import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from torch.autograd import Variable from torch.distributions.categorical import Categorical from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence import torch_ac from utils.babyai_utils.supervised_losses import required_heads import gym.spaces as spaces def safe_relu(x): return torch.maximum(x, torch.zeros_like(x)) # From https://github.com/ikostrikov/pytorch-a2c-ppo-acktr/blob/master/model.py def initialize_parameters(m): classname = m.__class__.__name__ if classname.find('Linear') != -1: m.weight.data.normal_(0, 1) m.weight.data *= 1 / torch.sqrt(m.weight.data.pow(2).sum(1, keepdim=True)) if m.bias is not None: m.bias.data.fill_(0) # Inspired by FiLMedBlock from https://arxiv.org/abs/1709.07871 class FiLM(nn.Module): def __init__(self, in_features, out_features, in_channels, imm_channels): super().__init__() self.conv1 = nn.Conv2d( in_channels=in_channels, out_channels=imm_channels, kernel_size=(3, 3), padding=1) self.bn1 = nn.BatchNorm2d(imm_channels) self.conv2 = nn.Conv2d( in_channels=imm_channels, out_channels=out_features, kernel_size=(3, 3), padding=1) self.bn2 = nn.BatchNorm2d(out_features) self.weight = nn.Linear(in_features, out_features) self.bias = nn.Linear(in_features, out_features) self.apply(initialize_parameters) def forward(self, x, y): x = F.relu(self.bn1(self.conv1(x))) x = self.conv2(x) weight = self.weight(y).unsqueeze(2).unsqueeze(3) bias = self.bias(y).unsqueeze(2).unsqueeze(3) out = x * weight + bias # return F.relu(self.bn2(out)) # this causes an error in the new version of pytorch -> replaced by safe_relu return safe_relu(self.bn2(out)) class ImageBOWEmbedding(nn.Module): def __init__(self, space, embedding_dim): super().__init__() # self.max_value = max(space) self.max_value = 255 # 255, because of "no_point" encoding, which is encoded as 255 self.space = space self.embedding_dim = embedding_dim self.embedding = nn.Embedding(self.space[-1] * self.max_value, embedding_dim) self.apply(initialize_parameters) def forward(self, inputs): offsets = torch.Tensor([x * self.max_value for x in range(self.space[-1])]).to(inputs.device) inputs = (inputs + offsets[None, :, None, None]).long() return self.embedding(inputs).sum(1).permute(0, 3, 1, 2) #notes: what they call instr is what we call text #class ACModel(nn.Module, babyai.rl.RecurrentACModel): # instr (them) == text (us) class MultiModalBaby11ACModel(nn.Module, torch_ac.RecurrentACModel): def __init__(self, obs_space, action_space, image_dim=128, memory_dim=128, text_dim=128, dialog_dim=128, use_text=False, use_dialogue=False, use_current_dialogue_only=False, lang_model="gru", use_memory=False, arch="bow_endpool_res", aux_info=None, num_films=2): super().__init__() # store config self.config = locals() # multi dim if action_space.shape == (): raise ValueError("The action space is not multi modal. Use ACModel instead.") if use_text: # for now we do not consider goal conditioned policies raise ValueError("You should not use text but dialogue. --text is cheating.") endpool = 'endpool' in arch use_bow = 'bow' in arch pixel = 'pixel' in arch self.res = 'res' in arch # Decide which components are enabled self.use_text = use_text self.use_dialogue = use_dialogue self.use_current_dialogue_only = use_current_dialogue_only self.use_memory = use_memory self.arch = arch self.lang_model = lang_model self.aux_info = aux_info if self.res and image_dim != 128: raise ValueError(f"image_dim is {image_dim}, expected 128") self.image_dim = image_dim self.memory_dim = memory_dim self.text_dim = text_dim self.dialog_dim = dialog_dim self.num_module = num_films self.n_primitive_actions = action_space.nvec[0] + 1 # not move action added self.move_switch_action = int(self.n_primitive_actions) - 1 self.n_utterance_actions = np.concatenate(([2], action_space.nvec[1:])) # binary to not speak self.talk_switch_subhead = 0 self.env_action_space = action_space self.model_raw_action_space = spaces.MultiDiscrete([self.n_primitive_actions, *self.n_utterance_actions]) self.obs_space = obs_space # transform given 3d obs_space into what babyai11 baseline uses, i.e. 1d embedding size n = obs_space["image"][0] m = obs_space["image"][1] nb_img_channels = self.obs_space['image'][2] self.obs_space = ((n-1)//2-2)*((m-1)//2-2)*64 for part in self.arch.split('_'): if part not in ['original', 'bow', 'pixels', 'endpool', 'res']: raise ValueError("Incorrect architecture name: {}".format(self.arch)) # if not self.use_text: # raise ValueError("FiLM architecture can be used when textuctions are enabled") self.image_conv = nn.Sequential(*[ *([ImageBOWEmbedding(obs_space['image'], 128)] if use_bow else []), *([nn.Conv2d( in_channels=nb_img_channels, out_channels=128, kernel_size=(8, 8), stride=8, padding=0)] if pixel else []), nn.Conv2d( in_channels=128 if use_bow or pixel else nb_img_channels, out_channels=128, kernel_size=(3, 3) if endpool else (2, 2), stride=1, padding=1), nn.BatchNorm2d(128), nn.ReLU(), *([] if endpool else [nn.MaxPool2d(kernel_size=(2, 2), stride=2)]), nn.Conv2d(in_channels=128, out_channels=128, kernel_size=(3, 3), padding=1), nn.BatchNorm2d(128), nn.ReLU(), *([] if endpool else [nn.MaxPool2d(kernel_size=(2, 2), stride=2)]) ]) self.film_pool = nn.MaxPool2d(kernel_size=(7, 7) if endpool else (2, 2), stride=2) # Define DIALOGUE embedding if self.use_dialogue or self.use_current_dialogue_only: if self.lang_model in ['gru', 'bigru', 'attgru']: #self.word_embedding = nn.Embedding(obs_space["instr"], self.dialog_dim) self.word_embedding = nn.Embedding(obs_space["text"], self.dialog_dim) if self.lang_model in ['gru', 'bigru', 'attgru']: gru_dim = self.dialog_dim if self.lang_model in ['bigru', 'attgru']: gru_dim //= 2 self.dialog_rnn = nn.GRU( self.dialog_dim, gru_dim, batch_first=True, bidirectional=(self.lang_model in ['bigru', 'attgru'])) self.final_dialog_dim = self.dialog_dim else: kernel_dim = 64 kernel_sizes = [3, 4] self.dialog_convs = nn.ModuleList([ nn.Conv2d(1, kernel_dim, (K, self.dialog_dim)) for K in kernel_sizes]) self.final_dialog_dim = kernel_dim * len(kernel_sizes) if self.lang_model == 'attgru': self.memory2key = nn.Linear(self.memory_size, self.final_dialog_dim) self.controllers = [] for ni in range(self.num_module): mod = FiLM( in_features=self.final_dialog_dim, out_features=128 if ni < self.num_module-1 else self.image_dim, in_channels=128, imm_channels=128) self.controllers.append(mod) self.add_module('FiLM_' + str(ni), mod) # Define memory and resize image embedding self.embedding_size = self.image_dim if self.use_memory: self.memory_rnn = nn.LSTMCell(self.image_dim, self.memory_dim) self.embedding_size = self.semi_memory_size # Define actor's model self.actor = nn.Sequential( nn.Linear(self.embedding_size, 64), nn.Tanh(), nn.Linear(64, self.n_primitive_actions) ) self.talker = nn.ModuleList([ nn.Sequential( nn.Linear(self.embedding_size, 64), nn.Tanh(), nn.Linear(64, n) ) for n in self.n_utterance_actions]) # Define critic's model self.critic = nn.Sequential( nn.Linear(self.embedding_size, 64), nn.Tanh(), nn.Linear(64, 1) ) # Initialize parameters correctly self.apply(initialize_parameters) # Define head for extra info if self.aux_info: self.extra_heads = None self.add_heads() def add_heads(self): ''' When using auxiliary tasks, the environment yields at each step some binary, continous, or multiclass information. The agent needs to predict those information. This function add extra heads to the model that output the predictions. There is a head per extra information (the head type depends on the extra information type). ''' self.extra_heads = nn.ModuleDict() for info in self.aux_info: if required_heads[info] == 'binary': self.extra_heads[info] = nn.Linear(self.embedding_size, 1) elif required_heads[info].startswith('multiclass'): n_classes = int(required_heads[info].split('multiclass')[-1]) self.extra_heads[info] = nn.Linear(self.embedding_size, n_classes) elif required_heads[info].startswith('continuous'): if required_heads[info].endswith('01'): self.extra_heads[info] = nn.Sequential(nn.Linear(self.embedding_size, 1), nn.Sigmoid()) else: raise ValueError('Only continous01 is implemented') else: raise ValueError('Type not supported') # initializing these parameters independently is done in order to have consistency of results when using # supervised-loss-coef = 0 and when not using any extra binary information self.extra_heads[info].apply(initialize_parameters) def add_extra_heads_if_necessary(self, aux_info): ''' This function allows using a pre-trained model without aux_info and add aux_info to it and still make it possible to finetune. ''' try: if not hasattr(self, 'aux_info') or not set(self.aux_info) == set(aux_info): self.aux_info = aux_info self.add_heads() except Exception: raise ValueError('Could not add extra heads') @property def memory_size(self): return 2 * self.semi_memory_size @property