighbors > 0: return torch.from_numpy(neighbors[:, :max_neighbors]) else: return torch.from_numpy(neighbors) def collate_fn_3dmatch(list_data, config, neighborhood_limits ): batched_points_list = [] batched_features_list = [] batched_lengths_list = [] correspondences_list = [] src_pcd_list = [] tgt_pcd_list = [] batched_rot = [] batched_trn = [] gt_cov_list = [] for ind, ( src_pcd, tgt_pcd, src_feats, tgt_feats, correspondences, rot, trn, gt_cov) in enumerate(list_data): correspondences_list.append(correspondences ) src_pcd_list.append(torch.from_numpy(src_pcd) ) tgt_pcd_list.append(torch.from_numpy(tgt_pcd) ) batched_points_list.append(src_pcd) batched_points_list.append(tgt_pcd) batched_features_list.append(src_feats) batched_features_list.append(tgt_feats) batched_lengths_list.append(len(src_pcd)) batched_lengths_list.append(len(tgt_pcd)) batched_rot.append( torch.from_numpy(rot).float()) batched_trn.append( torch.from_numpy(trn).float()) gt_cov_list.append(gt_cov) gt_cov_list = None if gt_cov_list[0] is None \ else np.stack(gt_cov_list, axis=0) # if timers: cnter['collate_load_batch'] = time.time() - st batched_features = torch.from_numpy(np.concatenate(batched_features_list, axis=0)) batched_points = torch.from_numpy(np.concatenate(batched_points_list, axis=0)) batched_lengths = torch.from_numpy(np.array(batched_lengths_list)).int() batched_rot = torch.stack(batched_rot,dim=0) batched_trn = torch.stack(batched_trn,dim=0) # Starting radius of convolutions r_normal = config.first_subsampling_dl * config.conv_radius # Starting layer layer_blocks = [] layer = 0 # Lists of inputs input_points = [] input_neighbors = [] input_pools = [] input_upsamples = [] input_batches_len = [] # construt kpfcn inds for block_i, block in enumerate(config.architecture): # Stop when meeting a global pooling or upsampling if 'global' in block or 'upsample' in block: break # Get all blocks of the layer if not ('pool' in block or 'strided' in block): layer_blocks += [block] if block_i < len(config.architecture) - 1 and not ('upsample' in config.architecture[block_i + 1]): continue # Convolution neighbors indices # ***************************** if layer_blocks: # Convolutions are done in this layer, compute the neighbors with the good radius if np.any(['deformable' in blck for blck in layer_blocks[:-1]]): r = r_normal * config.deform_radius / config.conv_radius else: r = r_normal conv_i = batch_neighbors_kpconv(batched_points, batched_points, batched_lengths, batched_lengths, r, neighborhood_limits[layer]) else: # This layer only perform pooling, no neighbors required conv_i = torch.zeros((0, 1), dtype=torch.int64) # Pooling neighbors indices # ************************* # If end of layer is a pooling operation if 'pool' in block or 'strided' in block: # New subsampling length dl = 2 * r_normal / config.conv_radius # Subsampled points pool_p, pool_b = batch_grid_subsampling_kpconv(batched_points, batched_lengths, sampleDl=dl) # Radius of pooled neighbors if 'deformable' in block: r = r_normal * config.deform_radius / config.conv_radius else: r = r_normal # Subsample indices pool_i = batch_neighbors_kpconv(pool_p, batched_points, pool_b, batched_lengths, r, neighborhood_limits[layer]) # Upsample indices (with the radius of the next layer to keep wanted density) up_i = batch_neighbors_kpconv(batched_points, pool_p, batched_lengths, pool_b, 2 * r, neighborhood_limits[layer]) else: # No pooling in the end of this layer, no pooling indices required pool_i = torch.zeros((0, 1), dtype=torch.int64) pool_p = torch.zeros((0, 3), dtype=torch.float32) pool_b = torch.zeros((0,), dtype=torch.int64) up_i = torch.zeros((0, 1), dtype=torch.int64) # Updating input lists input_points += [batched_points.float()] input_neighbors += [conv_i.long()] input_pools += [pool_i.long()] input_upsamples += [up_i.long()] input_batches_len += [batched_lengths] # New points for next layer batched_points = pool_p batched_lengths = pool_b # Update radius and reset blocks r_normal *= 2 layer += 1 layer_blocks = [] # coarse infomation coarse_level = config.coarse_level pts_num_coarse = input_batches_len[coarse_level].view(-1, 2) b_size = pts_num_coarse.shape[0] src_pts_max, tgt_pts_max = pts_num_coarse.amax(dim=0) coarse_pcd = input_points[coarse_level] # .numpy() coarse_matches= [] src_ind_coarse_split= [] # src_feats shape :[b_size * src_pts_max] src_ind_coarse = [] tgt_ind_coarse_split= [] tgt_ind_coarse = [] accumu = 0 src_mask = torch.zeros([b_size, src_pts_max], dtype=torch.bool) tgt_mask = torch.zeros([b_size, tgt_pts_max], dtype=torch.bool) #grid subsample fine level points for differentiable matching fine_pts, fine_length = batch_grid_subsampling_kpconv(input_points[0], input_batches_len[0], sampleDl=dl*0.5*0.85) fine_ind = batch_neighbors_kpconv(fine_pts, input_points[0], fine_length, input_batches_len[0], dl*0.5*0.85, 1).squeeze().long() for entry_id, cnt in enumerate( pts_num_coarse ): #input_batches_len[-1].numpy().reshape(-1,2)) : n_s_pts, n_t_pts = cnt '''split mask for bottlenect feats''' src_mask[entry_id][:n_s_pts] = 1 tgt_mask[entry_id][:n_t_pts] = 1 '''split indices of bottleneck feats''' src_ind_coarse_split.append( torch.arange( n_s_pts ) + entry_id * src_pts_max ) tgt_ind_coarse_split.append( torch.arange( n_t_pts ) + entry_id * tgt_pts_max ) src_ind_coarse.append( torch.arange( n_s_pts ) + accumu ) tgt_ind_coarse.append( torch.arange( n_t_pts ) + accumu + n_s_pts ) '''get match at coarse level''' c_src_pcd = coarse_pcd[accumu : accumu + n_s_pts] c_tgt_pcd = coarse_pcd[accumu + n_s_pts: accumu + n_s_pts + n_t_pts] s_pc_wrapped = (torch.matmul( batched_rot[entry_id], c_src_pcd.T ) + batched_trn [entry_id]).T coarse_match_gt = torch.from_numpy( multual_nn_correspondence(s_pc_wrapped.numpy(), c_tgt_pcd.numpy(), search_radius=config['coarse_match_radius']) )# 0.1m scaled coarse_matches.append(coarse_match_gt) accumu = accumu + n_s_pts + n_t_pts vis=False # for debug if vis : viz_coarse_nn_correspondence_mayavi(c_src_pcd, c_tgt_pcd, coarse_match_gt, scale_factor=0.04) vis=False # for debug if vis : pass import mayavi.mlab as mlab # src_nei_valid = src_nei_mask[coarse_match_gt[0]].view(-1) # tgt_nei_valid = tgt_nei_mask[coarse_match_gt[1]].view(-1) # # f_src_pcd = src_m_nei_pts.view(-1, 3)[src_nei_valid] # f_tgt_pcd = tgt_m_nei_pts.view(-1,3)[tgt_nei_valid] # # mlab.points3d(f_src_pcd[:, 0], f_src_pcd[:, 1], f_src_pcd[:, 2], scale_factor=0.02,color=c_gray1) # mlab.points3d(f_tgt_pcd[:, 0], f_tgt_pcd[:, 1], f_tgt_pcd[:, 2], scale_factor=0.02,color=c_gray2) # # src_m_nn_pts =src_m_nn_pts.view(-1, 3) # src_m_nn_pts_wrapped = src_m_nn_pts_wrapped.view(-1,3) # tgt_m_nn_pts = tgt_m_nei_pts [ torch.arange(tgt_m_nei_pts.shape[0]), nni.view(-1), ... ] # mlab.points3d(src_m_nn_pts[:, 0], src_m_nn_pts[:, 1], src_m_nn_pts[:, 2], scale_factor=0.04,color=c_red) # mlab.points3d(src_m_nn_pts_wrapped[:, 0], src_m_nn_pts_wrapped[:, 1], src_m_nn_pts_wrapped[:, 2], scale_factor=0.04,color=c_red) # mlab.points3d(tgt_m_nn_pts[:, 0], tgt_m_nn_pts[:, 1], tgt_m_nn_pts[:, 2], scale_factor=0.04 ,color=c_blue) # mlab.show() # viz_coarse_nn_correspondence_mayavi(c_src_pcd, c_tgt_pcd, coarse_match_gt, # f_src_pcd=src_m_nei_pts.view(-1,3)[src_nei_valid], # f_tgt_pcd=tgt_m_nei_pts.view(-1,3)[tgt_nei_valid], scale_factor=0.08) src_ind_coarse_split = torch.cat(src_ind_coarse_split) tgt_ind_coarse_split = torch.cat(tgt_ind_coarse_split) src_ind_coarse = torch.cat(src_ind_coarse) tgt_ind_coarse = torch.cat(tgt_ind_coarse) dict_inputs = { 'src_pcd_list': src_pcd_list, 'tgt_pcd_list': tgt_pcd_list, 'points': input_points, 'neighbors': input_neighbors, 'pools': input_pools, 'upsamples': input_upsamples, 'features': batched_features.float(), 'stack_lengths': input_batches_len, 'coarse_matches': coarse_matches, 'src_mask': src_mask, 'tgt_mask': tgt_mask, 'src_ind_coarse_split': src_ind_coarse_split, 'tgt_ind_coarse_split': tgt_ind_coarse_split, 'src_ind_coarse': src_ind_coarse, 'tgt_ind_coarse': tgt_ind_coarse, 'batched_rot': batched_rot, 'batched_trn': batched_trn, 'gt_cov': gt_cov_list, #for refine 'correspondences_list': correspondences_list, 'fine_ind': fine_ind, 'fine_pts': fine_pts, 'fine_length': fine_length } return dict_inputs def collate_fn_4dmatch(list_data, config, neighborhood_limits ): batched_points_list = [] batched_features_list = [] batched_lengths_list = [] correspondences_list = [] src_pcd_list = [] tgt_pcd_list = [] batched_rot = [] batched_trn = [] sflow_list = [] metric_index_list = [] #for feature matching recall computation for ind, ( src_pcd, tgt_pcd, src_feats, tgt_feats, correspondences, rot, trn, s2t_flow, metric_index) in enumerate(list_data): correspondences_list.append(correspondences ) src_pcd_list.append(torch.from_numpy(src_pcd) ) tgt_pcd_list.append(torch.from_numpy(tgt_pcd) ) batched_points_list.append(src_pcd) batched_points_list.append(tgt_pcd) batched_features_list.append(src_feats) batched_features_list.append(tgt_feats) batched_lengths_list.append(len(src_pcd)) batched_lengths_list.append(len(tgt_pcd)) batched_rot.append( torch.from_numpy(rot).float()) batched_trn.append( torch.from_numpy(trn).float()) # gt_cov_list.append(gt_cov) sflow_list.append( torch.from_numpy(s2t_flow).float() ) if metric_index is None: metric_index_list = None else : metric_index_list.append ( torch.from_numpy(metric_index)) # if timers: cnter['collate_load_batch'] = time.time() - st batched_features = torch.from_numpy(np.concatenate(batched_features_list, axis=0)) batched_points = torch.from_numpy(np.concatenate(batched_points_list, axis=0)) batched_lengths = torch.from_numpy(np.array(batched_lengths_list)).int() batched_rot = torch.stack(batched_rot,dim=0) batched_trn = torch.stack(batched_trn,dim=0) # Starting radius of convolutions r_normal = config.first_subsampling_dl * config.conv_radius # Starting layer layer_blocks = [] layer = 0 # Lists of inputs input_points = [] input_neighbors = [] input_pools = [] input_upsamples = [] input_batches_len = [] # construt kpfcn inds for block_i, block in enumerate(config.architecture): # Stop when meeting a global pooling or upsampling if 'global' in block or 'upsample' in block: break # Get all blocks of the layer if not ('pool' in block or 'strided' in block): layer_blocks += [block] if block_i < len(config.architecture) - 1 and not ('upsample' in config.architecture[block_i + 1]): continue # Convolution neighbors indices # ***************************** if layer_blocks: # Convolutions are done in this layer, compute the neighbors with the good radius if np.any(['deformable' in blck for blck in layer_blocks[:-1]]): r = r_normal * config.deform_radius / config.conv_radius else: r = r_normal conv_i = batch_neighbors_kpconv(batched_points, batched_points, batched_lengths, batched_lengths, r, neighborhood_limits[layer]) else: # This layer only perform pooling, no neighbors required conv_i = torch.zeros((0, 1), dtype=torch.int64) # Pooling neighbors indices # ************************* # If end of layer is a pooling operation if 'pool' in block or 'strided' in block: # New subsampling length dl = 2 * r_normal / config.conv_radius # Subsampled points pool_p, pool_b = batch_grid_subsampling_kpconv(batched_points, batched_lengths, sampleDl=dl) # Radius of pooled neighbors if 'deformable' in block: r = r_normal * config.deform_radius / config.conv_radius else: r = r_normal # Subsample indices pool_i = batch_neighbors_kpconv(pool_p, batched_points, pool_b, batched_lengths, r, neighborhood_limits[layer]) # Upsample indices (with the radius of the next layer to keep wanted density) up_i = batch_neighbors_kpconv(batched_points, pool_p, batched_lengths, pool_b, 2 * r, neighborhood_limits[layer]) else: # No pooling in the end of this layer, no pooling indices required pool_i = torch.zeros((0, 1), dtype=torch.int64) pool_p = torch.zeros((0, 3), dtype=torch.float32) pool_b = torch.zeros((0,), dtype=torch.int64) up_i = torch.zeros((0, 1), dtype=torch.int64) # Updating input lists input_points += [batched_points.float()] input_neighbors += [conv_i.long()] input_pools += [pool_i.long()] input_upsamples += [up_i.long()] input_batches_len += [batched_lengths] # New points for next layer batched_points = pool_p batched_lengths = pool_b # Update radius and reset blocks r_normal *= 2 layer += 1 layer_blocks = [] # coarse infomation coarse_level = config.coarse_level pts_num_coarse = input_batches_len[coarse_level].view(-1, 2) b_size = pts_num_coarse.shape[0] src_pts_max, tgt_pts_max = pts_num_coarse.amax(dim=0) coarse_pcd = input_points[coarse_level] # .numpy() coarse_matches= [] coarse_flow = [] src_ind_coarse_split= [] # src_feats shape :[b_size * src_pts_max] src_ind_coarse = [] tgt_ind_coarse_split= [] tgt_ind_coarse = [] accumu = 0 src_mask = torch.zeros([b_size, src_pts_max], dtype=torch.bool) tgt_mask = torch.zeros([b_size, tgt_pts_max], dtype=torch.bool) for entry_id, cnt in enumerate( pts_num_coarse ): #input_batches_len[-1].numpy().reshape(-1,2)) : n_s_pts, n_t_pts = cnt '''split mask for bottlenect feats''' src_mask[entry_id][:n_s_pts] = 1 tgt_mask[entry_id][:n_t_pts] = 1 '''split indices of bottleneck feats''' src_ind_coarse_split.append( torch.arange( n_s_pts ) + entry_id * src_pts_max ) tgt_ind_coarse_split.append( torch.arange( n_t_pts ) + entry_id * tgt_pts_max ) src_ind_coarse.append( torch.arange( n_s_pts ) + accumu ) tgt_ind_coarse.append( torch.arange( n_t_pts ) + accumu + n_s_pts ) '''get match at coarse level''' c_src_pcd_np = coarse_pcd[accumu : accumu + n_s_pts].numpy() c_tgt_pcd_np = coarse_pcd[accumu + n_s_pts: accumu + n_s_pts + n_t_pts].numpy() #interpolate flow f_src_pcd = batched_points_list[entry_id * 2] c_flow = blend_scene_flow( c_src_pcd_np, f_src_pcd, sflow_list[entry_id].numpy(), knn=3) c_src_pcd_deformed = c_src_pcd_np + c_flow s_pc_wrapped = (np.matmul( batched_rot[entry_id].numpy(), c_src_pcd_deformed.T ) + batched_trn [entry_id].numpy()).T coarse_match_gt = torch.from_numpy( multual_nn_correspondence(s_pc_wrapped , c_tgt_pcd_np , search_radius=config['coarse_match_radius']) )# 0.1m scaled coarse_matches.append(coarse_match_gt) coarse_flow.append(torch.from_numpy(c_flow) ) accumu = accumu + n_s_pts + n_t_pts vis=False # for debug if vis : viz_coarse_nn_correspondence_mayavi(c_src_pcd_np, c_tgt_pcd_np, coarse_match_gt, scale_factor=0.02) src_ind_coarse_split = torch.cat(src_ind_coarse_split) tgt_ind_coarse_split = torch.cat(tgt_ind_coarse_split) src_ind_coarse = torch.cat(src_ind_coarse) tgt_ind_coarse = torch.cat(tgt_ind_coarse) dict_inputs = { 'src_pcd_list': src_pcd_list, 'tgt_pcd_list': tgt_pcd_list, 'points': input_points, 'neighbors': input_neighbors, 'pools': input_pools, 'upsamples': input_upsamples, 'features': batched_features.float(), 'stack_lengths': input_batches_len, 'coarse_matches': coarse_matches, 'coarse_flow' : coarse_flow, 'src_mask': src_mask, 'tgt_mask': tgt_mask, 'src_ind_coarse_split': src_ind_coarse_split, 'tgt_ind_coarse_split': tgt_ind_coarse_split, 'src_ind_coarse': src_ind_coarse, 'tgt_ind_coarse': tgt_ind_coarse, 'batched_rot': batched_rot, 'batched_trn': batched_trn, 'sflow_list': sflow_list, "metric_index_list": metric_index_list } return dict_inputs def calibrate_neighbors(dataset, config, collate_fn, keep_ratio=0.8, samples_threshold=2000): # From config parameter, compute higher bound of neighbors number in a neighborhood hist_n = int(np.ceil(4 / 3 * np.pi * (config.deform_radius + 1) ** 3)) neighb_hists = np.zeros((config.num_layers, hist_n), dtype=np.int32) # Get histogram of neighborhood sizes i in 1 epoch max. for i in range(len(dataset)): batched_input = collate_fn([dataset[i]], config, neighborhood_limits=[hist_n] * 5) # update histogram counts = [torch.sum(neighb_mat < neighb_mat.shape[0], dim=1).numpy() for neighb_mat in batched_input['neighbors']] hists = [np.bincount(c, minlength=hist_n)[:hist_n] for c in counts] neighb_hists += np.vstack(hists) # if timer.total_time - last_display > 0.1: # last_display = timer.total_time # print(f"Calib Neighbors {i:08d}: timings {timer.total_time:4.2f}s") if np.min(np.sum(neighb_hists, axis=1)) > samples_threshold: break cumsum = np.cumsum(neighb_hists.T, axis=0) percentiles = np.sum(cumsum < (keep_ratio * cumsum[hist_n - 1, :]), axis=0) neighborhood_limits = percentiles print('\n') return neighborhood_limits def get_datasets(config): if (config.dataset == '3dmatch'): train_set = _3DMatch(config, 'train', data_augmentation=True) val_set = _3DMatch(config, 'val', data_augmentation=False) test_set = _3DMatch(config, 'test', data_augmentation=False) elif(config.dataset == '4dmatch'): train_set = _4DMatch(config, 'train', data_augmentation=True) val_set = _4DMatch(config, 'val', data_augmentation=False) test_set = _4DMatch(config, 'test', data_augmentation=False) else: raise NotImplementedError return train_set, val_set, test_set def get_dataloader(dataset, config, shuffle=True, neighborhood_limits=None): if config.dataset=='4dmatch': collate_fn = collate_fn_4dmatch elif config.dataset == '3dmatch': collate_fn = collate_fn_3dmatch else: raise NotImplementedError() if neighborhood_limits is None: neighborhood_limits = calibrate_neighbors(dataset, config['kpfcn_config'], collate_fn=collate_fn) print("neighborhood:", neighborhood_limits) dataloader = torch.utils.data.DataLoader( dataset, batch_size=config['batch_size'], shuffle=shuffle, num_workers=config['num_workers'], collate_fn=partial(collate_fn, config=config['kpfcn_config'], neighborhood_limits=neighborhood_limits ), drop_last=False ) return dataloader, neighborhood_limits if __name__ == '__main__': pass
24,996
37.875583
171
py
lepard
lepard-main/datasets/_3dmatch.py
import os, sys, glob, torch # sys.path.append("../") [sys.path.append(i) for i in ['.', '..']] import numpy as np import torch import random from scipy.spatial.transform import Rotation from torch.utils.data import Dataset from lib.benchmark_utils import to_o3d_pcd, to_tsfm, KDTree_corr from lib.utils import load_obj from lib.benchmark_utils import to_o3d_pcd, to_tsfm, get_correspondences class _3DMatch(Dataset): def __init__(self, config,split, data_augmentation=True): super(_3DMatch, self).__init__() assert split in ['train','val','test'] if 'overfit' in config.exp_dir: d_slice = config.batch_size else : d_slice = None self.infos = self.read_entries( config.split[split] , config.data_root, d_slice=d_slice ) self.base_dir = config.data_root self.data_augmentation = data_augmentation self.config = config self.rot_factor = 1. self.augment_noise = config.augment_noise self.max_points = 30000 self.overlap_radius = 0.0375 def read_entries (self, split, data_root, d_slice=None, shuffle= True): infos = load_obj(split) # we use the split prepared by Predator if d_slice: for k, v in infos.items(): infos[k] = v[:d_slice] return infos def __len__(self): return len(self.infos['rot']) def __getitem__(self, item, debug=False): # get transformation rot = self.infos['rot'][item] trans = self.infos['trans'][item] if 'gt_cov' in self.infos: gt_cov = self.infos['gt_cov'][item] else : gt_cov = None # get pointcloud src_path = os.path.join(self.base_dir, self.infos['src'][item]) tgt_path = os.path.join(self.base_dir, self.infos['tgt'][item]) src_pcd = torch.load(src_path) tgt_pcd = torch.load(tgt_path) # if we get too many points, we do some downsampling if (src_pcd.shape[0] > self.max_points): idx = np.random.permutation(src_pcd.shape[0])[:self.max_points] src_pcd = src_pcd[idx] if (tgt_pcd.shape[0] > self.max_points): idx = np.random.permutation(tgt_pcd.shape[0])[:self.max_points] tgt_pcd = tgt_pcd[idx] if debug: import mayavi.mlab as mlab c_red = (224. / 255., 0 / 255., 125 / 255.) c_pink = (224. / 255., 75. / 255., 232. / 255.) c_blue = (0. / 255., 0. / 255., 255. / 255.) scale_factor = 0.02 # mlab.points3d(s_pc[ :, 0] , s_pc[ :, 1], s_pc[:, 2], scale_factor=scale_factor , color=c_blue) mlab.points3d(src_pcd[ :, 0] , src_pcd[ :, 1], src_pcd[:, 2], scale_factor=scale_factor , color=c_red) mlab.points3d(tgt_pcd[ :, 0] , tgt_pcd[ :, 1], tgt_pcd[:, 2], scale_factor=scale_factor , color=c_blue) mlab.show() # add gaussian noise if self.data_augmentation: # rotate the point cloud euler_ab = np.random.rand(3) * np.pi * 2 / self.rot_factor # anglez, angley, anglex rot_ab = Rotation.from_euler('zyx', euler_ab).as_matrix() if (np.random.rand(1)[0] > 0.5): src_pcd = np.matmul(rot_ab, src_pcd.T).T rot = np.matmul(rot, rot_ab.T) else: tgt_pcd = np.matmul(rot_ab, tgt_pcd.T).T rot = np.matmul(rot_ab, rot) trans = np.matmul(rot_ab, trans) src_pcd += (np.random.rand(src_pcd.shape[0], 3) - 0.5) * self.augment_noise tgt_pcd += (np.random.rand(tgt_pcd.shape[0], 3) - 0.5) * self.augment_noise # get correspondence at fine level tsfm = to_tsfm(rot, trans) correspondences = get_correspondences(to_o3d_pcd(src_pcd), to_o3d_pcd(tgt_pcd), tsfm,self.overlap_radius) if debug: import mayavi.mlab as mlab c_red = (224. / 255., 0 / 255., 125 / 255.) c_pink = (224. / 255., 75. / 255., 232. / 255.) c_blue = (0. / 255., 0. / 255., 255. / 255.) scale_factor = 0.02 # mlab.points3d(s_pc[ :, 0] , s_pc[ :, 1], s_pc[:, 2], scale_factor=scale_factor , color=c_blue) mlab.points3d(src_pcd[ :, 0] , src_pcd[ :, 1], src_pcd[:, 2], scale_factor=scale_factor , color=c_red) mlab.points3d(tgt_pcd[ :, 0] , tgt_pcd[ :, 1], tgt_pcd[:, 2], scale_factor=scale_factor , color=c_blue) mlab.show() if (trans.ndim == 1): trans = trans[:, None] src_feats = np.ones_like(src_pcd[:, :1]).astype(np.float32) tgt_feats = np.ones_like(tgt_pcd[:, :1]).astype(np.float32) rot = rot.astype(np.float32) trans = trans.astype(np.float32) return src_pcd, tgt_pcd, src_feats, tgt_feats, correspondences, rot, trans, gt_cov if __name__ == '__main__': from lib.utils import load_config from easydict import EasyDict as edict from lib.tictok import Timers import yaml def join(loader, node): seq = loader.construct_sequence(node) return '_'.join([str(i) for i in seq]) yaml.add_constructor('!join', join) config = "/home/liyang/workspace/Regformer/configs/train/3dmatch.yaml" with open(config,'r') as f: config = yaml.load(f, Loader=yaml.Loader) config = edict(config) config.timers=Timers() D = _3DMatch(config, "test") for i in range (len(D)): try: if i%1000 == 0 : print (i,"/",len(D)) D.__getitem__(i, debug=True) except: pass # print ( D.data_entries[i] ) # print (os.remove(D.data_entries[i]) )
5,766
33.327381
116
py
lepard
lepard-main/lib/tester.py
from lib.trainer import Trainer import torch from tqdm import tqdm from models.loss import MatchMotionLoss as MML import numpy as np from models.matching import Matching as CM import math class _3DMatchTester(Trainer): """ 3DMatch tester """ def __init__(self,args): Trainer.__init__(self, args) def test(self): n = 3 afmr = 0. arr = 0 air = 0 for i in range(n): # combat ransac nondeterministic thr =0.05 rr, ir, fmr = self.test_thr(thr) afmr+=fmr arr+=rr air+=ir print( "conf_threshold", thr, "registration recall:", rr, " Inlier rate:", ir, "FMR:", fmr) print("average registration recall:", arr / n, afmr/n, air/n) # print ("registration recall:", self.test_thr()) def test_thr(self, conf_threshold=None): # print('Start to evaluate on test datasets...') # os.makedirs(f'{self.snapshot_dir}/{self.config.dataset}',exist_ok=True) num_iter = math.ceil(len(self.loader['test'].dataset) // self.loader['test'].batch_size) c_loader_iter = self.loader['test'].__iter__() self.model.eval() success1 = 0. IR=0. FMR=0. with torch.no_grad(): for idx in tqdm(range(num_iter)): # loop through this epoch ################################## if self.timers: self.timers.tic('load batch') inputs = c_loader_iter.next() for k, v in inputs.items(): if type(v) == list: inputs[k] = [item.to(self.device) for item in v] elif type(v) in [dict, float, type(None), np.ndarray]: pass else: inputs[k] = v.to(self.device) if self.timers: self.timers.toc('load batch') ################################## if self.timers: self.timers.tic('forward pass') data = self.model(inputs, timers=self.timers) # [N1, C1], [N2, C2] if self.timers: self.timers.toc('forward pass') match_pred, _, _ = CM.get_match(data['conf_matrix_pred'], thr=conf_threshold, mutual=False) rot, trn = MML.ransac_regist_coarse(data['s_pcd'], data['t_pcd'], data['src_mask'], data['tgt_mask'], match_pred) ir = MML.compute_inlier_ratio(match_pred, data, inlier_thr=0.1).mean() rr1 = MML.compute_registration_recall(rot, trn, data, thr=0.2) # 0.2m vis = False if vis: pcd = data['points'][0].cpu().numpy() lenth = data['stack_lengths'][0][0] spcd, tpcd = pcd[:lenth] , pcd[lenth:] import mayavi.mlab as mlab c_red = (224. / 255., 0 / 255., 125 / 255.) c_pink = (224. / 255., 75. / 255., 232. / 255.) c_blue = (0. / 255., 0. / 255., 255. / 255.) scale_factor = 0.02 # mlab.points3d(s_pc[ :, 0] , s_pc[ :, 1], s_pc[:, 2], scale_factor=scale_factor , color=c_blue) mlab.points3d(spcd[:, 0], spcd[:, 1], spcd[:, 2], scale_factor=scale_factor, color=c_red) mlab.points3d(tpcd[:, 0], tpcd[:, 1], tpcd[:, 2], scale_factor=scale_factor, color=c_blue) mlab.show() spcd = ( np.matmul(rot, spcd.T) + trn ).T mlab.points3d(spcd[:, 0], spcd[:, 1], spcd[:, 2], scale_factor=scale_factor, color=c_red) mlab.points3d(tpcd[:, 0], tpcd[:, 1], tpcd[:, 2], scale_factor=scale_factor, color=c_blue) mlab.show() bs = len(rot) assert bs==1 success1 += bs * rr1 IR += bs*ir FMR += (ir>0.05).float() recall1 = success1/len(self.loader['test'].dataset) IRate = IR/len(self.loader['test'].dataset) FMR = FMR/len(self.loader['test'].dataset) return recall1, IRate, FMR def blend_anchor_motion (query_loc, reference_loc, reference_flow , knn=3, search_radius=0.1) : '''approximate flow on query points this function assume query points are sub- or un-sampled from reference locations @param query_loc:[m,3] @param reference_loc:[n,3] @param reference_flow:[n,3] @param knn: @return: blended_flow:[m,3] ''' from datasets.utils import knn_point_np dists, idx = knn_point_np (knn, reference_loc, query_loc) dists[dists < 1e-10] = 1e-10 mask = dists>search_radius dists[mask] = 1e+10 weight = 1.0 / dists weight = weight / np.sum(weight, -1, keepdims=True) # [B,N,3] blended_flow = np.sum (reference_flow [idx] * weight.reshape ([-1, knn, 1]), axis=1, keepdims=False) mask = mask.sum(axis=1)<3 return blended_flow, mask def compute_nrfmr( match_pred, data, recall_thr=0.04): s_pcd, t_pcd = data['s_pcd'], data['t_pcd'] s_pcd_raw = data ['src_pcd_list'] sflow_list = data['sflow_list'] metric_index_list = data['metric_index_list'] batched_rot = data['batched_rot'] # B,3,3 batched_trn = data['batched_trn'] nrfmr = 0. for i in range ( len(s_pcd_raw)): # get the metric points' transformed position metric_index = metric_index_list[i] sflow = sflow_list[i] s_pcd_raw_i = s_pcd_raw[i] metric_pcd = s_pcd_raw_i [ metric_index ] metric_sflow = sflow [ metric_index ] metric_pcd_deformed = metric_pcd + metric_sflow metric_pcd_wrapped_gt = ( torch.matmul( batched_rot[i], metric_pcd_deformed.T) + batched_trn[i] ).T # use the match prediction as the motion anchor match_pred_i = match_pred[ match_pred[:, 0] == i ] s_id , t_id = match_pred_i[:,1], match_pred_i[:,2] s_pcd_matched= s_pcd[i][s_id] t_pcd_matched= t_pcd[i][t_id] motion_pred = t_pcd_matched - s_pcd_matched metric_motion_pred, valid_mask = blend_anchor_motion( metric_pcd.cpu().numpy(), s_pcd_matched.cpu().numpy(), motion_pred.cpu().numpy(), knn=3, search_radius=0.1) metric_pcd_wrapped_pred = metric_pcd + torch.from_numpy(metric_motion_pred).to(metric_pcd) debug = False if debug: import mayavi.mlab as mlab c_red = (224. / 255., 0 / 255., 125 / 255.) c_pink = (224. / 255., 75. / 255., 232. / 255.) c_blue = (0. / 255., 0. / 255., 255. / 255.) scale_factor = 0.013 metric_pcd_wrapped_gt = metric_pcd_wrapped_gt.cpu() metric_pcd_wrapped_pred = metric_pcd_wrapped_pred.cpu() err = metric_pcd_wrapped_pred - metric_pcd_wrapped_gt mlab.points3d(metric_pcd_wrapped_gt[:, 0], metric_pcd_wrapped_gt[:, 1], metric_pcd_wrapped_gt[:, 2], scale_factor=scale_factor, color=c_pink) mlab.points3d(metric_pcd_wrapped_pred[ :, 0] , metric_pcd_wrapped_pred[ :, 1], metric_pcd_wrapped_pred[:, 2], scale_factor=scale_factor , color=c_blue) mlab.quiver3d(metric_pcd_wrapped_gt[:, 0], metric_pcd_wrapped_gt[:, 1], metric_pcd_wrapped_gt[:, 2], err[:, 0], err[:, 1], err[:, 2], scale_factor=1, mode='2ddash', line_width=1.) mlab.show() dist = torch.sqrt( torch.sum( (metric_pcd_wrapped_pred - metric_pcd_wrapped_gt)**2, dim=1 ) ) r = (dist < recall_thr).float().sum() / len(dist) nrfmr = nrfmr + r nrfmr = nrfmr /len(s_pcd_raw) return nrfmr class _4DMatchTester(Trainer): """ 3DMatch tester """ def __init__(self,args): Trainer.__init__(self, args) def test(self): for thr in [ 0.05, 0.1, 0.2]: # for thr in [ 0.1 ]: import time start = time.time() ir, fmr, nspl = self.test_thr(thr) print( "conf_threshold", thr, "NFMR:", fmr, " Inlier rate:", ir, "Number sample:", nspl) print( "time costs:", time.time() - start) def test_thr(self, conf_threshold=None): num_iter = math.ceil(len(self.loader['test'].dataset) // self.loader['test'].batch_size) c_loader_iter = self.loader['test'].__iter__() self.model.eval() assert self.loader['test'].batch_size == 1 IR=0. NR_FMR=0. inlier_thr = recall_thr = 0.04 n_sample = 0. with torch.no_grad(): for idx in tqdm(range(num_iter)): # loop through this epoch ################################## if self.timers: self.timers.tic('load batch') inputs = c_loader_iter.next() for k, v in inputs.items(): if type(v) == list: inputs[k] = [item.to(self.device) for item in v] elif type(v) in [ dict, float, type(None), np.ndarray]: pass else: inputs[k] = v.to(self.device) if self.timers: self.timers.toc('load batch') ################################## if self.timers: self.timers.tic('forward pass') data = self.model(inputs, timers=self.timers) # [N1, C1], [N2, C2] if self.timers: self.timers.toc('forward pass') match_pred, _, _ = CM.get_match(data['conf_matrix_pred'], thr=conf_threshold, mutual=True) ir = MML.compute_inlier_ratio(match_pred, data, inlier_thr=inlier_thr, s2t_flow=data['coarse_flow'][0][None] )[0] nrfmr = compute_nrfmr(match_pred, data, recall_thr=recall_thr) IR += ir NR_FMR += nrfmr n_sample += match_pred.shape[0] IRate = IR/len(self.loader['test'].dataset) NR_FMR = NR_FMR/len(self.loader['test'].dataset) n_sample = n_sample/len(self.loader['test'].dataset) if self.timers: self.timers.print() return IRate, NR_FMR, n_sample def get_trainer(config): if config.dataset == '3dmatch': return _3DMatchTester(config) elif config.dataset == '4dmatch': return _4DMatchTester(config) else: raise NotImplementedError
10,544
34.385906
164
py
lepard
lepard-main/lib/benchmark_utils.py
import os,re,sys,json,yaml,random, glob, argparse, torch, pickle from tqdm import tqdm import numpy as np from scipy.spatial.transform import Rotation import open3d as o3d _EPS = 1e-7 # To prevent division by zero def viz_coarse_nn_correspondence_mayavi(s_pc, t_pc, good_c, bad_c, f_src_pcd=None, f_tgt_pcd=None, scale_factor=0.02): ''' @param s_pc: [S,3] @param t_pc: [T,3] @param correspondence: [2,K] @param f_src_pcd: [S1,3] @param f_tgt_pcd: [T1,3] @param scale_factor: @return: ''' import mayavi.mlab as mlab c_red = (224. / 255., 0 / 255., 0 / 255.) c_pink = (224. / 255., 75. / 255., 232. / 255.) c_blue = (0. / 255., 0. / 255., 255. / 255.) c_green = (0. / 255., 255. / 255., 0. / 255.) c_gray1 = (255 / 255., 255 / 255., 125 / 255.) c_gray2 = (125. / 255., 125. / 255., 255. / 255.) if f_src_pcd is not None: mlab.points3d(f_src_pcd[:, 0], f_src_pcd[:, 1], f_src_pcd[:, 2], scale_factor=scale_factor * 0.25, color=c_gray1) else: mlab.points3d(s_pc[:, 0], s_pc[:, 1], s_pc[:, 2], scale_factor=scale_factor * 0.75, color=c_gray1) if f_tgt_pcd is not None: mlab.points3d(f_tgt_pcd[:, 0], f_tgt_pcd[:, 1], f_tgt_pcd[:, 2], scale_factor=scale_factor * 0.25, color=c_gray2) else: mlab.points3d(t_pc[:, 0], t_pc[:, 1], t_pc[:, 2], scale_factor=scale_factor * 0.75, color=c_gray2) s_cpts_god = s_pc[good_c[0]] t_cpts_god = t_pc[good_c[1]] flow_good = t_cpts_god - s_cpts_god s_cpts_bd = s_pc[bad_c[0]] t_cpts_bd = t_pc[bad_c[1]] flow_bad = t_cpts_bd - s_cpts_bd def match_draw(s_cpts, t_cpts, flow, color): mlab.points3d(s_cpts[:, 0], s_cpts[:, 1], s_cpts[:, 2], scale_factor=scale_factor * 0.35, color=c_blue) mlab.points3d(t_cpts[:, 0], t_cpts[:, 1], t_cpts[:, 2], scale_factor=scale_factor * 0.35, color=c_pink) mlab.quiver3d(s_cpts[:, 0], s_cpts[:, 1], s_cpts[:, 2], flow[:, 0], flow[:, 1], flow[:, 2], scale_factor=1, mode='2ddash', line_width=1., color=color) match_draw(s_cpts_god, t_cpts_god, flow_good, c_green) match_draw(s_cpts_bd, t_cpts_bd, flow_bad, c_red) mlab.show() def correspondence_viz(src_raw, tgt_raw, src_pcd, tgt_pcd, corrs, inlier_mask, max=200): perm = np.random.permutation(corrs.shape[1]) ind = perm[:max] corrs = corrs[:, ind] inlier_mask = inlier_mask[ind] good_c = corrs[:, inlier_mask] bad_c = corrs[:, ~inlier_mask] offset = np.array([[1.45, 0, 0]]) # src_pcd = src_pcd + offset # src_raw = src_raw + offset tgt_pcd = tgt_pcd + offset tgt_raw = tgt_raw + offset viz_coarse_nn_correspondence_mayavi(src_pcd, tgt_pcd, good_c, bad_c, src_raw, tgt_raw, scale_factor=0.07) def fmr_wrt_distance(data,split,inlier_ratio_threshold=0.05): """ calculate feature match recall wrt distance threshold """ fmr_wrt_distance =[] for distance_threshold in range(1,21): inlier_ratios =[] distance_threshold /=100.0 for idx in range(data.shape[0]): inlier_ratio = (data[idx] < distance_threshold).mean() inlier_ratios.append(inlier_ratio) fmr = 0 for ele in split: fmr += (np.array(inlier_ratios[ele[0]:ele[1]]) > inlier_ratio_threshold).mean() fmr /= 8 fmr_wrt_distance.append(fmr*100) return fmr_wrt_distance def fmr_wrt_inlier_ratio(data, split, distance_threshold=0.1): """ calculate feature match recall wrt inlier ratio threshold """ fmr_wrt_inlier =[] for inlier_ratio_threshold in range(1,21): inlier_ratios =[] inlier_ratio_threshold /=100.0 for idx in range(data.shape[0]): inlier_ratio = (data[idx] < distance_threshold).mean() inlier_ratios.append(inlier_ratio) fmr = 0 for ele in split: fmr += (np.array(inlier_ratios[ele[0]:ele[1]]) > inlier_ratio_threshold).mean() fmr /= 8 fmr_wrt_inlier.append(fmr*100) return fmr_wrt_inlier def to_tensor(array): """ Convert array to tensor """ if(not isinstance(array,torch.Tensor)): return torch.from_numpy(array).float() else: return array def to_array(tensor): """ Conver tensor to array """ if(not isinstance(tensor,np.ndarray)): if(tensor.device == torch.device('cpu')): return tensor.numpy() else: return tensor.cpu().numpy() else: return tensor def to_tsfm(rot,trans): tsfm = np.eye(4) tsfm[:3,:3]=rot tsfm[:3,3]=trans.flatten() return tsfm def to_o3d_pcd(xyz): """ Convert tensor/array to open3d PointCloud xyz: [N, 3] """ pcd = o3d.geometry.PointCloud() pcd.points = o3d.utility.Vector3dVector(to_array(xyz)) return pcd def to_o3d_feats(embedding): """ Convert tensor/array to open3d features embedding: [N, 3] """ feats = o3d.registration.Feature() feats.data = to_array(embedding).T return feats def get_correspondences(src_pcd, tgt_pcd, trans, search_voxel_size, K=None): src_pcd.transform(trans) correspondences = KDTree_corr ( src_pcd, tgt_pcd, search_voxel_size, K=None) correspondences = torch.from_numpy(correspondences) return correspondences def KDTree_corr ( src_pcd_transformed, tgt_pcd, search_voxel_size, K=None): pcd_tree = o3d.geometry.KDTreeFlann(tgt_pcd) correspondences = [] for i, point in enumerate(src_pcd_transformed.points): [count, idx, _] = pcd_tree.search_radius_vector_3d(point, search_voxel_size) if K is not None: idx = idx[:K] for j in idx: correspondences.append([i, j]) correspondences = np.array(correspondences) return correspondences def get_blue(): """ Get color blue for rendering """ return [0, 0.651, 0.929] def get_yellow(): """ Get color yellow for rendering """ return [1, 0.706, 0] def random_sample(pcd, feats, N): """ Do random sampling to get exact N points and associated features pcd: [N,3] feats: [N,C] """ if(isinstance(pcd,torch.Tensor)): n1 = pcd.size(0) elif(isinstance(pcd, np.ndarray)): n1 = pcd.shape[0] if n1 == N: return pcd, feats if n1 > N: choice = np.random.permutation(n1)[:N] else: choice = np.random.choice(n1, N) return pcd[choice], feats[choice] def get_angle_deviation(R_pred,R_gt): """ Calculate the angle deviation between two rotaion matrice The rotation error is between [0,180] Input: R_pred: [B,3,3] R_gt : [B,3,3] Return: degs: [B] """ R=np.matmul(R_pred,R_gt.transpose(0,2,1)) tr=np.trace(R,0,1,2) rads=np.arccos(np.clip((tr-1)/2,-1,1)) # clip to valid range degs=rads/np.pi*180 return degs def ransac_pose_estimation(src_pcd, tgt_pcd, src_feat, tgt_feat, mutual = False, distance_threshold = 0.05, ransac_n = 3): """ RANSAC pose estimation with two checkers We follow D3Feat to set ransac_n = 3 for 3DMatch and ransac_n = 4 for KITTI. For 3DMatch dataset, we observe significant improvement after changing ransac_n from 4 to 3. """ if(mutual): if(torch.cuda.device_count()>=1): device = torch.device('cuda') else: device = torch.device('cpu') src_feat, tgt_feat = to_tensor(src_feat), to_tensor(tgt_feat) scores = torch.matmul(src_feat.to(device), tgt_feat.transpose(0,1).to(device)).cpu() selection = mutual_selection(scores[None,:,:])[0] row_sel, col_sel = np.where(selection) corrs = o3d.utility.Vector2iVector(np.array([row_sel,col_sel]).T) src_pcd = to_o3d_pcd(src_pcd) tgt_pcd = to_o3d_pcd(tgt_pcd) result_ransac = o3d.registration.registration_ransac_based_on_correspondence( source=src_pcd, target=tgt_pcd,corres=corrs, max_correspondence_distance=distance_threshold, estimation_method=o3d.registration.TransformationEstimationPointToPoint(False), ransac_n=4, criteria=o3d.registration.RANSACConvergenceCriteria(50000, 1000)) else: src_pcd = to_o3d_pcd(src_pcd) tgt_pcd = to_o3d_pcd(tgt_pcd) src_feats = to_o3d_feats(src_feat) tgt_feats = to_o3d_feats(tgt_feat) result_ransac = o3d.registration.registration_ransac_based_on_feature_matching( src_pcd, tgt_pcd, src_feats, tgt_feats,distance_threshold, o3d.registration.TransformationEstimationPointToPoint(False), ransac_n, [o3d.registration.CorrespondenceCheckerBasedOnEdgeLength(0.9), o3d.registration.CorrespondenceCheckerBasedOnDistance(distance_threshold)], o3d.registration.RANSACConvergenceCriteria(50000, 1000)) return result_ransac.transformation def get_inlier_ratio(src_pcd, tgt_pcd, src_feat, tgt_feat, rot, trans, inlier_distance_threshold = 0.1): """ Compute inlier ratios with and without mutual check, return both """ src_pcd = to_tensor(src_pcd) tgt_pcd = to_tensor(tgt_pcd) src_feat = to_tensor(src_feat) tgt_feat = to_tensor(tgt_feat) rot, trans = to_tensor(rot), to_tensor(trans) results =dict() results['w']=dict() results['wo']=dict() if(torch.cuda.device_count()>=1): device = torch.device('cuda') else: device = torch.device('cpu') src_pcd = (torch.matmul(rot, src_pcd.transpose(0,1)) + trans).transpose(0,1) scores = torch.matmul(src_feat.to(device), tgt_feat.transpose(0,1).to(device)).cpu() ######################################## # 1. calculate inlier ratios wo mutual check _, idx = scores.max(-1) dist = torch.norm(src_pcd- tgt_pcd[idx],dim=1) results['wo']['distance'] = dist.numpy() c_inlier_ratio = (dist < inlier_distance_threshold).float().mean() results['wo']['inlier_ratio'] = c_inlier_ratio ######################################## # 2. calculate inlier ratios w mutual check selection = mutual_selection(scores[None,:,:])[0] row_sel, col_sel = np.where(selection) dist = torch.norm(src_pcd[row_sel]- tgt_pcd[col_sel],dim=1) results['w']['distance'] = dist.numpy() c_inlier_ratio = (dist < inlier_distance_threshold).float().mean() results['w']['inlier_ratio'] = c_inlier_ratio return results def mutual_selection(score_mat): """ Return a {0,1} matrix, the element is 1 if and only if it's maximum along both row and column Args: np.array() score_mat: [B,N,N] Return: mutuals: [B,N,N] """ score_mat=to_array(score_mat) if(score_mat.ndim==2): score_mat=score_mat[None,:,:] mutuals=np.zeros_like(score_mat) for i in range(score_mat.shape[0]): # loop through the batch c_mat=score_mat[i] flag_row=np.zeros_like(c_mat) flag_column=np.zeros_like(c_mat) max_along_row=np.argmax(c_mat,1)[:,None] max_along_column=np.argmax(c_mat,0)[None,:] np.put_along_axis(flag_row,max_along_row,1,1) np.put_along_axis(flag_column,max_along_column,1,0) mutuals[i]=(flag_row.astype(np.bool)) & (flag_column.astype(np.bool)) return mutuals.astype(np.bool)
11,442
31.882184
122
py
lepard
lepard-main/lib/utils.py
import os,re,sys,json,yaml,random, argparse, torch, pickle import torch.nn as nn import torch.nn.functional as F import torch.optim as optim import numpy as np from scipy.spatial.transform import Rotation from sklearn.neighbors import NearestNeighbors from scipy.spatial.distance import minkowski _EPS = 1e-7 # To prevent division by zero class Logger: def __init__(self, path): self.path = path log_path = self.path + '/log' if os.path.exists(log_path): os.remove(log_path) self.fw = open(log_path,'a') def write(self, text): self.fw.write(text) self.fw.flush() def close(self): self.fw.close() def save_obj(obj, path ): """ save a dictionary to a pickle file """ with open(path, 'wb') as f: pickle.dump(obj, f) def load_obj(path): """ read a dictionary from a pickle file """ with open(path, 'rb') as f: return pickle.load(f) def load_config(path): """ Loads config file: Args: path (str): path to the config file Returns: config (dict): dictionary of the configuration parameters, merge sub_dicts """ with open(path,'r') as f: cfg = yaml.safe_load(f) config = dict() for key, value in cfg.items(): for k,v in value.items(): config[k] = v return config def setup_seed(seed): """ fix random seed for deterministic training """ torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) np.random.seed(seed) random.seed(seed) torch.backends.cudnn.deterministic = True def square_distance(src, dst, normalised = False): """ Calculate Euclid distance between each two points. Args: src: source points, [B, N, C] dst: target points, [B, M, C] Returns: dist: per-point square distance, [B, N, M] """ B, N, _ = src.shape _, M, _ = dst.shape dist = -2 * torch.matmul(src, dst.permute(0, 2, 1)) if(normalised): dist += 2 else: dist += torch.sum(src ** 2, dim=-1)[:, :, None] dist += torch.sum(dst ** 2, dim=-1)[:, None, :] dist = torch.clamp(dist, min=1e-12, max=None) return dist def validate_gradient(model): """ Confirm all the gradients are non-nan and non-inf """ for name, param in model.named_parameters(): if param.grad is not None: if torch.any(torch.isnan(param.grad)): return False if torch.any(torch.isinf(param.grad)): return False return True def natural_key(string_): """ Sort strings by numbers in the name """ return [int(s) if s.isdigit() else s for s in re.split(r'(\d+)', string_)]
2,759
23.424779
82
py
lepard
lepard-main/lib/trainer.py
import gc import os import torch import torch.nn as nn import numpy as np from tensorboardX import SummaryWriter from tqdm import tqdm from lib.timer import AverageMeter from lib.utils import Logger, validate_gradient from lib.tictok import Timers class Trainer(object): def __init__(self, args): self.config = args # parameters self.start_epoch = 1 self.max_epoch = args.max_epoch self.save_dir = args.save_dir self.device = args.device self.verbose = args.verbose self.model = args.model self.model = self.model.to(self.device) self.optimizer = args.optimizer self.scheduler = args.scheduler self.scheduler_freq = args.scheduler_freq self.snapshot_dir = args.snapshot_dir self.iter_size = args.iter_size self.verbose_freq = args.verbose_freq // args.batch_size + 1 if 'overfit' in self.config.exp_dir: self.verbose_freq = 1 self.loss = args.desc_loss self.best_loss = 1e5 self.best_recall = -1e5 self.summary_writer = SummaryWriter(log_dir=args.tboard_dir) self.logger = Logger(args.snapshot_dir) self.logger.write(f'#parameters {sum([x.nelement() for x in self.model.parameters()]) / 1000000.} M\n') if (args.pretrain != ''): self._load_pretrain(args.pretrain) self.loader = dict() self.loader['train'] = args.train_loader self.loader['val'] = args.val_loader self.loader['test'] = args.test_loader self.timers = args.timers with open(f'{args.snapshot_dir}/model', 'w') as f: f.write(str(self.model)) f.close() def _snapshot(self, epoch, name=None): state = { 'epoch': epoch, 'state_dict': self.model.state_dict(), 'optimizer': self.optimizer.state_dict(), 'scheduler': self.scheduler.state_dict(), 'best_loss': self.best_loss, 'best_recall': self.best_recall } if name is None: filename = os.path.join(self.save_dir, f'model_{epoch}.pth') else: filename = os.path.join(self.save_dir, f'model_{name}.pth') self.logger.write(f"Save model to {filename}\n") torch.save(state, filename, _use_new_zipfile_serialization=False) def _load_pretrain(self, resume): print ("loading pretrained", resume) if os.path.isfile(resume): state = torch.load(resume) self.model.load_state_dict(state['state_dict']) self.start_epoch = state['epoch'] self.scheduler.load_state_dict(state['scheduler']) self.optimizer.load_state_dict(state['optimizer']) self.best_loss = state['best_loss'] self.best_recall = state['best_recall'] self.logger.write(f'Successfully load pretrained model from {resume}!\n') self.logger.write(f'Current best loss {self.best_loss}\n') self.logger.write(f'Current best recall {self.best_recall}\n') else: raise ValueError(f"=> no checkpoint found at '{resume}'") def _get_lr(self, group=0): return self.optimizer.param_groups[group]['lr'] def inference_one_batch(self, inputs, phase): assert phase in ['train', 'val', 'test'] inputs ['phase'] = phase if (phase == 'train'): self.model.train() if self.timers: self.timers.tic('forward pass') data = self.model(inputs, timers=self.timers) # [N1, C1], [N2, C2] if self.timers: self.timers.toc('forward pass') if self.timers: self.timers.tic('compute loss') loss_info = self.loss( data) if self.timers: self.timers.toc('compute loss') if self.timers: self.timers.tic('backprop') loss_info['loss'].backward() if self.timers: self.timers.toc('backprop') else: self.model.eval() with torch.no_grad(): data = self.model(inputs, timers=self.timers) # [N1, C1], [N2, C2] loss_info = self.loss(data) return loss_info def inference_one_epoch(self, epoch, phase): gc.collect() assert phase in ['train', 'val', 'test'] # init stats meter stats_meter = None # self.stats_meter() num_iter = int(len(self.loader[phase].dataset) // self.loader[phase].batch_size) # drop last incomplete batch c_loader_iter = self.loader[phase].__iter__() self.optimizer.zero_grad() for c_iter in tqdm(range(num_iter)): # loop through this epoch if self.timers: self.timers.tic('one_iteration') ################################## if self.timers: self.timers.tic('load batch') inputs = c_loader_iter.next() # for gpu_div_i, _ in enumerate(inputs): for k, v in inputs.items(): if type(v) == list: inputs [k] = [item.to(self.device) for item in v] elif type(v) in [ dict, float, type(None), np.ndarray]: pass else: inputs [k] = v.to(self.device) if self.timers: self.timers.toc('load batch') ################################## if self.timers: self.timers.tic('inference_one_batch') loss_info = self.inference_one_batch(inputs, phase) if self.timers: self.timers.toc('inference_one_batch') ################################################### # run optimisation # if self.timers: self.timers.tic('run optimisation') if ((c_iter + 1) % self.iter_size == 0 and phase == 'train'): gradient_valid = validate_gradient(self.model) if (gradient_valid): self.optimizer.step() else: self.logger.write('gradient not valid\n') self.optimizer.zero_grad() # if self.timers: self.timers.toc('run optimisation') ################################ torch.cuda.empty_cache() if stats_meter is None: stats_meter = dict() for key, _ in loss_info.items(): stats_meter[key] = AverageMeter() for key, value in loss_info.items(): stats_meter[key].update(value) if phase == 'train' : if (c_iter + 1) % self.verbose_freq == 0 and self.verbose : curr_iter = num_iter * (epoch - 1) + c_iter for key, value in stats_meter.items(): self.summary_writer.add_scalar(f'{phase}/{key}', value.avg, curr_iter) dump_mess=True if dump_mess: message = f'{phase} Epoch: {epoch} [{c_iter + 1:4d}/{num_iter}]' for key, value in stats_meter.items(): message += f'{key}: {value.avg:.2f}\t' self.logger.write(message + '\n') if self.timers: self.timers.toc('one_iteration') # report evaluation score at end of each epoch if phase in ['val', 'test']: for key, value in stats_meter.items(): self.summary_writer.add_scalar(f'{phase}/{key}', value.avg, epoch) message = f'{phase} Epoch: {epoch}' for key, value in stats_meter.items(): message += f'{key}: {value.avg:.2f}\t' self.logger.write(message + '\n') return stats_meter def train(self): print('start training...') for epoch in range(self.start_epoch, self.max_epoch): with torch.autograd.set_detect_anomaly(True): if self.timers: self.timers.tic('run one epoch') stats_meter = self.inference_one_epoch(epoch, 'train') if self.timers: self.timers.toc('run one epoch') self.scheduler.step() if 'overfit' in self.config.exp_dir : if stats_meter['loss'].avg < self.best_loss: self.best_loss = stats_meter['loss'].avg self._snapshot(epoch, 'best_loss') if self.timers: self.timers.print() else : # no validation step for overfitting if self.config.do_valid: stats_meter = self.inference_one_epoch(epoch, 'val') if stats_meter['loss'].avg < self.best_loss: self.best_loss = stats_meter['loss'].avg self._snapshot(epoch, 'best_loss') if self.timers: self.timers.print() # finish all epoch print("Training finish!")
8,861
34.590361
117
py
sngan.pytorch
sngan.pytorch-master/test.py
# -*- coding: utf-8 -*- # @Date : 2019-07-25 # @Author : Xinyu Gong (xy_gong@tamu.edu) # @Link : None # @Version : 0.0 from __future__ import absolute_import from __future__ import division from __future__ import print_function import cfg import models from functions import validate from utils.utils import set_log_dir, create_logger from utils.inception_score import _init_inception from utils.fid_score import create_inception_graph, check_or_download_inception import torch import os import numpy as np from tensorboardX import SummaryWriter torch.backends.cudnn.enabled = True torch.backends.cudnn.benchmark = True def main(): args = cfg.parse_args() torch.cuda.manual_seed(args.random_seed) assert args.exp_name assert args.load_path.endswith('.pth') assert os.path.exists(args.load_path) args.path_helper = set_log_dir('logs_eval', args.exp_name) logger = create_logger(args.path_helper['log_path'], phase='test') # set tf env _init_inception() inception_path = check_or_download_inception(None) create_inception_graph(inception_path) # import network gen_net = eval('models.'+args.model+'.Generator')(args=args).cuda() # fid stat if args.dataset.lower() == 'cifar10': fid_stat = 'fid_stat/fid_stats_cifar10_train.npz' else: raise NotImplementedError(f'no fid stat for {args.dataset.lower()}') assert os.path.exists(fid_stat) # initial fixed_z = torch.cuda.FloatTensor(np.random.normal(0, 1, (25, args.latent_dim))) # set writer logger.info(f'=> resuming from {args.load_path}') checkpoint_file = args.load_path assert os.path.exists(checkpoint_file) checkpoint = torch.load(checkpoint_file) if 'avg_gen_state_dict' in checkpoint: gen_net.load_state_dict(checkpoint['avg_gen_state_dict']) epoch = checkpoint['epoch'] logger.info(f'=> loaded checkpoint {checkpoint_file} (epoch {epoch})') else: gen_net.load_state_dict(checkpoint) logger.info(f'=> loaded checkpoint {checkpoint_file}') logger.info(args) writer_dict = { 'writer': SummaryWriter(args.path_helper['log_path']), 'valid_global_steps': 0, } inception_score, fid_score = validate(args, fixed_z, fid_stat, gen_net, writer_dict) logger.info(f'Inception score: {inception_score}, FID score: {fid_score}.') if __name__ == '__main__': main()
2,425
29.708861
88
py
sngan.pytorch
sngan.pytorch-master/functions.py
# -*- coding: utf-8 -*- # @Date : 2019-07-25 # @Author : Xinyu Gong (xy_gong@tamu.edu) # @Link : None # @Version : 0.0 import os import numpy as np import torch import torch.nn as nn from torchvision.utils import make_grid from imageio import imsave from tqdm import tqdm from copy import deepcopy import logging from utils.inception_score import get_inception_score from utils.fid_score import calculate_fid_given_paths logger = logging.getLogger(__name__) def train(args, gen_net: nn.Module, dis_net: nn.Module, gen_optimizer, dis_optimizer, gen_avg_param, train_loader, epoch, writer_dict, schedulers=None): writer = writer_dict['writer'] gen_step = 0 # train mode gen_net = gen_net.train() dis_net = dis_net.train() for iter_idx, (imgs, _) in enumerate(tqdm(train_loader)): global_steps = writer_dict['train_global_steps'] # Adversarial ground truths real_imgs = imgs.type(torch.cuda.FloatTensor) # Sample noise as generator input z = torch.cuda.FloatTensor(np.random.normal(0, 1, (imgs.shape[0], args.latent_dim))) # --------------------- # Train Discriminator # --------------------- dis_optimizer.zero_grad() real_validity = dis_net(real_imgs) fake_imgs = gen_net(z).detach() assert fake_imgs.size() == real_imgs.size() fake_validity = dis_net(fake_imgs) # cal loss d_loss = torch.mean(nn.ReLU(inplace=True)(1.0 - real_validity)) + \ torch.mean(nn.ReLU(inplace=True)(1 + fake_validity)) d_loss.backward() dis_optimizer.step() writer.add_scalar('d_loss', d_loss.item(), global_steps) # ----------------- # Train Generator # ----------------- if global_steps % args.n_critic == 0: gen_optimizer.zero_grad() gen_z = torch.cuda.FloatTensor(np.random.normal(0, 1, (args.gen_batch_size, args.latent_dim))) gen_imgs = gen_net(gen_z) fake_validity = dis_net(gen_imgs) # cal loss g_loss = -torch.mean(fake_validity) g_loss.backward() gen_optimizer.step() # adjust learning rate if schedulers: gen_scheduler, dis_scheduler = schedulers g_lr = gen_scheduler.step(global_steps) d_lr = dis_scheduler.step(global_steps) writer.add_scalar('LR/g_lr', g_lr, global_steps) writer.add_scalar('LR/d_lr', d_lr, global_steps) # moving average weight for p, avg_p in zip(gen_net.parameters(), gen_avg_param): avg_p.mul_(0.999).add_(0.001, p.data) writer.add_scalar('g_loss', g_loss.item(), global_steps) gen_step += 1 # verbose if gen_step and iter_idx % args.print_freq == 0: tqdm.write( "[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]" % (epoch, args.max_epoch, iter_idx % len(train_loader), len(train_loader), d_loss.item(), g_loss.item())) writer_dict['train_global_steps'] = global_steps + 1 def validate(args, fixed_z, fid_stat, gen_net: nn.Module, writer_dict): writer = writer_dict['writer'] global_steps = writer_dict['valid_global_steps'] # eval mode gen_net = gen_net.eval() # generate images sample_imgs = gen_net(fixed_z) img_grid = make_grid(sample_imgs, nrow=5, normalize=True, scale_each=True) # get fid and inception score fid_buffer_dir = os.path.join(args.path_helper['sample_path'], 'fid_buffer') os.makedirs(fid_buffer_dir) eval_iter = args.num_eval_imgs // args.eval_batch_size img_list = list() for iter_idx in tqdm(range(eval_iter), desc='sample images'): z = torch.cuda.FloatTensor(np.random.normal(0, 1, (args.eval_batch_size, args.latent_dim))) # Generate a batch of images gen_imgs = gen_net(z).mul_(127.5).add_(127.5).clamp_(0.0, 255.0).permute(0, 2, 3, 1).to('cpu', torch.uint8).numpy() for img_idx, img in enumerate(gen_imgs): file_name = os.path.join(fid_buffer_dir, f'iter{iter_idx}_b{img_idx}.png') imsave(file_name, img) img_list.extend(list(gen_imgs)) # get inception score logger.info('=> calculate inception score') mean, std = get_inception_score(img_list) # get fid score logger.info('=> calculate fid score') fid_score = calculate_fid_given_paths([fid_buffer_dir, fid_stat], inception_path=None) os.system('rm -r {}'.format(fid_buffer_dir)) writer.add_image('sampled_images', img_grid, global_steps) writer.add_scalar('Inception_score/mean', mean, global_steps) writer.add_scalar('Inception_score/std', std, global_steps) writer.add_scalar('FID_score', fid_score, global_steps) writer_dict['valid_global_steps'] = global_steps + 1 return mean, fid_score class LinearLrDecay(object): def __init__(self, optimizer, start_lr, end_lr, decay_start_step, decay_end_step): assert start_lr > end_lr self.optimizer = optimizer self.delta = (start_lr - end_lr) / (decay_end_step - decay_start_step) self.decay_start_step = decay_start_step self.decay_end_step = decay_end_step self.start_lr = start_lr self.end_lr = end_lr def step(self, current_step): if current_step <= self.decay_start_step: lr = self.start_lr elif current_step >= self.decay_end_step: lr = self.end_lr else: lr = self.start_lr - self.delta * (current_step - self.decay_start_step) for param_group in self.optimizer.param_groups: param_group['lr'] = lr return lr def load_params(model, new_param): for p, new_p in zip(model.parameters(), new_param): p.data.copy_(new_p) def copy_params(model): flatten = deepcopy(list(p.data for p in model.parameters())) return flatten
6,022
32.837079
123
py
sngan.pytorch
sngan.pytorch-master/datasets.py
import torch import torchvision.datasets as datasets import torchvision.transforms as transforms from torch.utils.data import Dataset class ImageDataset(object): def __init__(self, args): if args.dataset.lower() == 'cifar10': Dt = datasets.CIFAR10 transform = transforms.Compose([ transforms.Resize(args.img_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), ]) args.n_classes = 10 elif args.dataset.lower() == 'stl10': Dt = datasets.STL10 transform = transforms.Compose([ transforms.Resize(args.img_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), ]) else: raise NotImplementedError('Unknown dataset: {}'.format(args.dataset)) if args.dataset.lower() == 'stl10': self.train = torch.utils.data.DataLoader( Dt(root=args.data_path, split='train+unlabeled', transform=transform, download=True), batch_size=args.dis_batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) self.valid = torch.utils.data.DataLoader( Dt(root=args.data_path, split='test', transform=transform), batch_size=args.dis_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) self.test = self.valid else: self.train = torch.utils.data.DataLoader( Dt(root=args.data_path, train=True, transform=transform, download=True), batch_size=args.dis_batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) self.valid = torch.utils.data.DataLoader( Dt(root=args.data_path, train=False, transform=transform), batch_size=args.dis_batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) self.test = self.valid
2,115
40.490196
101
py
sngan.pytorch
sngan.pytorch-master/train.py
# -*- coding: utf-8 -*- # @Date : 2019-07-25 # @Author : Xinyu Gong (xy_gong@tamu.edu) # @Link : None # @Version : 0.0 from __future__ import absolute_import from __future__ import division from __future__ import print_function import cfg import models import datasets from functions import train, validate, LinearLrDecay, load_params, copy_params from utils.utils import set_log_dir, save_checkpoint, create_logger from utils.inception_score import _init_inception from utils.fid_score import create_inception_graph, check_or_download_inception import torch import os import numpy as np import torch.nn as nn from tensorboardX import SummaryWriter from tqdm import tqdm from copy import deepcopy torch.backends.cudnn.enabled = True torch.backends.cudnn.benchmark = True def main(): args = cfg.parse_args() torch.cuda.manual_seed(args.random_seed) # set tf env _init_inception() inception_path = check_or_download_inception(None) create_inception_graph(inception_path) # import network gen_net = eval('models.'+args.model+'.Generator')(args=args).cuda() dis_net = eval('models.'+args.model+'.Discriminator')(args=args).cuda() # weight init def weights_init(m): classname = m.__class__.__name__ if classname.find('Conv2d') != -1: if args.init_type == 'normal': nn.init.normal_(m.weight.data, 0.0, 0.02) elif args.init_type == 'orth': nn.init.orthogonal_(m.weight.data) elif args.init_type == 'xavier_uniform': nn.init.xavier_uniform(m.weight.data, 1.) else: raise NotImplementedError('{} unknown inital type'.format(args.init_type)) elif classname.find('BatchNorm2d') != -1: nn.init.normal_(m.weight.data, 1.0, 0.02) nn.init.constant_(m.bias.data, 0.0) gen_net.apply(weights_init) dis_net.apply(weights_init) # set optimizer gen_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, gen_net.parameters()), args.g_lr, (args.beta1, args.beta2)) dis_optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, dis_net.parameters()), args.d_lr, (args.beta1, args.beta2)) gen_scheduler = LinearLrDecay(gen_optimizer, args.g_lr, 0.0, 0, args.max_iter * args.n_critic) dis_scheduler = LinearLrDecay(dis_optimizer, args.d_lr, 0.0, 0, args.max_iter * args.n_critic) # set up data_loader dataset = datasets.ImageDataset(args) train_loader = dataset.train # fid stat if args.dataset.lower() == 'cifar10': fid_stat = 'fid_stat/fid_stats_cifar10_train.npz' elif args.dataset.lower() == 'stl10': fid_stat = 'fid_stat/stl10_train_unlabeled_fid_stats_48.npz' else: raise NotImplementedError(f'no fid stat for {args.dataset.lower()}') assert os.path.exists(fid_stat) # epoch number for dis_net args.max_epoch = args.max_epoch * args.n_critic if args.max_iter: args.max_epoch = np.ceil(args.max_iter * args.n_critic / len(train_loader)) # initial fixed_z = torch.cuda.FloatTensor(np.random.normal(0, 1, (25, args.latent_dim))) gen_avg_param = copy_params(gen_net) start_epoch = 0 best_fid = 1e4 # set writer if args.load_path: print(f'=> resuming from {args.load_path}') assert os.path.exists(args.load_path) checkpoint_file = os.path.join(args.load_path, 'Model', 'checkpoint.pth') assert os.path.exists(checkpoint_file) checkpoint = torch.load(checkpoint_file) start_epoch = checkpoint['epoch'] best_fid = checkpoint['best_fid'] gen_net.load_state_dict(checkpoint['gen_state_dict']) dis_net.load_state_dict(checkpoint['dis_state_dict']) gen_optimizer.load_state_dict(checkpoint['gen_optimizer']) dis_optimizer.load_state_dict(checkpoint['dis_optimizer']) avg_gen_net = deepcopy(gen_net) avg_gen_net.load_state_dict(checkpoint['avg_gen_state_dict']) gen_avg_param = copy_params(avg_gen_net) del avg_gen_net args.path_helper = checkpoint['path_helper'] logger = create_logger(args.path_helper['log_path']) logger.info(f'=> loaded checkpoint {checkpoint_file} (epoch {start_epoch})') else: # create new log dir assert args.exp_name args.path_helper = set_log_dir('logs', args.exp_name) logger = create_logger(args.path_helper['log_path']) logger.info(args) writer_dict = { 'writer': SummaryWriter(args.path_helper['log_path']), 'train_global_steps': start_epoch * len(train_loader), 'valid_global_steps': start_epoch // args.val_freq, } # train loop lr_schedulers = (gen_scheduler, dis_scheduler) if args.lr_decay else None for epoch in tqdm(range(int(start_epoch), int(args.max_epoch)), desc='total progress'): train(args, gen_net, dis_net, gen_optimizer, dis_optimizer, gen_avg_param, train_loader, epoch, writer_dict, lr_schedulers) if epoch and epoch % args.val_freq == 0 or epoch == int(args.max_epoch)-1: backup_param = copy_params(gen_net) load_params(gen_net, gen_avg_param) inception_score, fid_score = validate(args, fixed_z, fid_stat, gen_net, writer_dict) logger.info(f'Inception score: {inception_score}, FID score: {fid_score} || @ epoch {epoch}.') load_params(gen_net, backup_param) if fid_score < best_fid: best_fid = fid_score is_best = True else: is_best = False else: is_best = False avg_gen_net = deepcopy(gen_net) load_params(avg_gen_net, gen_avg_param) save_checkpoint({ 'epoch': epoch + 1, 'model': args.model, 'gen_state_dict': gen_net.state_dict(), 'dis_state_dict': dis_net.state_dict(), 'avg_gen_state_dict': avg_gen_net.state_dict(), 'gen_optimizer': gen_optimizer.state_dict(), 'dis_optimizer': dis_optimizer.state_dict(), 'best_fid': best_fid, 'path_helper': args.path_helper }, is_best, args.path_helper['ckpt_path']) del avg_gen_net if __name__ == '__main__': main()
6,393
37.287425
116
py
sngan.pytorch
sngan.pytorch-master/models/sngan_64.py
import torch.nn as nn class GenBlock(nn.Module): def __init__(self, in_channels, out_channels, hidden_channels=None, ksize=3, pad=1, activation=nn.ReLU(), upsample=False, n_classes=0): super(GenBlock, self).__init__() self.activation = activation self.upsample = upsample self.learnable_sc = in_channels != out_channels or upsample hidden_channels = out_channels if hidden_channels is None else hidden_channels self.n_classes = n_classes self.c1 = nn.Conv2d(in_channels, hidden_channels, kernel_size=ksize, padding=pad) self.c2 = nn.Conv2d(hidden_channels, out_channels, kernel_size=ksize, padding=pad) self.b1 = nn.BatchNorm2d(in_channels) self.b2 = nn.BatchNorm2d(hidden_channels) if self.learnable_sc: self.c_sc = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0) def upsample_conv(self, x, conv): return conv(nn.UpsamplingNearest2d(scale_factor=2)(x)) def residual(self, x): h = x h = self.b1(h) h = self.activation(h) h = self.upsample_conv(h, self.c1) if self.upsample else self.c1(h) h = self.b2(h) h = self.activation(h) h = self.c2(h) return h def shortcut(self, x): if self.learnable_sc: x = self.upsample_conv(x, self.c_sc) if self.upsample else self.c_sc(x) return x else: return x def forward(self, x): return self.residual(x) + self.shortcut(x) class Generator(nn.Module): def __init__(self, args, activation=nn.ReLU(), n_classes=0): super(Generator, self).__init__() self.bottom_width = args.bottom_width self.activation = activation self.n_classes = n_classes self.ch = args.gf_dim self.l1 = nn.Linear(args.latent_dim, (self.bottom_width ** 2) * self.ch) self.block2 = GenBlock(self.ch, self.ch, activation=activation, upsample=True, n_classes=n_classes) self.block3 = GenBlock(self.ch, self.ch, activation=activation, upsample=True, n_classes=n_classes) self.block4 = GenBlock(self.ch, self.ch, activation=activation, upsample=True, n_classes=n_classes) self.b5 = nn.BatchNorm2d(self.ch) self.c5 = nn.Conv2d(self.ch, 3, kernel_size=3, stride=1, padding=1) def forward(self, z): h = z h = self.l1(h).view(-1, self.ch, self.bottom_width, self.bottom_width) h = self.block2(h) h = self.block3(h) h = self.block4(h) h = self.b5(h) h = self.activation(h) h = nn.Tanh()(self.c5(h)) return h """Discriminator""" def _downsample(x): # Downsample (Mean Avg Pooling with 2x2 kernel) return nn.AvgPool2d(kernel_size=2)(x) class OptimizedDisBlock(nn.Module): def __init__(self, args, in_channels, out_channels, ksize=3, pad=1, activation=nn.ReLU()): super(OptimizedDisBlock, self).__init__() self.activation = activation self.c1 = nn.Conv2d(in_channels, out_channels, kernel_size=ksize, padding=pad) self.c2 = nn.Conv2d(out_channels, out_channels, kernel_size=ksize, padding=pad) self.c_sc = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0) if args.d_spectral_norm: self.c1 = nn.utils.spectral_norm(self.c1) self.c2 = nn.utils.spectral_norm(self.c2) self.c_sc = nn.utils.spectral_norm(self.c_sc) def residual(self, x): h = x h = self.c1(h) h = self.activation(h) h = self.c2(h) h = _downsample(h) return h def shortcut(self, x): return self.c_sc(_downsample(x)) def forward(self, x): return self.residual(x) + self.shortcut(x) class DisBlock(nn.Module): def __init__(self, args, in_channels, out_channels, hidden_channels=None, ksize=3, pad=1, activation=nn.ReLU(), downsample=False): super(DisBlock, self).__init__() self.activation = activation self.downsample = downsample self.learnable_sc = (in_channels != out_channels) or downsample hidden_channels = in_channels if hidden_channels is None else hidden_channels self.c1 = nn.Conv2d(in_channels, hidden_channels, kernel_size=ksize, padding=pad) self.c2 = nn.Conv2d(hidden_channels, out_channels, kernel_size=ksize, padding=pad) if args.d_spectral_norm: self.c1 = nn.utils.spectral_norm(self.c1) self.c2 = nn.utils.spectral_norm(self.c2) if self.learnable_sc: self.c_sc = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0) if args.d_spectral_norm: self.c_sc = nn.utils.spectral_norm(self.c_sc) def residual(self, x): h = x h = self.activation(h) h = self.c1(h) h = self.activation(h) h = self.c2(h) if self.downsample: h = _downsample(h) return h def shortcut(self, x): if self.learnable_sc: x = self.c_sc(x) if self.downsample: return _downsample(x) else: return x else: return x def forward(self, x): return self.residual(x) + self.shortcut(x) class Discriminator(nn.Module): def __init__(self, args, activation=nn.ReLU()): super(Discriminator, self).__init__() self.ch = args.df_dim self.activation = activation self.block1 = OptimizedDisBlock(args, 3, self.ch) self.block2 = DisBlock(args, self.ch, self.ch, activation=activation, downsample=True) self.block3 = DisBlock(args, self.ch, self.ch, activation=activation, downsample=False) self.block4 = DisBlock(args, self.ch, self.ch, activation=activation, downsample=False) self.l5 = nn.Linear(self.ch, 1, bias=False) if args.d_spectral_norm: self.l5 = nn.utils.spectral_norm(self.l5) def forward(self, x): h = x h = self.block1(h) h = self.block2(h) h = self.block3(h) h = self.block4(h) h = self.activation(h) # Global average pooling h = h.sum(2).sum(2) output = self.l5(h) return output
6,296
34.778409
107
py
sngan.pytorch
sngan.pytorch-master/models/sngan_stl10.py
import torch.nn as nn class GenBlock(nn.Module): def __init__(self, in_channels, out_channels, hidden_channels=None, ksize=3, pad=1, activation=nn.ReLU(), upsample=False, n_classes=0): super(GenBlock, self).__init__() self.activation = activation self.upsample = upsample self.learnable_sc = in_channels != out_channels or upsample hidden_channels = out_channels if hidden_channels is None else hidden_channels self.n_classes = n_classes self.c1 = nn.Conv2d(in_channels, hidden_channels, kernel_size=ksize, padding=pad) self.c2 = nn.Conv2d(hidden_channels, out_channels, kernel_size=ksize, padding=pad) self.b1 = nn.BatchNorm2d(in_channels) self.b2 = nn.BatchNorm2d(hidden_channels) if self.learnable_sc: self.c_sc = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0) def upsample_conv(self, x, conv): return conv(nn.UpsamplingNearest2d(scale_factor=2)(x)) def residual(self, x): h = x h = self.b1(h) h = self.activation(h) h = self.upsample_conv(h, self.c1) if self.upsample else self.c1(h) h = self.b2(h) h = self.activation(h) h = self.c2(h) return h def shortcut(self, x): if self.learnable_sc: x = self.upsample_conv(x, self.c_sc) if self.upsample else self.c_sc(x) return x else: return x def forward(self, x): return self.residual(x) + self.shortcut(x) class Generator(nn.Module): def __init__(self, args, activation=nn.ReLU(), n_classes=0): super(Generator, self).__init__() self.bottom_width = args.bottom_width self.activation = activation self.n_classes = n_classes self.ch = 512 self.l1 = nn.Linear(args.latent_dim, (self.bottom_width ** 2) * self.ch) self.block2 = GenBlock(512, 256, activation=activation, upsample=True, n_classes=n_classes) self.block3 = GenBlock(256, 128, activation=activation, upsample=True, n_classes=n_classes) self.block4 = GenBlock(128, 64, activation=activation, upsample=True, n_classes=n_classes) self.b5 = nn.BatchNorm2d(64) self.c5 = nn.Conv2d(64, 3, kernel_size=3, stride=1, padding=1) def forward(self, z): h = z h = self.l1(h).view(-1, self.ch, self.bottom_width, self.bottom_width) h = self.block2(h) h = self.block3(h) h = self.block4(h) h = self.b5(h) h = self.activation(h) h = nn.Tanh()(self.c5(h)) return h """Discriminator""" def _downsample(x): # Downsample (Mean Avg Pooling with 2x2 kernel) return nn.AvgPool2d(kernel_size=2)(x) class OptimizedDisBlock(nn.Module): def __init__(self, args, in_channels, out_channels, ksize=3, pad=1, activation=nn.ReLU()): super(OptimizedDisBlock, self).__init__() self.activation = activation self.c1 = nn.Conv2d(in_channels, out_channels, kernel_size=ksize, padding=pad) self.c2 = nn.Conv2d(out_channels, out_channels, kernel_size=ksize, padding=pad) self.c_sc = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0) if args.d_spectral_norm: self.c1 = nn.utils.spectral_norm(self.c1) self.c2 = nn.utils.spectral_norm(self.c2) self.c_sc = nn.utils.spectral_norm(self.c_sc) def residual(self, x): h = x h = self.c1(h) h = self.activation(h) h = self.c2(h) h = _downsample(h) return h def shortcut(self, x): return self.c_sc(_downsample(x)) def forward(self, x): return self.residual(x) + self.shortcut(x) class DisBlock(nn.Module): def __init__(self, args, in_channels, out_channels, hidden_channels=None, ksize=3, pad=1, activation=nn.ReLU(), downsample=False): super(DisBlock, self).__init__() self.activation = activation self.downsample = downsample self.learnable_sc = (in_channels != out_channels) or downsample hidden_channels = in_channels if hidden_channels is None else hidden_channels self.c1 = nn.Conv2d(in_channels, hidden_channels, kernel_size=ksize, padding=pad) self.c2 = nn.Conv2d(hidden_channels, out_channels, kernel_size=ksize, padding=pad) if args.d_spectral_norm: self.c1 = nn.utils.spectral_norm(self.c1) self.c2 = nn.utils.spectral_norm(self.c2) if self.learnable_sc: self.c_sc = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0) if args.d_spectral_norm: self.c_sc = nn.utils.spectral_norm(self.c_sc) def residual(self, x): h = x h = self.activation(h) h = self.c1(h) h = self.activation(h) h = self.c2(h) if self.downsample: h = _downsample(h) return h def shortcut(self, x): if self.learnable_sc: x = self.c_sc(x) if self.downsample: return _downsample(x) else: return x else: return x def forward(self, x): return self.residual(x) + self.shortcut(x) class Discriminator(nn.Module): def __init__(self, args, activation=nn.ReLU()): super(Discriminator, self).__init__() self.activation = activation self.block1 = OptimizedDisBlock(args, 3, 64) self.block2 = DisBlock(args, 64, 128, activation=activation, downsample=True) self.block3 = DisBlock(args, 128, 256, activation=activation, downsample=True) self.block4 = DisBlock(args, 256, 512, activation=activation, downsample=True) self.block5 = DisBlock(args, 512, 1024, activation=activation, downsample=False) self.l6 = nn.Linear(1024, 1, bias=False) if args.d_spectral_norm: self.l6 = nn.utils.spectral_norm(self.l6) def forward(self, x): h = x h = self.block1(h) h = self.block2(h) h = self.block3(h) h = self.block4(h) h = self.block5(h) h = self.activation(h) # Global average pooling h = h.sum(2).sum(2) output = self.l6(h) return output
6,305
34.426966
99
py
sngan.pytorch
sngan.pytorch-master/models/sngan_cifar10.py
import torch.nn as nn from .gen_resblock import GenBlock class Generator(nn.Module): def __init__(self, args, activation=nn.ReLU(), n_classes=0): super(Generator, self).__init__() self.bottom_width = args.bottom_width self.activation = activation self.n_classes = n_classes self.ch = args.gf_dim self.l1 = nn.Linear(args.latent_dim, (self.bottom_width ** 2) * self.ch) self.block2 = GenBlock(self.ch, self.ch, activation=activation, upsample=True, n_classes=n_classes) self.block3 = GenBlock(self.ch, self.ch, activation=activation, upsample=True, n_classes=n_classes) self.block4 = GenBlock(self.ch, self.ch, activation=activation, upsample=True, n_classes=n_classes) self.b5 = nn.BatchNorm2d(self.ch) self.c5 = nn.Conv2d(self.ch, 3, kernel_size=3, stride=1, padding=1) def forward(self, z): h = z h = self.l1(h).view(-1, self.ch, self.bottom_width, self.bottom_width) h = self.block2(h) h = self.block3(h) h = self.block4(h) h = self.b5(h) h = self.activation(h) h = nn.Tanh()(self.c5(h)) return h """Discriminator""" def _downsample(x): # Downsample (Mean Avg Pooling with 2x2 kernel) return nn.AvgPool2d(kernel_size=2)(x) class OptimizedDisBlock(nn.Module): def __init__(self, args, in_channels, out_channels, ksize=3, pad=1, activation=nn.ReLU()): super(OptimizedDisBlock, self).__init__() self.activation = activation self.c1 = nn.Conv2d(in_channels, out_channels, kernel_size=ksize, padding=pad) self.c2 = nn.Conv2d(out_channels, out_channels, kernel_size=ksize, padding=pad) self.c_sc = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0) if args.d_spectral_norm: self.c1 = nn.utils.spectral_norm(self.c1) self.c2 = nn.utils.spectral_norm(self.c2) self.c_sc = nn.utils.spectral_norm(self.c_sc) def residual(self, x): h = x h = self.c1(h) h = self.activation(h) h = self.c2(h) h = _downsample(h) return h def shortcut(self, x): return self.c_sc(_downsample(x)) def forward(self, x): return self.residual(x) + self.shortcut(x) class DisBlock(nn.Module): def __init__(self, args, in_channels, out_channels, hidden_channels=None, ksize=3, pad=1, activation=nn.ReLU(), downsample=False): super(DisBlock, self).__init__() self.activation = activation self.downsample = downsample self.learnable_sc = (in_channels != out_channels) or downsample hidden_channels = in_channels if hidden_channels is None else hidden_channels self.c1 = nn.Conv2d(in_channels, hidden_channels, kernel_size=ksize, padding=pad) self.c2 = nn.Conv2d(hidden_channels, out_channels, kernel_size=ksize, padding=pad) if args.d_spectral_norm: self.c1 = nn.utils.spectral_norm(self.c1) self.c2 = nn.utils.spectral_norm(self.c2) if self.learnable_sc: self.c_sc = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0) if args.d_spectral_norm: self.c_sc = nn.utils.spectral_norm(self.c_sc) def residual(self, x): h = x h = self.activation(h) h = self.c1(h) h = self.activation(h) h = self.c2(h) if self.downsample: h = _downsample(h) return h def shortcut(self, x): if self.learnable_sc: x = self.c_sc(x) if self.downsample: return _downsample(x) else: return x else: return x def forward(self, x): return self.residual(x) + self.shortcut(x) class Discriminator(nn.Module): def __init__(self, args, activation=nn.ReLU()): super(Discriminator, self).__init__() self.ch = args.df_dim self.activation = activation self.block1 = OptimizedDisBlock(args, 3, self.ch) self.block2 = DisBlock(args, self.ch, self.ch, activation=activation, downsample=True) self.block3 = DisBlock(args, self.ch, self.ch, activation=activation, downsample=False) self.block4 = DisBlock(args, self.ch, self.ch, activation=activation, downsample=False) self.l5 = nn.Linear(self.ch, 1, bias=False) if args.d_spectral_norm: self.l5 = nn.utils.spectral_norm(self.l5) def forward(self, x): h = x h = self.block1(h) h = self.block2(h) h = self.block3(h) h = self.block4(h) h = self.activation(h) # Global average pooling h = h.sum(2).sum(2) output = self.l5(h) return output
4,805
34.338235
107
py
sngan.pytorch
sngan.pytorch-master/models/gen_resblock.py
# -*- coding: utf-8 -*- # @Date : 3/26/20 # @Author : Xinyu Gong (xy_gong@tamu.edu) # @Link : None # @Version : 0.0 import torch.nn as nn class GenBlock(nn.Module): def __init__(self, in_channels, out_channels, hidden_channels=None, ksize=3, pad=1, activation=nn.ReLU(), upsample=False, n_classes=0): super(GenBlock, self).__init__() self.activation = activation self.upsample = upsample self.learnable_sc = in_channels != out_channels or upsample hidden_channels = out_channels if hidden_channels is None else hidden_channels self.n_classes = n_classes self.c1 = nn.Conv2d(in_channels, hidden_channels, kernel_size=ksize, padding=pad) self.c2 = nn.Conv2d(hidden_channels, out_channels, kernel_size=ksize, padding=pad) self.b1 = nn.BatchNorm2d(in_channels) self.b2 = nn.BatchNorm2d(hidden_channels) if self.learnable_sc: self.c_sc = nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0) def upsample_conv(self, x, conv): return conv(nn.UpsamplingNearest2d(scale_factor=2)(x)) def residual(self, x): h = x h = self.b1(h) h = self.activation(h) h = self.upsample_conv(h, self.c1) if self.upsample else self.c1(h) h = self.b2(h) h = self.activation(h) h = self.c2(h) return h def shortcut(self, x): if self.learnable_sc: x = self.upsample_conv(x, self.c_sc) if self.upsample else self.c_sc(x) return x else: return x def forward(self, x): return self.residual(x) + self.shortcut(x)
1,671
33.833333
90
py
sngan.pytorch
sngan.pytorch-master/utils/utils.py
# -*- coding: utf-8 -*- # @Date : 2019-07-25 # @Author : Xinyu Gong (xy_gong@tamu.edu) # @Link : None # @Version : 0.0 import os import torch import dateutil.tz from datetime import datetime import time import logging def create_logger(log_dir, phase='train'): time_str = time.strftime('%Y-%m-%d-%H-%M') log_file = '{}_{}.log'.format(time_str, phase) final_log_file = os.path.join(log_dir, log_file) head = '%(asctime)-15s %(message)s' logging.basicConfig(filename=str(final_log_file), format=head) logger = logging.getLogger() logger.setLevel(logging.INFO) console = logging.StreamHandler() logging.getLogger('').addHandler(console) return logger def set_log_dir(root_dir, exp_name): path_dict = {} os.makedirs(root_dir, exist_ok=True) # set log path exp_path = os.path.join(root_dir, exp_name) now = datetime.now(dateutil.tz.tzlocal()) timestamp = now.strftime('%Y_%m_%d_%H_%M_%S') prefix = exp_path + '_' + timestamp os.makedirs(prefix) path_dict['prefix'] = prefix # set checkpoint path ckpt_path = os.path.join(prefix, 'Model') os.makedirs(ckpt_path) path_dict['ckpt_path'] = ckpt_path log_path = os.path.join(prefix, 'Log') os.makedirs(log_path) path_dict['log_path'] = log_path # set sample image path for fid calculation sample_path = os.path.join(prefix, 'Samples') os.makedirs(sample_path) path_dict['sample_path'] = sample_path return path_dict def save_checkpoint(states, is_best, output_dir, filename='checkpoint.pth'): torch.save(states, os.path.join(output_dir, filename)) if is_best: torch.save(states, os.path.join(output_dir, 'checkpoint_best.pth'))
1,772
26.703125
75
py
neu-nbv
neu-nbv-main/scripts/planning/dtu_experiment.py
import sys import os root_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..")) sys.path.insert(0, root_dir) from neural_rendering.evaluation.pretrained_model import PretrainedModel from neural_rendering.data import get_data from neural_rendering.utils import parser, util import yaml from dotmap import DotMap import torch import warnings import numpy as np import pandas import seaborn as sb import copy from scipy.spatial import distance from datetime import datetime import random import pickle from dotmap import DotMap warnings.filterwarnings("ignore") # follow pixelnerf setup candidate_index_list = [ 6, 7, 8, 9, 10, 13, 14, 15, 16, 17, 21, 22, 23, 24, 25, 31, 32, 33, 34, 35, 41, 42, 43, 44, 45, ] def setup_random_seed(seed): torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) np.random.seed(seed) random.seed(seed) torch.backends.cudnn.deterministic = True def get_nbv_ref_index( model, images, poses, focal, c, z_near, z_far, candidate_list, budget, ref_index ): _, _, H, W = images.shape for i in range(budget): remain_candidate_list = list(set(candidate_list) - set(ref_index)) reward_list = [] model.network.encode( images[ref_index].unsqueeze(0), poses[ref_index].unsqueeze(0), focal.unsqueeze(0), c.unsqueeze(0), ) for target_view in remain_candidate_list: novel_pose = poses[target_view] target_rays = util.gen_rays( novel_pose.unsqueeze(0), W, H, focal, z_near, z_far, c ) target_rays = target_rays.reshape(1, H * W, -1) predict = DotMap(model.renderer_par(target_rays)) uncertainty = predict["uncertainty"][0] reward = torch.sum(uncertainty**2).cpu().numpy() reward_list.append(reward) nbv_index = np.argmax(reward_list) new_ref_index = remain_candidate_list[nbv_index] ref_index.append(new_ref_index) return ref_index def get_camera_view_direction(poses): poses = poses.cpu().numpy() view_direction = -poses[..., :3, 2] view_direction = view_direction / np.linalg.norm(view_direction) return view_direction def get_max_dist_ref_index(poses, ref_index, candidate_list, budget): view_direction = get_camera_view_direction(poses) for i in range(budget): remain_candidate_list = list(set(candidate_list) - set(ref_index)) cos_distance_list = [] for idx in remain_candidate_list: cos_dist = 0.0 for image_idx in ref_index: cos_dist += distance.cosine( view_direction[idx], view_direction[image_idx] ) cos_distance_list.append(cos_dist) new_ref_index = remain_candidate_list[np.argmax(cos_distance_list)] ref_index.append(new_ref_index) return ref_index def main(): # planning experiment on DTU using baseline planners and our planner setup_random_seed(10) args = parser.parse_args(planning_args) dtu_nbv_planner = DTUNBVPlanning(args) experiment_path = args.experiment_path if args.evaluation_only: with open(f"{experiment_path}/saved_index_dict.pkl", "rb") as f: index_record = pickle.load(f) else: experiment_path = os.path.join( root_dir, "experiments", "dtu", datetime.now().strftime("%d-%m-%Y-%H-%M"), ) os.makedirs(experiment_path) index_record = dtu_nbv_planner.planning() with open(f"{experiment_path}/saved_index_dict.pkl", "wb") as f: pickle.dump(index_record, f) total_df = dtu_nbv_planner.evaluation(index_record) total_df.to_csv(f"{experiment_path}/dataframe.csv") class DTUNBVPlanning: """ planning on DTU using different view selection methods: max_view_distance, random, and our uncertainty guided """ def __init__(self, args): log_path = os.path.join(root_dir, "neural_rendering", "logs", args.model_name) assert os.path.exists(log_path), "experiment does not exist" with open(f"{log_path}/training_setup.yaml", "r") as config_file: cfg = yaml.safe_load(config_file) checkpoint_path = os.path.join(log_path, "checkpoints", "best.ckpt") assert os.path.exists(checkpoint_path), "checkpoint does not exist" ckpt_file = torch.load(checkpoint_path) gpu_id = list(map(int, args.gpu_id.split())) self.device = util.get_cuda(gpu_id[0]) self.repeat = args.repeat self.model = PretrainedModel(cfg["model"], ckpt_file, self.device, gpu_id) cfg["data"]["dataset"]["data_rootdir"] = os.path.join( root_dir, "neural_rendering/data/dataset/dtu_dataset/rs_dtu_4/DTU" ) datamodule = get_data(cfg["data"]) self.dataset = datamodule.load_dataset("val") self.z_near = self.dataset.z_near self.z_far = self.dataset.z_far def planning(self): print(f"---------- planning ---------- \n") ON = len(self.dataset) selection_type = ["Max. View Distance", "Random", "Ours"] nview_list = [2, 3, 4, 5, 6, 7, 8, 9] # maximal budget = 9 scene_index = range(ON) ref_index_record = {} with torch.no_grad(): for nviews in nview_list: ref_index_record[nviews] = {} print(f"---------- {nviews} views experiment---------- \n") for i in scene_index: data_instance = self.dataset.__getitem__(i) scene_title = data_instance["scan_name"] ref_index_record[nviews][i] = {} print(f"test on {scene_title}") images = data_instance["images"].to(self.device) focal = data_instance["focal"].to(self.device) c = data_instance["c"].to(self.device) poses = data_instance["poses"].to(self.device) # random initialize first 2 ref images for all methods for r in range(self.repeat): ref_index_record[nviews][i][r] = {} initial_ref_index = list( np.random.choice(candidate_index_list, 2, replace=False) ) candidate_list = list( set(candidate_index_list) - set(initial_ref_index) ) budget = nviews - 2 for stype in selection_type: print(f"---------- repeat: {r}, {stype} ---------- \n") if stype == "Max. View Distance": ref_index = get_max_dist_ref_index( poses, copy.deepcopy(initial_ref_index), candidate_list, budget, ) print(ref_index) elif stype == "Random": random_ref_index = list( np.random.choice( candidate_index_list, budget, replace=True ) ) ref_index = initial_ref_index + random_ref_index print(ref_index) ref_index = np.unique(ref_index) elif stype == "Ours": ref_index = get_nbv_ref_index( self.model, images, poses, focal, c, self.z_near, self.z_far, candidate_list, budget, copy.deepcopy(initial_ref_index), ) print(ref_index) ref_index_record[nviews][i][r][stype] = ref_index return ref_index_record def evaluation(self, index_record): print(f"---------- evaluation ---------- \n") total_df = pandas.DataFrame( { "Planning Type": [], "Reference Image Number": [], "PSNR": [], "SSIM": [], "Scene": [], } ) with torch.no_grad(): for nviews, nviews_dict in index_record.items(): print(f"---------- {nviews} views experiment---------- \n") for scene_id, scene_dict in nviews_dict.items(): data_instance = self.dataset.__getitem__(scene_id) scene_title = data_instance["scan_name"] print(f"test on {scene_title}") images = data_instance["images"].to(self.device) images_0to1 = images * 0.5 + 0.5 _, _, H, W = images.shape focal = data_instance["focal"].to(self.device) c = data_instance["c"].to(self.device) poses = data_instance["poses"].to(self.device) psnr_per_scene = [] ssim_per_scene = [] # random initialize first 2 ref images for all methods for repeat, repeat_dict in scene_dict.items(): for stype, ref_index in repeat_dict.items(): print(f"---------- repeat: {repeat}, {stype} ---------- \n") print(ref_index) self.model.network.encode( images[ref_index].unsqueeze(0), poses[ref_index].unsqueeze(0), focal.unsqueeze(0), c.unsqueeze(0), ) test_index = list( set(candidate_index_list) - set(ref_index) ) psnr_per_test = [] ssim_per_test = [] for target_view in test_index: gt = ( images_0to1[target_view] .permute(1, 2, 0) .cpu() .numpy() ) novel_pose = poses[target_view] target_rays = util.gen_rays( novel_pose.unsqueeze(0), W, H, focal, self.z_near, self.z_far, c, ) target_rays = target_rays.reshape(1, H * W, -1) predict = DotMap(self.model.renderer_par(target_rays)) metrics_dict = util.calc_metrics( predict, torch.tensor(gt) ) psnr_per_test.append(metrics_dict["psnr"]) ssim_per_test.append(metrics_dict["ssim"]) psnr_per_scene = np.mean(psnr_per_test) ssim_per_scene = np.mean(ssim_per_test) print(psnr_per_scene, ssim_per_scene) dataframe = pandas.DataFrame( { "Planning Type": stype, "Reference Image Number": nviews, "PSNR": psnr_per_scene, "SSIM": ssim_per_scene, "Scene": scene_id, }, index=[repeat], ) total_df = total_df.append(dataframe) return total_df def planning_args(parser): """ Parse arguments for evaluation setup. """ parser.add_argument( "--model_name", "-M", type=str, required=True, help="model name of pretrained model", ) parser.add_argument( "--repeat", "-R", type=int, default=5, help="repeat times for planning experiment", ) # arguments with default values parser.add_argument( "--evaluation_only", action="store_true", help="evaluation mode" ) parser.add_argument( "--experiment_path", type=str, default="not defined", help="must be defined in evaluation mode", ) parser.add_argument( "--gpu_id", type=str, default="0", help="GPU(s) to use, space delimited" ) return parser if __name__ == "__main__": main()
13,632
34.046272
113
py
neu-nbv
neu-nbv-main/scripts/planning/simulator_experiment.py
import rospy import os import sys root_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..")) sys.path.insert(0, root_dir) import yaml import argparse from planner import get_planner from planner.utils import uniform_sampling import numpy as np import scipy.spatial as spatial from datetime import datetime import imageio import glob from dotmap import DotMap import torch from neural_rendering.utils import util from neural_rendering.evaluation.pretrained_model import PretrainedModel import pandas import torch.nn.functional as F planner_title = { "max_distance": "Max. View Distance", "random": "Random", "neural_nbv": "Ours", } def setup_random_seed(seed): np.random.seed(seed) def main(): # planning experiment in simulator using baseline planners and our planner setup_random_seed(10) rospy.init_node("simulator_experiment") args = parse_args() planner_type_list = ["max_distance", "random", "neural_nbv"] repeat = args.repeat experiment_path = args.experiment_path if not args.evaluation_only: experiment_path = os.path.join( root_dir, "experiments", "simulator", datetime.now().strftime("%d-%m-%Y-%H-%M"), ) os.makedirs(experiment_path, exist_ok=True) print("---------- planning ----------") for i in range(repeat): # initialize planning with 2 same views random_initial_view = [] for _ in range(2): random_initial_view.append( uniform_sampling(radius=2, phi_min=0.15) ) # hard-coded, should be the same for config file for planner_type in planner_type_list: # find planner configuration file print( f"---------- {planner_type} planner, experiment ID {i} ----------\n" ) planner_cfg_path = os.path.join( "planning/config", f"{planner_type}_planner.yaml" ) assert os.path.exists(planner_cfg_path) with open(planner_cfg_path, "r") as config_file: planner_cfg = yaml.safe_load(config_file) planner_cfg.update(args.__dict__) planner_cfg["planner_type"] = planner_type planner_cfg["experiment_path"] = experiment_path planner_cfg["experiment_id"] = i nbv_planner = get_planner(planner_cfg) nbv_planner.start(initial_view=random_initial_view) print("---------- evaluation ----------") gpu_id = list(map(int, args.gpu_id.split())) device = util.get_cuda(gpu_id[0]) log_path = os.path.join(root_dir, "neural_rendering", "logs", args.model_name) assert os.path.exists(log_path), "experiment does not exist" with open(f"{log_path}/training_setup.yaml", "r") as config_file: cfg = yaml.safe_load(config_file) checkpoint_path = os.path.join(log_path, "checkpoints", "best.ckpt") assert os.path.exists(checkpoint_path), "checkpoint does not exist" ckpt_file = torch.load(checkpoint_path) model = PretrainedModel(cfg["model"], ckpt_file, device, gpu_id) # load test view data as ground truth test_rgbs, test_poses, focal, c = get_image_data( args.test_data_path, "normal", device ) # configure rendering information nview = int(args.nviews) _, _, H, W = test_rgbs.shape z_near = cfg["data"]["dataset"]["z_near"] z_far = cfg["data"]["dataset"]["z_far"] step_list = [2, 4, 6, 8, 10, 12, 14, 16, 18, 20] total_df = pandas.DataFrame( { "Planning Type": [], "Reference Image Num.": [], "PSNR": [], "SSIM": [], } ) for r in range(repeat): for planner_type in planner_type_list: ref_data_path = os.path.join(experiment_path, planner_type, str(r)) ref_rgbs, ref_poses, _, _ = get_image_data(ref_data_path, "normal", device) for step in step_list: print( f"---------- planner:{planner_type}, repeat {r}, step {step} ----------\n" ) ref_kd_tree = spatial.KDTree(ref_poses[:step, :3, 3].cpu().numpy()) psnr_list = [] ssim_list = [] with torch.no_grad(): for i, rgb in enumerate(test_rgbs): pose = test_poses[i] gt = rgb * 0.5 + 0.5 gt = gt.permute(1, 2, 0).cpu().numpy() _, ref_index = ref_kd_tree.query( pose[:3, 3].cpu().numpy(), np.minimum(nview, step) ) model.network.encode( ref_rgbs[ref_index].unsqueeze(0), ref_poses[ref_index].unsqueeze(0), focal.unsqueeze(0), c.unsqueeze(0), ) target_rays = util.gen_rays( pose.unsqueeze(0), W, H, focal, z_near, z_far, c ) target_rays = target_rays.reshape(1, H * W, -1) predict = DotMap(model.renderer_par(target_rays)) metrics_dict = util.calc_metrics(predict, torch.tensor(gt)) psnr_list.append(metrics_dict["psnr"]) ssim_list.append(metrics_dict["ssim"]) psnr_mean = np.mean(psnr_list) ssim_mean = np.mean(ssim_list) print("psnr:", psnr_mean, "ssim:", ssim_mean) dataframe = pandas.DataFrame( { "Planning Type": planner_title[planner_type], "Reference Image Num.": step, "PSNR": psnr_mean, "SSIM": ssim_mean, }, index=[r], ) total_df = total_df.append(dataframe) total_df.to_csv(f"{experiment_path}/dataframe.csv") image_to_tensor = util.get_image_to_tensor_balanced() def get_image_data(data_path, coordinate_format, device, rescale=0.5): assert os.path.exists(data_path) rgb_paths = [ x for x in glob.glob(f"{data_path}/images/*") if (x.endswith(".jpg") or x.endswith(".png")) ] rgb_paths = sorted(rgb_paths) images = [] poses = [] for image_path in rgb_paths: image = imageio.imread(image_path)[..., :3] image = image_to_tensor(image) images.append(image) pose_list = np.load(f"{data_path}/trajectory.npy") for pose in pose_list: pose = util.coordinate_transformation(pose, format=coordinate_format) poses.append(pose) with open(f"{data_path}/camera_info.yaml") as file: intrinsic = yaml.safe_load(file) images = torch.stack(images).to(device) poses = torch.stack(poses).to(device) if rescale != 1: _, _, H, W = images.shape H = int(rescale * H) W = int(rescale * W) images = F.interpolate(images, size=[W, H], mode="area") focal = rescale * torch.tensor(intrinsic["focal"], dtype=torch.float32).to(device) c = rescale * torch.tensor(intrinsic["c"], dtype=torch.float32).to(device) assert len(images) == len(poses) return images, poses, focal, c def test_visualize(results_dict): import matplotlib.pyplot as plt H = 400 W = 400 rgb = results_dict.rgb[0].cpu().numpy().reshape(H, W, 3) depth = results_dict.depth[0].cpu().numpy().reshape(H, W) uncertainty = results_dict.uncertainty[0].cpu().numpy().reshape(H, W) fig, axs = plt.subplots(1, 3) axs[0].imshow(rgb) axs[1].imshow(uncertainty) axs[2].imshow(depth) plt.show() def parse_args(): parser = argparse.ArgumentParser() parser.add_argument( "--model_name", "-M", type=str, required=True, help="model name of pretrained model", ) parser.add_argument( "--test_data_path", "-TD", type=str, required=True, help="data path", ) # mandatory arguments parser.add_argument( "--repeat", "-rp", type=int, default=10, help="repeat experiment", ) # arguments with default values parser.add_argument( "--nviews", "-nv", type=int, default=5, help="number of reference views" ) parser.add_argument( "--planning_budget", "-BG", type=int, default=20, help="maximal measurments for the mission", ) parser.add_argument( "--device", type=str, default="cuda", help="config file path", ) parser.add_argument( "--gpu_id", type=str, default="0", help="gpu to use, space delimited", ) parser.add_argument( "--evaluation_only", action="store_true", help="evaluation mode" ) parser.add_argument( "--experiment_path", type=str, default="not defined", help="must be defined in evaluation mode", ) args = parser.parse_args() return args if __name__ == "__main__": main()
9,450
29.685065
94
py
neu-nbv
neu-nbv-main/scripts/planning/planner/neural_nbv/neural_nbv_planner.py
import numpy as np from scipy.spatial.transform import Rotation as R from planner.planner import Planner from planner.utils import view_to_pose_batch, random_view, uniform_sampling from neural_rendering.evaluation.pretrained_model import PretrainedModel import torch from dotmap import DotMap from neural_rendering.utils import util import torch.nn.functional as F import scipy.spatial as spatial import matplotlib.pyplot as plt import time import yaml import os class NeuralNBVPlanner(Planner): def __init__(self, cfg): super().__init__(cfg) self.device = cfg["device"] self.gpu_id = list(map(int, cfg["gpu_id"].split())) self.init_sensor_model(cfg) self.image_to_tensor = util.get_image_to_tensor_balanced() self.num_candidates = cfg["num_candidates"] self.sample_type = cfg["sample_type"] self.view_change = cfg["view_change"] self.local_view_change = cfg["local_view_change"] self.selection_range = cfg["selection_range"] self.hierachical_sampling = cfg["use_hierachical_sampling"] self.sample_ratio = cfg["sample_ratio"] self.K = cfg["top_k"] self.max_ref_num = cfg["maximal_ref"] self.reward_type = cfg["reward_type"] self.render_batch_size = cfg["render_batch_size"] self.uncertainty_th = cfg["uncertainty_threshold"] self.candidate_views = None self.candidate_poses = None self.render_pairs = None self.trajectory_kdtree = None # self.depth_for_renderer = torch.empty( # (self.planning_budget, self.H, self.W) # ).to(self.device) def init_sensor_model(self, cfg): assert os.path.exists(cfg["config_path"]) assert os.path.exists(cfg["checkpoint_path"]) with open(cfg["config_path"], "r") as config_file: model_cfg = yaml.safe_load(config_file)["model"] ckpt_file = torch.load(cfg["checkpoint_path"]) self.model = PretrainedModel(model_cfg, ckpt_file, self.device, self.gpu_id) # original image format H, W = self.camera_info["image_resolution"] # (H, W) focal = self.camera_info["focal"] # (f_x, f_y) c = self.camera_info["c"] # (c_x, c_y) # desired image format for redendering input render_info = cfg["render_info"] H_ref, W_ref = render_info["ref_image_resolution"] ref_focal = [0, 0] ref_c = [0, 0] if np.any([H, W] != [H_ref, W_ref]): scale_h = H_ref / H scale_w = W_ref / W ref_focal[0] = scale_w * focal[0] ref_focal[1] = scale_h * focal[1] ref_c[0] = scale_w * c[0] ref_c[1] = scale_h * c[1] self.ref_focal = torch.tensor(ref_focal, dtype=torch.float32).to(self.device) self.ref_c = torch.tensor(ref_c, dtype=torch.float32).to(self.device) self.ref_image_resolution = (H_ref, W_ref) self.trajectory_for_renderer = torch.empty((self.planning_budget, 4, 4)).to( self.device ) self.rgb_for_renderer = torch.empty((self.planning_budget, 3, H_ref, W_ref)).to( self.device ) # desired image format for redendering output render_scale = render_info["render_scale"] self.H_render = int(render_scale * H_ref) self.W_render = int(render_scale * W_ref) render_scale = torch.tensor( [ self.W_render / W_ref, self.H_render / H_ref, ] ).to(self.device) self.render_focal = render_scale * self.ref_focal self.render_c = render_scale * self.ref_c self.z_near, self.z_far = render_info["scene_range"] def render_novel_views(self, candidate_poses): candidate_num = len(candidate_poses) reward_list = np.zeros(candidate_num) distance_all, ref_index_all = self.trajectory_kdtree.query( candidate_poses[:, :3, 3], np.minimum(self.max_ref_num, self.step) ) # distance_all = torch.tensor(distance_all) # ref_index_all = torch.tensor(ref_index_all) bool_mask = ~np.isinf(distance_all) novel_poses = util.coordinate_transformation( candidate_poses, format="normal" ).to(self.device) # render novel view in batch split_novel_view = torch.split( torch.arange(candidate_num), self.render_batch_size, dim=0 ) for i in split_novel_view: ref_index = torch.tensor(ref_index_all[i] * bool_mask[i]) ref_images = self.rgb_for_renderer[ref_index] ref_poses = self.trajectory_for_renderer[ref_index] render_results = self.rendering(ref_images, ref_poses, novel_poses[i]) reward_list[i] = self.cal_reward(render_results) return reward_list def rendering(self, ref_images, ref_poses, novel_poses): NP = len(novel_poses) with torch.no_grad(): self.model.network.encode( ref_images, ref_poses, self.ref_focal.unsqueeze(0), self.ref_c.unsqueeze(0), ) target_rays = util.gen_rays( novel_poses, self.W_render, self.H_render, self.render_focal, self.z_near, self.z_far, self.render_c, ) # (IN, H, W, 8) target_rays = target_rays.reshape(NP, self.H_render * self.W_render, -1) predict = DotMap(self.model.renderer_par(target_rays)) return predict def cal_reward(self, render_results): uncertainty = render_results["uncertainty"] reward = torch.mean(uncertainty**2, dim=-1).cpu().numpy() reward = np.log10(reward) return reward # one stage planning def start_planning(self): candidate_views, candidate_poses = self.local_sampling( self.num_candidates, self.current_pose[:3, 3], view_change=self.view_change ) reward_list = self.render_novel_views(candidate_poses) nbv_index = np.argmax(reward_list) return candidate_views[nbv_index] def global_sampling(self, num): view_list = np.empty((num, 2)) for i in range(num): view_list[i] = uniform_sampling(self.radius, self.phi_min) pose_list = view_to_pose_batch(view_list, self.radius) return view_list, pose_list def local_sampling(self, num, xyz, view_change, min_view_change=0.2): view_list = np.empty((num, 2)) for i in range(num): view_list[i] = random_view( xyz, self.radius, self.phi_min, min_view_change, view_change ) pose_list = view_to_pose_batch(view_list, self.radius) return view_list, pose_list def plan_next_view(self): import time if self.step > 1: t1 = time.time() nbv = self.start_planning() t2 = time.time() print((t2 - t1)) return nbv # need at least two views to start the planning else: random_next_view = random_view( self.current_pose[:3, 3], self.radius, self.phi_min, self.view_change - 0.1, self.view_change, ) return random_next_view def record_trajectory(self, view, pose): self.view_trajectory[self.step] = view self.trajectory[self.step] = pose # maintain current measurment positions in kd tree self.trajectory_kdtree = spatial.KDTree(self.trajectory[: self.step + 1, :3, 3]) self.trajectory_for_renderer[self.step] = util.coordinate_transformation( pose, format="normal" ).to(self.device) def record_rgb_measurement(self, rgb): rgb = np.clip(rgb, a_min=0, a_max=255) rgb = rgb / 255 self.rgb_measurements[self.step] = rgb ref_image = self.image_to_tensor(rgb).to(self.device) ref_image = F.interpolate( ref_image.unsqueeze(0), size=self.ref_image_resolution, mode="area" ).squeeze(0) self.rgb_for_renderer[self.step] = ref_image def test_visualize(self, ref_images, results_dict): import matplotlib.pyplot as plt H = 60 W = 60 for i in range(self.render_batch_size): rgb = results_dict.rgb[i].cpu().numpy().reshape(H, W, 3) depth = results_dict.depth[i].cpu().numpy().reshape(H, W) uncertainty = results_dict.uncertainty[i].cpu().numpy().reshape(H, W) fig, axs = plt.subplots(1, 3) axs[0].imshow(rgb) axs[1].imshow(uncertainty) axs[2].imshow(depth) plt.show()
8,864
33.901575
88
py
fitclip
fitclip-main/util/structured_group_utils.py
"""Useful utils when using `DataModuleStructuredGroup`.""" from typing import Any, Mapping, Sequence, Tuple import torch from aligner.video_text_module import TYPE_INPUT from util.tensor_utils import pad TYPE_MULTI_INPUT = Mapping[str, TYPE_INPUT] # It's like `default_collate` but instead of a sequence we have a mapping, and we do `cat` instead of `stack`. # It makes sense to be similar because we're merging multiple batches together. # Note that using collate from the dataloader side. It's simpler, and more GPU-memory efficient. def _cat_collate(batch: Sequence[Any]) -> Any: elem = batch[0] elem_type = type(batch) if isinstance(elem, torch.Tensor): return torch.cat(batch) # noqa elif isinstance(elem, Mapping): return {k: _cat_collate([d[k] for d in batch]) for k in elem} elif isinstance(elem, (float, int, bytes, str)): return batch elif isinstance(elem, tuple) and hasattr(elem, '_fields'): # namedtuple return elem_type(*(_cat_collate(samples) for samples in zip(*batch))) # noqa elif isinstance(elem, Sequence): return [x for d in batch for x in d] else: raise TypeError(f"Not sure how to collate type {elem_type}") def _merge_datasets_batch(batches_by_dataset: TYPE_MULTI_INPUT) -> Tuple[TYPE_INPUT, Sequence[int]]: lengths = [len(batch["video"]) for batch in batches_by_dataset.values()] max_text_len = max(batch["text"]["input_ids"].shape[-1] for batch in batches_by_dataset.values()) for batch in batches_by_dataset.values(): batch["text"] = {k: pad(v, min_size=max_text_len, dim=-1) for k, v in batch["text"].items()} batch = _cat_collate(list(batches_by_dataset.values())) return batch, lengths
1,737
40.380952
110
py
fitclip
fitclip-main/util/viz_utils.py
import numpy as np import torch import torchvision from matplotlib import pyplot as plt from matplotlib.pyplot import subplots_adjust from torchvision.transforms.functional import to_pil_image from aligner.encoder.video_text_encoder import VideoTextEncoder def visualize_images_tensor(images: torch.Tensor) -> plt.Axes: """`images` has shape (N, C, H, W).""" grid = torchvision.utils.make_grid(images) fig, ax = plt.subplots() fig.tight_layout() subplots_adjust(left=None, bottom=None, right=None, top=None, wspace=None, hspace=None) ax.autoscale_view("tight") ax.imshow(np.asarray(to_pil_image(grid))) ax.set(xticklabels=[], yticklabels=[], xticks=[], yticks=[]) return ax def debug_batch(video: torch.Tensor, text: torch.Tensor, encoder: VideoTextEncoder) -> None: video, text = video.detach().cpu(), text.detach().cpu() video = encoder.to_bchw(video) denormalized_images = encoder.denormalize_video_tensor(video).reshape(-1, *video.shape[2:]) visualize_images_tensor(denormalized_images) plt.show() for decoded in encoder.decode_text(text): print(decoded)
1,139
29
95
py
fitclip
fitclip-main/util/tensor_utils.py
from typing import Any, Mapping, Optional, Sequence, TypeVar, Union import pytorch_lightning as pl import torch import torch.nn.functional as F from pytorch_lightning.utilities.apply_func import apply_to_collection T = TypeVar("T") def pad(t: torch.Tensor, min_size: int, dim: int = 1, value: Any = 0) -> torch.Tensor: """Pads the dim `dim` in `t` with the value `value` so the size is at least `min_size`.""" if dim < 0: dim += len(t.shape) if (count := t.shape[dim]) < min_size: # `pad` keyword arg goes from the last dim to the first one in pairs, where the first value of the pair is # for left padding and the other one for right padding. return F.pad(t, pad=(0, 0) * (len(t.shape) - 1 - dim) + (0, min_size - count), value=value) else: return t def split_in_collection(data: T, split_size_or_sections: Union[int, Sequence[int]]) -> Sequence[T]: """Applies `split` to the inside tensors of the collections and also generates one collection for each of the returned elements from `split`.""" type_ = type(data) if isinstance(data, torch.Tensor): return data.split(split_size_or_sections) elif isinstance(data, Mapping): zipped = zip(*(split_in_collection(v, split_size_or_sections) for v in data.values())) return [type_((k, v) for k, v in zip(data.keys(), z)) for z in zipped] elif isinstance(data, Sequence): return [type_(z) for z in zip(*(split_in_collection(e, split_size_or_sections) for e in data))] else: raise ValueError(f"Unsupported type for split: {type_}") def _first_tensor_in_collection(data: Any) -> torch.Tensor: if isinstance(data, torch.Tensor): return data elif isinstance(data, Mapping): return _first_tensor_in_collection(data.values()) else: return _first_tensor_in_collection(next(iter(data))) def all_gather(lightning_module: pl.LightningModule, data: Any, group: Optional[Any] = None, sync_grads: bool = False, return_world_size_dim: bool = False) -> Any: """Gathers a tensor, or multiple tensors inside a collection, so that the output number of dimensions is the same regardless of the accelerator. Note this is different from `pl.LightningModule.all_gather`, that for a single GPU it doesn't return a new dimension but for the parallel settings it does. """ first_tensor_old_shape = _first_tensor_in_collection(data).shape output = lightning_module.all_gather(data, group=group, sync_grads=sync_grads) if len(first_tensor_new_shape := _first_tensor_in_collection(output).shape) == len(first_tensor_old_shape) + 1: return output if return_world_size_dim else apply_to_collection(output, torch.Tensor, lambda t: t.view(-1, *t.shape[2:])) elif len(first_tensor_new_shape) == len(first_tensor_old_shape): return apply_to_collection(output, torch.Tensor, torch.Tensor.unsqueeze, 0) if return_world_size_dim else output else: raise ValueError(f"Unexpected new shape for the first tensor in the collection: {first_tensor_new_shape} (old " f"was {first_tensor_old_shape}). " f"The new shape was expected to have the same number of dimensions or one more.")
3,355
49.089552
120
py
fitclip
fitclip-main/util/checkpoint_utils.py
from typing import MutableMapping import torch from cached_path import cached_path from util.typing_utils import TYPE_PATH def state_dict_from_checkpoint_path(checkpoint_path: TYPE_PATH, prefix: str = "") -> MutableMapping[str, torch.Tensor]: prefix += ("" if prefix.endswith(".") or not prefix else ".") checkpoint = torch.load(cached_path(checkpoint_path)) return {k[len(prefix):]: v for k, v in checkpoint["state_dict"].items() if k.startswith(prefix)}
472
35.384615
119
py
fitclip
fitclip-main/util/video_utils.py
import os from typing import Any, Callable, Iterable, Iterator, Optional, Sequence from torchvision.datasets.video_utils import VideoClips from util.typing_utils import TYPE_PATH # From https://en.wikipedia.org/wiki/Video_file_format VIDEO_FILE_EXTENSIONS = (".3g2", ".3gp", ".amv", ".asf", ".avi", ".drc", ".f4a", ".f4b", ".f4p", ".f4v", ".flv", ".gif", ".gifv", ".m2ts", ".m2v", ".m4p", ".m4v", ".mkv", ".mng", ".mov", ".mp2", ".mp4", ".mpe", ".mpeg", ".mpg", ".mpv", ".mts", ".mxf", ".nsv", ".ogg", ".ogv", ".qt", ".rm", ".rmvb", ".roq", ".svi", ".ts", ".viv", ".vob", ".webm", ".wmv", ".yuv") def get_videos_in_folder(path: TYPE_PATH, extensions: Optional[Iterable[str]] = VIDEO_FILE_EXTENSIONS) -> Iterator[str]: extensions = None if extensions is None else tuple(extensions) for folder, _, filenames in os.walk(path, followlinks=True): for filename in filenames: if os.path.isfile(full_path := os.path.join(folder, filename)) \ and (not extensions or filename.lower().endswith(extensions)): yield full_path def get_sorted_videos_in_folder(path: TYPE_PATH, extensions: Optional[Iterable[str]] = VIDEO_FILE_EXTENSIONS, key: Optional[Callable[[str], Any]] = None, reverse: bool = False) -> Iterator[str]: """Returns a sorted version of `get_videos_in_folder`. Even though this can be simply applied by the caller, the fact that the main use case of `get_videos_in_folder` is from a video dataset and that its order should be deterministic (but that `get_videos_in_folder` doesn't guarantee it) makes this function handy and a wake-up call for this issue. The videos in a PyTorch `Dataset` need to be deterministic e.g. for a distributed setting, when e.g. using `DistributedSampler` for it to guarantee each data sample is used once and only once between all processes. """ return sorted(get_videos_in_folder(path, extensions), key=key, reverse=reverse) def resample(num_frames: int, original_fps: float, new_fps: float) -> Sequence[int]: """Returns essentially the same as `VideoClips._resample_video_idx`. Unlike it, it always checks for the max frames (the mentioned function doesn't do it when it returns a `slice`).""" indices = VideoClips._resample_video_idx(num_frames, original_fps, new_fps) if isinstance(indices, slice) and indices.stop is None: indices = range(*indices.indices((indices.start or 0) + num_frames * indices.step)) return indices
2,659
53.285714
119
py
fitclip
fitclip-main/scripts/apply_wise_ft.py
#!/usr/bin/env python import argparse import torch from aligner.encoder.clip_video_text_encoder import load_clip_model from aligner.wise import wise_state_dict from util.argparse_with_defaults import ArgumentParserWithDefaults def parse_args() -> argparse.Namespace: parser = ArgumentParserWithDefaults("Applies weight-space ensembles for fine-tuning (WiSE-FT) on 2 CLIP " "checkpoints.", description="See https://arxiv.org/abs/2109.01903 for more info.") parser.add_argument("input_path_or_name1", metavar="INPUT_FILE_OR_NAME_1") parser.add_argument("input_path_or_name2", metavar="INPUT_FILE_OR_NAME_2") parser.add_argument("output_path", metavar="OUTPUT_FILE") parser.add_argument("--weight-for-2", type=float, default=0.5) return parser.parse_args() def main() -> None: args = parse_args() model1 = load_clip_model(args.input_path_or_name1) model2 = load_clip_model(args.input_path_or_name2) # We don't use the logic scale from CLIP but ours, so we had deleted it. Here we need to re-create the variable, # so it doesn't fail when using the checkpoints. model1.logit_scale = getattr(model1, "logit_scale", torch.tensor(float("nan"))) model2.logit_scale = getattr(model2, "logit_scale", torch.tensor(float("nan"))) state_dict = wise_state_dict(model1, model2, weight_for_2=args.weight_for_2) torch.save(state_dict, args.output_path) if __name__ == "__main__": main()
1,526
37.175
116
py
fitclip
fitclip-main/scripts/subcorr.py
#!/usr/bin/env python import argparse import sys from typing import Any, Callable, Iterable, MutableMapping, Optional, Sequence, Union import PIL.Image import clip import decord import numpy as np import seaborn as sns import torch from clip.model import CLIP from matplotlib import pyplot as plt from matplotlib.offsetbox import AnnotationBbox, OffsetImage from spacy.tokens import Doc, Span def get_video_info(path: str) -> MutableMapping[str, Any]: video_reader = decord.VideoReader(path) frame_indices = list(range(0, len(video_reader), 10)) frames = [PIL.Image.fromarray(f) for f in video_reader.get_batch(frame_indices).asnumpy()] thumbnails_frame_indices = video_reader.get_key_indices() thumbnails = [PIL.Image.fromarray(f) for f in video_reader.get_batch(thumbnails_frame_indices).asnumpy()] thumbnails = [f.copy() for f in thumbnails] for thumbnail in thumbnails: thumbnail.thumbnail((64, 64)) return { "frames": frames, "frame_times": video_reader.get_frame_timestamp(frame_indices).mean(axis=-1), # noqa "thumbnails": thumbnails, "thumbnail_times": video_reader.get_frame_timestamp(thumbnails_frame_indices).mean(axis=-1), # noqa } def encode_visual(images: Iterable[PIL.Image.Image], clip_model: CLIP, image_preprocessor: Callable[[PIL.Image.Image], torch.Tensor], device: Optional[Any] = None) -> torch.Tensor: images = torch.stack([image_preprocessor(image) for image in images]) if device is not None: images = images.to(device) with torch.inference_mode(): encoded_images = clip_model.encode_image(images) return encoded_images / encoded_images.norm(dim=-1, keepdim=True) def encode_text(text: str, clip_model: CLIP, device: Optional[Any] = None) -> torch.Tensor: tokenized_texts = clip.tokenize([text]) if device is not None: tokenized_texts = tokenized_texts.to(device) with torch.inference_mode(): encoded_texts = clip_model.encode_text(tokenized_texts) return encoded_texts / encoded_texts.norm(dim=-1, keepdim=True) def text_probs(encoded_images: torch.Tensor, encoded_texts: torch.Tensor) -> np.ndarray: with torch.inference_mode(): # clip_model.logit_scale.exp() == 100 return (100 * encoded_images @ encoded_texts.T).softmax(dim=0).squeeze(-1).cpu().numpy() # noqa def create_figure(times: Sequence[float], probs: Sequence[float], thumbnail_times: Sequence[float], thumbnails: Iterable[PIL.Image.Image], title: Union[Doc, Span, str]) -> plt.Axes: # noinspection SpellCheckingInspection sns.set(rc={"figure.figsize": (1.0 * len(thumbnail_times), 1.5)}) ax = sns.lineplot(x=times, y=probs) plt.xticks(thumbnail_times) ax.set_title(title.text if isinstance(title, (Doc, Span)) else title, fontsize=35, y=0.6) ax.set(xlabel="time", ylabel="probability") plt.fill_between(times, probs) if isinstance(title, (Doc, Span)): start_time = title[0]._.start_time end_time = title[-1]._.end_time plt.axvspan(start_time, end_time, alpha=0.5, color="red") for i, (time, thumbnail) in enumerate(zip(thumbnail_times, thumbnails)): im = OffsetImage(thumbnail, axes=ax) ab = AnnotationBbox(im, (time, 0), xybox=(0, -60), frameon=False, boxcoords="offset points", pad=0) ax.add_artist(ab) plt.margins(x=0, tight=True) plt.subplots_adjust(left=0, bottom=0, right=1, top=1, wspace=0, hspace=0) return ax def create_figure_for_text(encoded_frames: torch.Tensor, text: Union[Doc, Span, str], clip_model: CLIP, times: Sequence[float], thumbnail_times: Sequence[float], thumbnails: Iterable[PIL.Image.Image]) -> plt.Axes: encoded_texts = encode_text(text.text if isinstance(text, (Doc, Span)) else text, clip_model, device=encoded_frames.device) probs = text_probs(encoded_frames, encoded_texts) return create_figure(times, probs, thumbnail_times, thumbnails, text) def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser() parser.add_argument("path", metavar="PATH") return parser.parse_args() def main() -> None: sns.set_theme() args = parse_args() device = "cuda" if torch.cuda.is_available() else "cpu" clip_model, image_preprocessor = clip.load("ViT-B/16", device=device) # noinspection SpellCheckingInspection video_info = get_video_info(args.path) encoded_frames = encode_visual(video_info["frames"], clip_model, image_preprocessor, device=device) for text in sys.stdin: if text := text.strip(): create_figure_for_text(encoded_frames, text, clip_model, video_info["frame_times"], video_info["thumbnail_times"], video_info["thumbnails"]) plt.show() if __name__ == "__main__": main()
4,981
35.101449
109
py
fitclip
fitclip-main/scripts/prepare_trained_clip_checkpoint_for_evaluation.py
#!/usr/bin/env python import argparse import torch from util.checkpoint_utils import state_dict_from_checkpoint_path def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser() parser.add_argument("input_path", metavar="INPUT_FILE") parser.add_argument("output_path", metavar="OUTPUT_FILE") parser.add_argument("--prefix", default="encoder.model.") return parser.parse_args() def main() -> None: args = parse_args() state_dict = state_dict_from_checkpoint_path(args.input_path, prefix=args.prefix) # We don't use the logic scale from CLIP but ours, so we had deleted it. Here we need to re-create the variable, # so it doesn't fail when loading this `state_dict`. state_dict["logit_scale"] = torch.tensor(float("nan")) torch.save(state_dict, args.output_path) if __name__ == "__main__": main()
868
27.032258
116
py
fitclip
fitclip-main/scripts/checkpoint_to_state_dict.py
#!/usr/bin/env python import argparse import sys import torch from util.checkpoint_utils import state_dict_from_checkpoint_path def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser() parser.add_argument("input_path", metavar="INPUT_FILE") parser.add_argument("--prefix", default="encoder.model.") return parser.parse_args() def main() -> None: args = parse_args() state_dict = state_dict_from_checkpoint_path(args.input_path, prefix=args.prefix) torch.save(state_dict, sys.stdout.buffer) if __name__ == "__main__": main()
582
22.32
85
py
fitclip
fitclip-main/scripts/prepare_trained_checkpoint_for_evaluation.py
#!/usr/bin/env python import argparse import torch from cached_path import cached_path def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser() parser.add_argument("input_path", metavar="INPUT_FILE", type=cached_path) parser.add_argument("output_path", metavar="OUTPUT_FILE") parser.add_argument("--prefix", default="encoder.model.") return parser.parse_args() def main() -> None: args = parse_args() checkpoint = torch.load(args.input_path) prefix = args.prefix + ("" if args.prefix.endswith(".") else ".") checkpoint["state_dict"] = {k[len(prefix):]: v for k, v in checkpoint["state_dict"].items() if k.startswith(prefix)} torch.save(checkpoint, args.output_path) if __name__ == "__main__": main()
771
26.571429
120
py
fitclip
fitclip-main/scripts/open_clip_checkpoint_to_model.py
#!/usr/bin/env python import argparse import torch from cached_path import cached_path def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser() parser.add_argument("input_path", metavar="INPUT_FILE", type=cached_path) parser.add_argument("output_path", metavar="OUTPUT_FILE") return parser.parse_args() def main() -> None: args = parse_args() checkpoint = torch.load(args.input_path) state_dict = checkpoint["state_dict"] first_key = next(iter(state_dict)) prefix = next(prefix for prefix in ["model", "module"] if first_key.startswith(prefix + ".")) torch.save({k[len(prefix + "."):]: v for k, v in state_dict.items()}, args.output_path) if __name__ == "__main__": main()
744
25.607143
97
py
fitclip
fitclip-main/aligner/video_text_module.py
from typing import Any, Literal, Mapping, MutableMapping, Optional, Sequence, Tuple, Union import math import pytorch_lightning as pl import torch.distributed.nn from overrides import overrides from torch import nn from torch.nn.modules.loss import _Loss from aligner.encoder.video_text_encoder import TYPE_OUTPUT, VideoTextEncoder from aligner.loss import NCELoss from util.tensor_utils import all_gather TYPE_INPUT = MutableMapping[str, Any] TYPE_SPLIT = Literal["train", "val"] def log_lr(pl_module: pl.LightningModule, **kwargs) -> None: for i, optimizer in enumerate(pl_module.trainer.optimizers): for j, param_group in enumerate(optimizer.param_groups): if (lr := param_group.get("lr")) is not None: # noqa pl_module.log(f"lr_{i}_group_{j}", lr, **kwargs) class VideoTextLightningModule(pl.LightningModule): # noqa def __init__(self, encoder: VideoTextEncoder, init_temperature: float = 0.05, min_temperature: float = 0.001, fit_temperature: bool = True, loss: Optional[_Loss] = None) -> None: super().__init__() self.encoder = encoder # Use the temperature as in CLIP: save it in log-space and fit it along with the model. self.logit_scale = nn.Parameter(torch.tensor([- math.log(init_temperature)]), requires_grad=fit_temperature) # The following constant is set also as a parameter, so it's moved to the correct device automatically. self.max_logit_scale = nn.Parameter(torch.tensor([- math.log(min_temperature)]), requires_grad=False) self.loss = loss or NCELoss() @overrides(check_signature=False) def forward(self, batch: TYPE_INPUT, _batch_idx: int = 0) -> Union[TYPE_OUTPUT, Tuple[torch.Tensor, torch.Tensor, Sequence[str]]]: batch.pop("video_id", None) return self.encoder(**batch) def _step(self, batch: TYPE_INPUT, batch_idx: int = 0) -> TYPE_OUTPUT: return self(batch, batch_idx) @overrides(check_signature=False) def training_step(self, batch: TYPE_INPUT, _batch_idx: int = 0) -> TYPE_OUTPUT: output = self._step(batch, _batch_idx) # Need to log the step because PL doesn't log it in Neptune. # See https://github.com/PyTorchLightning/pytorch-lightning/pull/5510 first_video_value = next(v for k, v in batch.items() if k.startswith("video")) self.log("step", float(self.global_step), batch_size=len(first_video_value)) return output def _step_end(self, output: TYPE_OUTPUT, split: TYPE_SPLIT, log_kwargs: Optional[Mapping[str, Any]] = None) -> Union[torch.Tensor, TYPE_OUTPUT]: log_kwargs = log_kwargs or {} encoded_video, encoded_text = all_gather(self, output, sync_grads=split == "train") batch_size = len(encoded_video) logit_scale = self.logit_scale.exp() scores = logit_scale * encoded_video @ encoded_text.T loss = self.loss(scores) # Note train loss it's already shown in the progress bar by PL by default. # # Note that we need to pass the batch size in the first step log # as it can't be easily inferred by PL in our case. self.log(f"loss/{split}", loss, prog_bar=split != "train", batch_size=batch_size, **log_kwargs) if split == "train": self.log("batch_size", float(batch_size), batch_size=batch_size) self.log("temperature", 1 / logit_scale, batch_size=batch_size) return loss if split == "train" else (encoded_video, encoded_text) @overrides(check_signature=False) def training_step_end(self, output: TYPE_OUTPUT) -> torch.Tensor: loss = self._step_end(output, split="train") log_lr(self) return loss @overrides(check_signature=False) def predict_step(self, batch: TYPE_INPUT, batch_idx: int = 0) -> Mapping[str, torch.Tensor]: encoded_video, encoded_text = self._step(batch, batch_idx) return { "encoded_videos": encoded_video, "encoded_texts": encoded_text, "video_ids": batch["video_id"] } @overrides(check_signature=False) def optimizer_step(self, *args, **kwargs) -> None: super().optimizer_step(*args, **kwargs) if self.logit_scale >= self.max_logit_scale: self.logit_scale.copy_(self.max_logit_scale)
4,376
43.663265
116
py
fitclip
fitclip-main/aligner/__main__.py
#!/usr/bin/env python import logging import os from time import strftime from typing import Mapping, Optional import hydra import torch from omegaconf import DictConfig from pytorch_lightning.loggers import NeptuneLogger, TensorBoardLogger from aligner.cli import create_model_data_module_trainer_and_ckpt_path, init_cli from aligner.logger_utils import get_logger_by_type # Note it's better to have this as a module, so it's importable and DDP works fine in debug mode. # Maybe this issue is caused by Hydra moving the CWD to somewhere else. LOGGER = logging.getLogger(__name__) # Set an env var, if empty, to the desired working directory in sweep mode. Then we read it from the config. # This way we make sure all processes use the same folder. # See https://github.com/PyTorchLightning/pytorch-lightning/issues/2727 os.environ.setdefault("SWEEP_DIR", f"multirun/{strftime('%Y-%m-%d')}/{strftime('%H-%M-%S')}") @hydra.main(config_path="../config", config_name="trainer") def main(cfg: DictConfig) -> Optional[float]: init_cli(cfg) if cfg.get("trainer", {}).get("strategy") == "dp": LOGGER.warning("DP strategy not supported by the current metric logging scheme." " See https://torchmetrics.readthedocs.io/en/stable/pages/lightning.html#logging-torchmetrics") model, data_module, trainer, ckpt_path = create_model_data_module_trainer_and_ckpt_path(cfg) output = None if cfg.command == "train": if cfg.get("validate_before_training"): LOGGER.info("Validation before training started.") with torch.inference_mode(): metrics_list = trainer.validate(model, datamodule=data_module, ckpt_path=ckpt_path) LOGGER.info("Validation before training finished.") if (tb_logger := get_logger_by_type(trainer, TensorBoardLogger)) and not tb_logger._default_hp_metric: tb_logger.log_hyperparams(model.hparams_initial, metrics={k: v for metrics in metrics_list for k, v in metrics.items()}) LOGGER.info("Training started.") trainer.fit(model, datamodule=data_module, ckpt_path=ckpt_path) if optimized_metric_name := cfg.get("optimized_metric_name"): output = trainer.callback_metrics.get(optimized_metric_name) elif cfg.command == "tune": assert ckpt_path is None, "Checkpoint path not supported when tuning." if trainer._accelerator_connector.is_distributed: LOGGER.warning("Tuning with the PL Trainer is known to have some issues in distributed settings." " See e.g. https://github.com/PyTorchLightning/pytorch-lightning/issues/4280") LOGGER.info("Tuning started.") trainer.tune(model, datamodule=data_module) elif cfg.command in {"evaluate", "validate"}: with torch.inference_mode(): trainer.validate(model, datamodule=data_module, ckpt_path=ckpt_path) elif cfg.command == "test": with torch.inference_mode(): trainer.test(model, datamodule=data_module, ckpt_path=ckpt_path) elif cfg.command == "predict": if trainer._accelerator_connector.is_distributed: LOGGER.warning("Predicting with the PL Trainer is known to have some issues in distributed settings." " See e.g. https://github.com/PyTorchLightning/pytorch-lightning/issues/10618") output_path = cfg.get("output_path", "predictions.pt") with torch.inference_mode(): predictions = trainer.predict(model, datamodule=data_module, ckpt_path=ckpt_path) assert predictions first_prediction = predictions[0] assert isinstance(first_prediction, Mapping) keys = first_prediction predictions_map = {k: torch.cat([prediction[k] for prediction in predictions]) if isinstance(first_prediction[k], torch.Tensor) else [p for prediction in predictions for p in prediction[k]] for k in keys} torch.save(predictions_map, output_path) else: raise ValueError(f"Unrecognized command: {cfg.command}") if (neptune_logger := get_logger_by_type(trainer, NeptuneLogger)) and trainer.is_global_zero: # In a Hydra multirun (sweep) scenario, Neptune experiments from finished runs are marked as still running # unless we stop them manually. See https://github.com/PyTorchLightning/pytorch-lightning/issues/11368 neptune_logger.run.stop() # Return the optimized metric value for hparam search. return output if __name__ == "__main__": main()
4,715
43.490566
118
py
fitclip
fitclip-main/aligner/logger_utils.py
from typing import Optional, Type, TypeVar import pytorch_lightning as pl from pytorch_lightning.loggers import LightningLoggerBase, LoggerCollection T = TypeVar("T", bound=LightningLoggerBase) def get_logger_by_type(trainer: pl.Trainer, logger_class: Type[T]) -> Optional[T]: if isinstance(trainer.logger, LoggerCollection): return next((logger for logger in trainer.logger._logger_iterable if isinstance(logger, logger_class)), None) elif isinstance(trainer.logger, logger_class): return trainer.logger else: return None
564
32.235294
117
py
fitclip
fitclip-main/aligner/teacher_student.py
import itertools from typing import Iterable, Mapping, MutableMapping, Optional, Tuple, Union import torch.distributed.nn from overrides import overrides from torch import nn from aligner.encoder import video_text_encoder from aligner.encoder.video_text_encoder import TYPE_TOKENIZER, VideoTextEncoder from aligner.loss import TeacherStudentNCELoss from aligner.text_video_retrieval import TextVideoRetrievalLightningModule from aligner.video_text_module import TYPE_INPUT, TYPE_SPLIT, log_lr from util.tensor_utils import all_gather, pad, split_in_collection TYPE_OUTPUT = Tuple[video_text_encoder.TYPE_OUTPUT, video_text_encoder.TYPE_OUTPUT] TYPE_MULTI_OUTPUT = Mapping[str, TYPE_OUTPUT] def _replace_in_tokenized_text(tokenized_text: MutableMapping[str, torch.Tensor], new_tokenized_text: Mapping[str, torch.Tensor], start_idx: int, end_idx: int, tokenizer: TYPE_TOKENIZER) -> None: """Replaces the content in the tensor `tokenized_text` from the index `start_idx` to `end_idx` (exclusive) for `new_tokenized_text`. When it needs to know details about the tokenization, it uses `tokenizer`. """ for k in tokenized_text: padding_value = 0 if "mask" in k else getattr(tokenizer, "pad_token_id", 0) # We suppose right padding. if tokenized_text[k].shape[1] > new_tokenized_text[k].shape[1]: padded = pad(new_tokenized_text[k], min_size=tokenized_text[k].shape[1], value=padding_value) tokenized_text[k] = torch.cat((tokenized_text[k][:start_idx], padded, tokenized_text[k][end_idx:])) elif tokenized_text[k].shape[1] < new_tokenized_text[k].shape[1]: padded = pad(tokenized_text[k], min_size=new_tokenized_text[k].shape[1], value=padding_value) tokenized_text[k] = torch.cat((padded[:start_idx], new_tokenized_text[k], padded[end_idx:])) else: tokenized_text[k] = torch.cat((tokenized_text[k][:start_idx], new_tokenized_text[k], tokenized_text[k][end_idx:])) class TeacherStudentLightningModule(TextVideoRetrievalLightningModule): # noqa """ Distillation training module. If specified, `prompts` is used with the unlabeled dataset videos instead of the labels it provides (if any). """ def __init__(self, encoder: VideoTextEncoder, teacher: VideoTextEncoder, labeled_dataset_name: str = "labeled", labeled_dataset_loss_share: Optional[float] = None, dataset_names: Iterable[str] = ("labeled", "unlabeled"), prompts: Optional[Iterable[str]] = None, **kwargs) -> None: super().__init__(encoder=encoder, dataset_names=dataset_names, **kwargs) self.teacher = teacher assert self.dataset_names, "This module uses dataset names." assert len(self.dataset_names) == 2, "The current implementation needs exactly 2 datasets." # FIXME: it doesn't work with different datasets for training and evaluation, because it needs certain names # for training; and this logic assumes the same dataset names for both. if labeled_dataset_loss_share is None: self.dataset_loss_share = {name: 1 / len(self.dataset_names) for name in self.dataset_names} else: self.dataset_loss_share = {labeled_dataset_name: labeled_dataset_loss_share} self.dataset_loss_share.update((name, (1 - labeled_dataset_loss_share) / (len(self.dataset_names) - 1)) for name in self.dataset_names if name != labeled_dataset_name) self.teacher_student_logit_scale = nn.Parameter(self.logit_scale.clone(), requires_grad=self.logit_scale.requires_grad) # noinspection SpellCheckingInspection self.teacher_student_loss = TeacherStudentNCELoss(reduction="batchmean") for p in self.teacher.parameters(): p.requires_grad = False self.labeled_dataset_name = labeled_dataset_name self.unlabeled_dataset_name = next(k for k in self.dataset_names if k != labeled_dataset_name) if prompts is None: self.tokenized_prompts = None self.teacher_tokenized_prompts = None else: prompts = list(prompts) # We use parameters so the device and dtype are moved correctly along with this module. self.tokenized_prompts = nn.ParameterDict((k, nn.Parameter(v, requires_grad=False)) # noqa for k, v in encoder.get_tokenizer()(prompts).items()) self.teacher_tokenized_prompts = nn.ParameterDict((k, nn.Parameter(v, requires_grad=False)) # noqa for k, v in teacher.get_tokenizer()(prompts).items()) @overrides(check_signature=False) def _step(self, batch: TYPE_INPUT, _batch_idx: int = 0) -> TYPE_OUTPUT: # Note we pass the labeled dataset portion to the teacher, but then we don't use it. return self({"video": batch["video_student"], "text": batch["text_student"]}), \ self.teacher(video=batch["video_teacher"], text=batch["text_teacher"]) @overrides(check_signature=False) def training_step(self, batch: TYPE_INPUT, _batch_idx: int = 0) -> TYPE_MULTI_OUTPUT: keys, lengths = zip(*((key, sum(1 for _ in group)) for key, group in itertools.groupby(dataset for dataset in batch.pop("dataset")))) assert len(keys) == len(self.dataset_names), "All datasets should be present in each batch." if self.tokenized_prompts is None: unlabeled_dataset_idx = None else: unlabeled_dataset_idx = keys.index(self.unlabeled_dataset_name) start_idx_in_batch = sum(lengths[i] for i in range(unlabeled_dataset_idx)) end_idx_in_batch = start_idx_in_batch + lengths[unlabeled_dataset_idx] _replace_in_tokenized_text(tokenized_text=batch["text_student"], new_tokenized_text=self.tokenized_prompts, start_idx=start_idx_in_batch, end_idx=end_idx_in_batch, tokenizer=self.encoder.get_tokenizer()) _replace_in_tokenized_text(tokenized_text=batch["text_teacher"], new_tokenized_text=self.teacher_tokenized_prompts, start_idx=start_idx_in_batch, end_idx=end_idx_in_batch, tokenizer=self.teacher.get_tokenizer()) output = self._step(batch, _batch_idx) # Need to log the step because PL doesn't log it in Neptune. # See https://github.com/PyTorchLightning/pytorch-lightning/pull/5510 first_video_value = next(v for k, v in batch.items() if k.startswith("video")) self.log(f"step", self.global_step, batch_size=len(first_video_value)) if self.tokenized_prompts is None: split_output = split_in_collection(output, lengths) else: text_split_sections = list(lengths) text_split_sections[unlabeled_dataset_idx] = len(next(iter(self.tokenized_prompts.values()))) student_video_sections = split_in_collection(output[0][0], lengths) student_text_sections = split_in_collection(output[0][1], text_split_sections) teacher_video_sections = split_in_collection(output[1][0], lengths) teacher_text_sections = split_in_collection(output[1][1], text_split_sections) split_output = (((student_video_sections[i], student_text_sections[i]), (teacher_video_sections[i], teacher_text_sections[i])) for i in range(len(student_video_sections))) return dict(zip(keys, split_output)) def _dataset_step_end(self, output: TYPE_OUTPUT, split: TYPE_SPLIT, dataset_name: Optional[str] = None) -> Union[torch.Tensor, video_text_encoder.TYPE_OUTPUT]: gathered_output = all_gather(self, output, sync_grads=split == "train") (encoded_video, encoded_text), (teacher_encoded_video, teacher_encoded_text) = gathered_output batch_size = len(encoded_video) logit_scale = self.logit_scale.exp() scores = logit_scale * encoded_video @ encoded_text.T if dataset_name == self.labeled_dataset_name: loss = self.loss(scores) else: teacher_student_logit_scale = self.teacher_student_logit_scale.exp() teacher_scores = teacher_student_logit_scale * teacher_encoded_video @ teacher_encoded_text.T loss = self.teacher_student_loss(scores, teacher_scores) * teacher_student_logit_scale ** 2 if split == "train": # Note that we need to pass the batch size in the first step log # as it can't be easily inferred by PL in our case. self.log("batch_size", float(batch_size), batch_size=batch_size) self.log("temperature/labeled", 1 / logit_scale) self.log("temperature/unlabeled", 1 / teacher_student_logit_scale) prefix = f"loss/{split}_{dataset_name}" if dataset_name else f"loss/{split}" # Note that we need to pass the batch size in the first step log # as it can't be easily inferred by PL in our case. self.log(prefix, loss, prog_bar=split != "train", batch_size=batch_size, add_dataloader_idx=False) return loss if split == "train" else (encoded_video, encoded_text) @overrides(check_signature=False) def training_step_end(self, output: TYPE_MULTI_OUTPUT) -> torch.Tensor: loss = sum(self._dataset_step_end(batch, split="train", dataset_name=name) * self.dataset_loss_share[name] for name, batch in output.items()) self.log("loss/train", loss) # Note train loss it's already shown in the progress bar by PL by default. log_lr(self) return loss @overrides(check_signature=False) def _validation_dataset_step_end(self, output: TYPE_OUTPUT, dataset_name: Optional[str] = None) -> video_text_encoder.TYPE_OUTPUT: return self._dataset_step_end(output, split="val", dataset_name=dataset_name) @overrides(check_signature=False) def optimizer_step(self, *args, **kwargs) -> None: super().optimizer_step(*args, **kwargs) if self.teacher_student_logit_scale >= self.max_logit_scale: self.teacher_student_logit_scale.copy_(self.max_logit_scale)
10,735
54.05641
117
py
fitclip
fitclip-main/aligner/loss.py
from typing import Literal import torch from overrides import overrides from torch.nn import functional as F from torch.nn.modules.loss import _Loss TYPE_REDUCTION = Literal["none", "mean", "sum"] # noinspection SpellCheckingInspection TYPE_REDUCTION_KL_DIV = Literal["none", "batchmean", "mean", "sum"] def _rows_to_columns_nce_loss(scores: torch.Tensor, reduction: TYPE_REDUCTION = "mean") -> torch.Tensor: loss = - F.log_softmax(scores, dim=-1).diag() if reduction == "mean": return loss.mean() elif reduction == "sum": return loss.sum() else: return loss def nce_loss(scores: torch.Tensor, reduction: TYPE_REDUCTION = "mean") -> torch.Tensor: return (_rows_to_columns_nce_loss(scores, reduction=reduction) + _rows_to_columns_nce_loss(scores.T, reduction=reduction)) def _rows_to_columns_teacher_student_nce_loss(scores: torch.Tensor, teacher_scores: torch.Tensor, reduction: TYPE_REDUCTION_KL_DIV = "mean") -> torch.Tensor: logits = F.log_softmax(scores, dim=-1) teacher_probs = F.softmax(teacher_scores, dim=-1) return F.kl_div(logits, teacher_probs, reduction=reduction) def teacher_student_nce_loss(scores: torch.Tensor, teacher_scores: torch.Tensor, reduction: TYPE_REDUCTION_KL_DIV = "mean") -> torch.Tensor: return (_rows_to_columns_teacher_student_nce_loss(scores, teacher_scores, reduction=reduction) + _rows_to_columns_teacher_student_nce_loss(scores.T, teacher_scores.T, reduction=reduction)) class NCELoss(_Loss): @overrides(check_signature=False) def forward(self, scores: torch.Tensor) -> torch.Tensor: return nce_loss(scores, reduction=self.reduction) # noqa class TeacherStudentNCELoss(_Loss): @overrides(check_signature=False) def forward(self, scores: torch.Tensor, teacher_scores: torch.Tensor) -> torch.Tensor: return teacher_student_nce_loss(scores, teacher_scores, reduction=self.reduction) # noqa class SimilarityLoss(_Loss): @overrides(check_signature=False) def forward(self, scores: torch.Tensor) -> torch.Tensor: # Note we actually don't need all the scores. loss = - torch.log(torch.sigmoid(scores.diag())) if self.reduction == "mean": return loss.mean() elif self.reduction == "sum": return loss.sum() else: return loss
2,447
36.090909
105
py
fitclip
fitclip-main/aligner/text_video_retrieval.py
from collections import OrderedDict from typing import Iterable, Mapping, Optional, Sequence, Tuple, Union import torch import torch.distributed.nn from overrides import overrides from torch import nn from torchmetrics import Metric, Recall from aligner.encoder.video_text_encoder import TYPE_OUTPUT from aligner.metrics import MedianRank, Rank from aligner.video_text_module import TYPE_INPUT, VideoTextLightningModule from util.tensor_utils import all_gather class TextVideoRetrievalLightningModule(VideoTextLightningModule): # noqa def __init__(self, *args, dataset_names: Optional[Iterable[str]] = None, compute_rank: bool = False, **kwargs) -> None: super().__init__(*args, **kwargs) metrics_dict = {"r1": Recall(), "r5": Recall(top_k=5), "r10": Recall(top_k=10), "mr": MedianRank()} if compute_rank: metrics_dict["rank"] = Rank() self.dataset_names = list(dataset_names) if dataset_names else None self.multiple_datasets = self.dataset_names is not None and len(self.dataset_names) > 1 if self.multiple_datasets: assert all("_" not in name for name in self.dataset_names), \ "Underscores in dataset names are problematic because of how we get their corresponding metrics." self.metrics: Mapping[str, Metric] = nn.ModuleDict((f"{name}_{dataset_name}", metric.clone()) # noqa for dataset_name in self.dataset_names for name, metric in metrics_dict.items()) else: self.metrics: Mapping[str, Metric] = nn.ModuleDict(metrics_dict) @overrides(check_signature=False) def validation_step(self, batch: TYPE_INPUT, batch_idx: int = 0, dataloader_idx: Optional[int] = None) -> Tuple[TYPE_OUTPUT, Optional[int]]: return self._step(batch, batch_idx), dataloader_idx def _validation_dataset_step_end(self, output: TYPE_OUTPUT, dataset_name: Optional[str] = None) -> TYPE_OUTPUT: encoded_video, encoded_text = all_gather(self, output) batch_size = len(encoded_video) logit_scale = self.logit_scale.exp() scores = logit_scale * encoded_video @ encoded_text.T loss = self.loss(scores) # Note that we need to pass the batch size in the first step log # as it can't be easily inferred by PL in our case. key = "loss/val" + ("" if dataset_name is None else f"_{dataset_name}") self.log(key, loss, prog_bar=True, batch_size=batch_size, add_dataloader_idx=False) return encoded_video, encoded_text @overrides(check_signature=False) def validation_step_end(self, output: Tuple[TYPE_OUTPUT, int]) -> TYPE_OUTPUT: step_output, data_loader_idx = output assert self.multiple_datasets == (data_loader_idx is not None) dataset_name = self.dataset_names[data_loader_idx] if self.multiple_datasets else None return self._validation_dataset_step_end(step_output, dataset_name=dataset_name) def _validate_dataset(self, outputs: Sequence[TYPE_OUTPUT], dataset_name: Optional[str] = None) -> None: assert self.multiple_datasets == (dataset_name is not None) encoded_videos, encoded_texts = (torch.cat(x) for x in zip(*outputs)) batch_size = len(encoded_videos) scores = encoded_texts @ encoded_videos.T target = torch.arange(scores.shape[-1], device=scores.device) for name, metric in self.metrics.items(): if not dataset_name or name.endswith(f"_{dataset_name}"): metric(scores, target) # Note that we need to pass the batch size in the first step log # as it can't be easily inferred by PL in our case. self.log(name, metric, prog_bar=True, batch_size=batch_size, add_dataloader_idx=False) @overrides(check_signature=False) def validation_epoch_end(self, outputs: Union[Sequence[TYPE_OUTPUT], Sequence[Sequence[TYPE_OUTPUT]]]) -> None: if self.multiple_datasets: for i, (name, dataset_output) in enumerate(zip(self.dataset_names, outputs)): # Necessary to set the current data loader ID so PL knows to which one the logged metrics belong # (because it returns the metrics by data loader). self._current_dataloader_idx = i self._validate_dataset(dataset_output, dataset_name=name) # noqa self._current_dataloader_idx = None else: self._validate_dataset(outputs) if "rank" in self.metrics: self.print(self.metrics["rank"].compute().tolist()) @overrides def load_state_dict(self, state_dict: "OrderedDict[str, torch.Tensor]", strict: bool = True): # If it's exactly this class, then ignore any teacher-related thing. # We do it here, so we can control it more, and avoid bugs with a general solution. if type(self) is TextVideoRetrievalLightningModule: incompatible_keys = super().load_state_dict(state_dict, strict=False) unexpected_keys = set(incompatible_keys.unexpected_keys) for key in incompatible_keys.unexpected_keys: if key.startswith("teacher"): unexpected_keys.remove(key) # We then do as in super: if strict: error_msgs = [] if unexpected_keys: unexpected_key_str = ", ".join(f'"{k}"' for k in unexpected_keys) error_msgs.append(f"Unexpected key(s) in state_dict: {unexpected_key_str}. ") if incompatible_keys.missing_keys: missing_keys_str = ', '.join(f'"{k}"' for k in incompatible_keys.missing_keys) error_msgs.append(f"Missing key(s) in state_dict: {missing_keys_str}. ") if error_msgs: error_msgs_str = "\n\t".join(error_msgs) raise RuntimeError(f"Error(s) in loading state_dict for {self.__class__.__name__}:\n\t" f"{error_msgs_str}") return incompatible_keys else: return super().load_state_dict(state_dict, strict)
6,325
46.924242
115
py
fitclip
fitclip-main/aligner/video_text_classification.py
import logging import math from typing import Any, Iterable, Mapping, Optional, Sequence, TypeVar import torch from overrides import overrides from pytorch_lightning.callbacks import RichProgressBar from pytorch_lightning.utilities.apply_func import apply_to_collection from torch import nn from torchmetrics import Accuracy, Metric from aligner.encoder.video_text_encoder import VideoTextEncoder from aligner.metrics import MedianRank from aligner.video_text_module import VideoTextLightningModule from util import iter_utils LOGGER = logging.getLogger(__name__) T = TypeVar("T") def batch_tokenized_text(tokenized: Mapping[str, Sequence[T]], n: int) -> Iterable[Mapping[str, T]]: tokenized_dicts = {k: iter(iter_utils.batch_sequence(v, n)) for k, v in tokenized.items()} length = math.ceil(len(next(iter(tokenized.values()))) / n) for _ in range(length): yield {k: next(tokenized_dicts[k]) for k in tokenized} class VideoTextClassificationLightningModule(VideoTextLightningModule): # noqa def __init__(self, encoder: VideoTextEncoder, labels: Iterable[str], templates: Optional[Iterable[str]], return_metrics_by_class: bool = False, **kwargs) -> None: super().__init__(encoder, **kwargs) labels = list(labels) label_count = len(labels) # If different templates are provided, we used them for each label # and reset the labels to be {labels} x {templates}. if templates: templates = list(templates) self.template_count = len(templates) labels = [template.format(label) for label in labels for template in templates] else: self.template_count = 1 # We tokenize all the labels but defer the encoding until the model is in the device. tokenized_labels = encoder.get_tokenizer()(labels) device = next(encoder.parameters()).device tokenized_labels = apply_to_collection(tokenized_labels, torch.Tensor, torch.Tensor.to, device) self.tokenized_labels = nn.ParameterDict(apply_to_collection(tokenized_labels, torch.Tensor, nn.Parameter, requires_grad=False)) # We encode just one label to allocate the size correctly. encoded_text = self.encoder.encode_text({k: v[:1] for k, v in tokenized_labels.items()}) self.encoded_labels = nn.Parameter(torch.empty(label_count, encoded_text.shape[-1]), requires_grad=False) self.metrics: Mapping[str, Metric] = nn.ModuleDict({"a1": Accuracy(), "a5": Accuracy(top_k=5), "mr": MedianRank()}) if return_metrics_by_class: self.metrics_by_class = nn.ModuleDict({f"a1_{k}": Accuracy() for k in range(label_count)}) else: self.metrics_by_class = None def _on_start(self) -> None: # Note that for training they should be encoded at running time, not here. # But we aren't training any text classification model but evaluating them. # # We compute them here and not during init because here the model is already in the device. # This is especially much faster than in CPU (init) when using templates. batch_size = 32 callback = next(callback for callback in self.trainer.callbacks if isinstance(callback, RichProgressBar)) progress = callback.progress if self.trainer.is_global_zero: progress_task = progress.add_task( description="Encoding the labels", total=math.ceil(len(next(iter(self.tokenized_labels.values()))) / batch_size)) else: progress_task = None encoded_label_list = [] for tokenized_labels_batch in batch_tokenized_text(self.tokenized_labels, batch_size): encoded_label_list.append(self.encoder.encode_text(tokenized_labels_batch)) if progress_task is not None: progress.update(progress_task, advance=1) encoded_labels = torch.cat(encoded_label_list) encoded_labels = encoded_labels.reshape(-1, self.template_count, encoded_labels.shape[1]).mean(dim=1) self.encoded_labels.copy_(encoded_labels) if progress_task is not None: # If we remove it, it later fails, not sure why. So we just hide it. progress.update(progress_task, visible=False) @overrides def on_validation_start(self) -> None: self._on_start() @overrides def on_test_start(self) -> None: self._on_start() @overrides def on_predict_start(self) -> None: self._on_start() @overrides(check_signature=False) def forward(self, video: torch.Tensor) -> torch.Tensor: return self.encoder.encode_video(video) @ self.encoded_labels.T @overrides(check_signature=False) def validation_step(self, batch: Mapping[str, Any], _batch_idx: int = 0) -> None: scores = self(batch["video"]) label_id = batch["target"][1] for name, metric in self.metrics.items(): metric(scores, label_id) # Note that we need to pass the batch size in the first step log # as it can't be easily inferred by PL in our case. self.log(name, metric, prog_bar=True, batch_size=len(batch["video"])) if self.metrics_by_class is not None: for scores_instance, label_id_instance in zip(scores, label_id): metric = self.metrics_by_class[f"a1_{label_id_instance}"] metric(scores_instance.unsqueeze(0), label_id_instance.unsqueeze(0)) self.log(f"a1_{label_id_instance}", metric, batch_size=1) @overrides(check_signature=False) def predict_step(self, batch: Mapping[str, Any], _batch_idx: int = 0) -> Mapping[str, torch.Tensor]: return { "predictions": self(batch["video"]).argmax(dim=-1), "labels": batch["target"][1], "video_ids": batch["video_id"], }
6,045
41.879433
114
py
fitclip
fitclip-main/aligner/cli.py
#!/usr/bin/env python import copy import logging import warnings from types import MethodType from typing import Any, Mapping, Optional, Tuple, Type import hydra import pytorch_lightning as pl from cached_path import cached_path from omegaconf import DictConfig from pytorch_lightning import seed_everything from torch.optim import Optimizer from aligner.data.data_module_group import DataModuleStructuredGroup, EvalDataModuleGroup, MixedBatchDataModule, \ TrainAndEvalDataModules from aligner.data.video_data_module import ENCODER_OR_ENCODER_MAP, VideoClassificationDataModule from aligner.encoder.video_text_encoder import VideoTextEncoder from aligner.video_text_classification import VideoTextClassificationLightningModule from aligner.video_text_module import VideoTextLightningModule LOGGER = logging.getLogger(__name__) # This is because PL can't automatically infer the batch size, that's needed for logging. But we set it manually # within the modules. warnings.filterwarnings("ignore", message=r"^Trying to infer the `batch_size` from an ambiguous collection\. .+") # From https://stackoverflow.com/a/2020083/1165181 def fullname(klass: Type[Any]) -> str: return f"{klass.__module__}.{klass.__qualname__}" def set_logging_level(level: int) -> None: logging.basicConfig(level=level) # `basicConfig` will only work for new loggers, so we also need to set up the existing ones: for logger in logging.root.manager.loggerDict.values(): if isinstance(logger, logging.Logger): # Otherwise, it could be a `logging.PlaceHolder`. logger.setLevel(level) logging.getLogger().setLevel(level) # The root logger is not present in the previous iterable. def init_cli(cfg: DictConfig) -> None: if cfg.get("silent"): set_logging_level(logging.WARNING) else: set_logging_level(logging.INFO) if "seed" in cfg: seed_everything(cfg.seed, workers=True) def instantiate_data_module(cfg: DictConfig, encoder: ENCODER_OR_ENCODER_MAP) -> pl.LightningDataModule: kwargs = {} if cfg._target_ in {fullname(klass) for klass in [DataModuleStructuredGroup, EvalDataModuleGroup, MixedBatchDataModule]}: if isinstance(cfg.data_modules, Mapping): kwargs["data_modules"] = {k: instantiate_data_module(v, encoder=encoder) # noqa for k, v in cfg.data_modules.items()} else: kwargs["data_modules"] = {instantiate_data_module(cfg_dm, encoder=encoder) for cfg_dm in cfg.data_modules} # Convert because otherwise the passed `data_modules` is a `DictConfig` instead of a `dict` and # `train_dataloader` can't respect the same collection type as `DictConfig` can't have normal classes. kwargs["_convert_"] = "all" elif cfg._target_ == fullname(TrainAndEvalDataModules): kwargs["train_data_module"] = instantiate_data_module(cfg.train_data_module, encoder=encoder) kwargs["eval_data_module"] = instantiate_data_module(cfg.eval_data_module, encoder=encoder) else: kwargs["encoder"] = encoder # Necessary as well when the encoder is a dict. kwargs["_convert_"] = "all" return hydra.utils.instantiate(cfg, **kwargs) def create_model_data_module_trainer_and_ckpt_path( cfg: DictConfig, model_kwargs: Optional[Mapping[str, Any]] = None) -> Tuple[VideoTextLightningModule, pl.LightningDataModule, pl.Trainer, str]: model_kwargs = model_kwargs or {} LOGGER.info(f"Instantiating encoder <{getattr(cfg.encoder, '_target_', type(cfg.encoder).__name__)}>…") encoder: ENCODER_OR_ENCODER_MAP = hydra.utils.instantiate(cfg.encoder) if isinstance(encoder, Mapping) and cfg.get("use_student_encoder_for_data_preprocessing"): encoder_for_data_preprocessing = encoder["student"] else: encoder_for_data_preprocessing = encoder LOGGER.info("Encoder instantiated.") LOGGER.info(f"Instantiating data module <{cfg.data._target_}>…") data_module = instantiate_data_module(cfg.data, encoder=encoder_for_data_preprocessing) LOGGER.info("Data module instantiated.") LOGGER.info(f"Instantiating model <{cfg.model._target_}>…") if isinstance(encoder, Mapping): model_kwargs.setdefault("encoder", encoder["student"]) model_kwargs.setdefault("teacher", encoder["teacher"]) else: model_kwargs.setdefault("encoder", encoder) if isinstance(data_module, VideoClassificationDataModule): assert isinstance(encoder_for_data_preprocessing, VideoTextEncoder), \ "Encoder can't be a mapping and has to support text when doing classification." cfg.model._target_ = fullname(VideoTextClassificationLightningModule) model_kwargs.setdefault("labels", data_module.categories) model_kwargs.setdefault("templates", data_module.templates) if prompts_path := cfg.get("prompts"): # noqa with open(cached_path(prompts_path)) as file: model_kwargs.setdefault("prompts", [stripped_line for line in file if (stripped_line := line.strip())]) # noqa model: VideoTextLightningModule = hydra.utils.instantiate(cfg.model, **model_kwargs) LOGGER.info("Model instantiated.") if "optimizer" in cfg: LOGGER.info(f"Instantiating Optimizer <{cfg.optimizer._target_}>…") def configure_optimizers(self: pl.LightningModule) -> Optimizer: if (lr_ := self.hparams.get("lr")) is not None: # To be used by auto LR find. cfg.optimizer["lr"] = lr_ return hydra.utils.instantiate(cfg.optimizer, self.parameters()) model.configure_optimizers = MethodType(configure_optimizers, model) LOGGER.info("Optimizer instantiated.") LOGGER.info(f"Instantiating trainer <{cfg.trainer._target_}>…") trainer: pl.Trainer = hydra.utils.instantiate(cfg.trainer) LOGGER.info("Trainer instantiated.") # We do what `model.save_hyperparameters(cfg)` would do but without needing a current frame to get the args from. # It turns out that, even if you provide args, it still checks the current frame for args, and set those # conditioned by the provided args. model._log_hyperparams = trainer.logger model._set_hparams(cfg) # noqa model._hparams_initial = copy.deepcopy(model._hparams) ckpt_path = cached_path(cfg.checkpoint_path) if cfg.get("path") else None return model, data_module, trainer, ckpt_path
6,809
44.099338
119
py
fitclip
fitclip-main/aligner/wise.py
import copy from typing import Mapping, TypeVar import torch from torch import nn T = TypeVar("T", bound=nn.Module) def wise_state_dict(model1: T, model2: T, weight_for_2: float = 0.5) -> Mapping[str, torch.Tensor]: state_dict1 = dict(model1.named_parameters()) state_dict2 = dict(model2.named_parameters()) assert set(state_dict1) == set(state_dict2) return {k: (1 - weight_for_2) * state_dict1[k] + weight_for_2 * state_dict2[k] for k in state_dict1} def wise(model1: T, model2: T, weight_for_2: float = 0.5, copy_model1: bool = True) -> T: assert type(model1) is type(model2) model = copy.deepcopy(model1 if copy_model1 else model2) model.load_state_dict(wise_state_dict(model1, model2, weight_for_2=weight_for_2)) # noqa return model
779
31.5
104
py
fitclip
fitclip-main/aligner/param_freezer.py
# Inspired from https://github.com/allenai/allennlp/blob/0d8c0fc/allennlp/training/optimizers.py import logging import re from typing import Iterable, Optional, Union import pytorch_lightning as pl from overrides import overrides LOGGER = logging.getLogger(__name__) class ParamFreezer(pl.Callback): def __init__(self, regexes: Iterable[Union[str, re.Pattern]]) -> None: super().__init__() self.regexes = [re.compile(regex) for regex in regexes] @overrides def setup(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule", stage: Optional[str] = None) -> None: unused_regexes = {p.pattern for p in self.regexes} params_to_tune = [] frozen_params = [] for name, param in pl_module.named_parameters(): for regex in self.regexes: if regex.search(name): param.requires_grad = False if regex.pattern in unused_regexes: unused_regexes.remove(regex.pattern) frozen_params.append(name) break else: params_to_tune.append(name) LOGGER.debug(f"Params to tune: {params_to_tune}") LOGGER.debug(f"Frozen params: {frozen_params}") if unused_regexes: LOGGER.warning(f"The following param regexes used for freezing didn't match any param name: " f"{unused_regexes}")
1,450
32.744186
113
py
fitclip
fitclip-main/aligner/metrics.py
import torch from overrides import overrides from torchmetrics import Metric class Rank(Metric): is_differentiable: bool = False higher_is_better: bool = False full_state_update: bool = False def __init__(self, **kwargs) -> None: super().__init__(**kwargs) self.add_state("ranks", default=[], dist_reduce_fx="cat") @overrides(check_signature=False) def update(self, predictions: torch.Tensor, target: torch.Tensor) -> None: sorted_predicted_positions = predictions.argsort(dim=1, descending=True) ranks = torch.where(sorted_predicted_positions == target.unsqueeze(-1))[1] # noqa self.ranks.append(ranks) @overrides def compute(self) -> torch.Tensor: # It could be already reduced depending on when we call it (e.g., at the epoch end). return self.ranks if isinstance(self.ranks, torch.Tensor) else torch.cat(self.ranks) class MeanRank(Rank): @overrides def compute(self) -> torch.Tensor: return super().compute().mean() + 1 class MedianRank(Rank): @overrides def compute(self) -> torch.Tensor: return super().compute().median() + 1
1,162
30.432432
92
py
fitclip
fitclip-main/aligner/transforms.py
"""From https://github.com/pytorch/vision/blob/993325d/references/video_classification/transforms.py""" import random from typing import Any import torch import torch.nn as nn from overrides import overrides from torchvision.transforms import InterpolationMode, RandomResizedCrop, functional as F from util.tensor_utils import pad class ConvertBHWCtoBCHW(nn.Module): """Convert tensor from (B, H, W, C) to (B, C, H, W).""" @overrides(check_signature=False) def forward(self, v: torch.Tensor) -> torch.Tensor: return v.permute(0, 3, 1, 2) class ConvertBCHWtoCBHW(nn.Module): """Convert tensor from (B, C, H, W) to (C, B, H, W).""" @overrides(check_signature=False) def forward(self, v: torch.Tensor) -> torch.Tensor: return v.permute(1, 0, 2, 3) # Added by me: class ConvertBHWCtoCBHW(nn.Module): """Convert tensor from (B, H, W, C) to (C, B, H, W).""" @overrides(check_signature=False) def forward(self, v: torch.Tensor) -> torch.Tensor: return v.permute(3, 0, 1, 2) class PadToMinFrames: def __init__(self, min_frames: int, frame_dim: int = 0, padding_value: Any = 0) -> None: self.min_frames = min_frames self.frame_dim = frame_dim self.padding_value = padding_value def __call__(self, video: torch.Tensor) -> torch.Tensor: return pad(video, min_size=self.min_frames, dim=self.frame_dim, value=self.padding_value) class MaxFrames: def __init__(self, max_frames: int, frame_dim: int = 0) -> None: self.max_frames = max_frames self.frame_dim = frame_dim def __call__(self, video: torch.Tensor) -> torch.Tensor: return video[(slice(None),) * self.frame_dim + (slice(None, self.max_frames),)] class RandomResizedCropWithRandomInterpolation(RandomResizedCrop): @overrides def forward(self, img: torch.Tensor) -> torch.Tensor: i, j, h, w = self.get_params(img, self.scale, self.ratio) # noqa interpolation = random.choice([InterpolationMode.BILINEAR, InterpolationMode.BICUBIC]) return F.resized_crop(img, i, j, h, w, self.size, interpolation)
2,123
33.258065
103
py
fitclip
fitclip-main/aligner/tests/data/multi_source_sampler_test.py
import string from typing import Literal from torch.utils.data import ConcatDataset, DataLoader, SequentialSampler from aligner.data.multi_source_sampler import RoundRobinMultiSourceSampler def _create_sample_data_loader(mode: Literal["min_size", "max_size_cycle"]) -> DataLoader: dataset1 = string.ascii_lowercase dataset2 = range(10) dataset = ConcatDataset([dataset1, dataset2]) # noqa sampler = RoundRobinMultiSourceSampler([SequentialSampler(dataset1), SequentialSampler(dataset2)], sequence_sizes=[4, 3], mode=mode) return DataLoader(dataset, sampler=sampler, batch_size=None) def test_multi_source_sampler_min_size() -> None: data_loader = _create_sample_data_loader(mode="min_size") expected_list = ["a", "b", "c", "d", 0, 1, 2, "e", "f", "g", "h", 3, 4, 5, "i", "j", "k", "l", 6, 7, 8, "m", "n", "o", "p", 9] assert len(data_loader) == len(expected_list) assert list(data_loader) == expected_list def test_multi_source_sampler_max_size_cycle() -> None: data_loader = _create_sample_data_loader(mode="max_size_cycle") expected_list = ["a", "b", "c", "d", 0, 1, 2, "e", "f", "g", "h", 3, 4, 5, "i", "j", "k", "l", 6, 7, 8, "m", "n", "o", "p", 9, 0, 1, "q", "r", "s", "t", 2, 3, 4, "u", "v", "w", "x", 5, 6, 7, "y", "z"] assert len(data_loader) == len(expected_list) assert list(data_loader) == expected_list
1,463
42.058824
117
py