def semi_memory_size(self): return self.memory_dim def forward(self, obs, memory, dialog_embedding=None, return_embeddings=False): if self.use_dialogue and dialog_embedding is None: if not hasattr(obs, "utterance_history"): raise ValueError("The environment need's to be updated to 'utterance' and 'utterance_history' keys'") dialog_embedding = self._get_dialog_embedding(obs.utterance_history) elif self.use_current_dialogue_only and dialog_embedding is None: if not hasattr(obs, "utterance"): raise ValueError("The environment need's to be updated to 'utterance' and 'utterance_history' keys'") dialog_embedding = self._get_dialog_embedding(obs.utterance) if (self.use_dialogue or self.use_current_dialogue_only) and self.lang_model == "attgru": # outputs: B x L x D # memory: B x M #mask = (obs.instr != 0).float() mask = (obs.utterance_history != 0).float() # The mask tensor has the same length as obs.instr, and # thus can be both shorter and longer than instr_embedding. # It can be longer if instr_embedding is computed # for a subbatch of obs.instr. # It can be shorter if obs.instr is a subbatch of # the batch that instr_embeddings was computed for. # Here, we make sure that mask and instr_embeddings # have equal length along dimension 1. mask = mask[:, :dialog_embedding.shape[1]] dialog_embedding = dialog_embedding[:, :mask.shape[1]] keys = self.memory2key(memory) pre_softmax = (keys[:, None, :] * dialog_embedding).sum(2) + 1000 * mask attention = F.softmax(pre_softmax, dim=1) dialog_embedding = (dialog_embedding * attention[:, :, None]).sum(1) x = torch.transpose(torch.transpose(obs.image, 1, 3), 2, 3) if 'pixel' in self.arch: x /= 256.0 x = self.image_conv(x) if (self.use_dialogue or self.use_current_dialogue_only): for controller in self.controllers: out = controller(x, dialog_embedding) if self.res: out += x x = out x = F.relu(self.film_pool(x)) x = x.reshape(x.shape[0], -1) if self.use_memory: hidden = (memory[:, :self.semi_memory_size], memory[:, self.semi_memory_size:]) hidden = self.memory_rnn(x, hidden) embedding = hidden[0] memory = torch.cat(hidden, dim=1) else: embedding = x if hasattr(self, 'aux_info') and self.aux_info: extra_predictions = {info: self.extra_heads[info](embedding) for info in self.extra_heads} else: extra_predictions = dict() # x = self.actor(embedding) # dist = Categorical(logits=F.log_softmax(x, dim=1)) x = self.actor(embedding) primitive_actions_dist = Categorical(logits=F.log_softmax(x, dim=1)) x = self.critic(embedding) value = x.squeeze(1) utterance_actions_dists = [ Categorical(logits=F.log_softmax( tal(embedding), dim=1, )) for tal in self.talker ] dist = [primitive_actions_dist] + utterance_actions_dists #return {'dist': dist, 'value': value, 'memory': memory, 'extra_predictions': extra_predictions} if return_embeddings: return dist, value, memory, embedding else: return dist, value, memory def _get_dialog_embedding(self, dialog): lengths = (dialog != 0).sum(1).long() if self.lang_model == 'gru': out, _ = self.dialog_rnn(self.word_embedding(dialog)) hidden = out[range(len(lengths)), lengths-1, :] return hidden elif self.lang_model in ['bigru', 'attgru']: masks = (dialog != 0).float() if lengths.shape[0] > 1: seq_lengths, perm_idx = lengths.sort(0, descending=True) iperm_idx = torch.LongTensor(perm_idx.shape).fill_(0) if dialog.is_cuda: iperm_idx = iperm_idx.cuda() for i, v in enumerate(perm_idx): iperm_idx[v.data] = i inputs = self.word_embedding(dialog) inputs = inputs[perm_idx] inputs = pack_padded_sequence(inputs, seq_lengths.data.cpu().numpy(), batch_first=True) outputs, final_states = self.dialog_rnn(inputs) else: dialog = dialog[:, 0:lengths[0]] outputs, final_states = self.dialog_rnn(self.word_embedding(dialog)) iperm_idx = None final_states = final_states.transpose(0, 1).contiguous() final_states = final_states.view(final_states.shape[0], -1) if iperm_idx is not None: outputs, _ = pad_packed_sequence(outputs, batch_first=True) outputs = outputs[iperm_idx] final_states = final_states[iperm_idx] return outputs if self.lang_model == 'attgru' else final_states else: ValueError("Undefined lang_model architecture: {}".format(self.lang_model)) # add action sampling to fit our interaction pipeline ## baby ai [[Categorical(logits: torch.Size([16, 8])), Categorical(logits: torch.Size([16, 2])), Categorical(logits: torch.Size([16, 2]))]] ## mh ac [Categorical(logits: torch.Size([16, 8])), Categorical(logits: torch.Size([16, 2])), Categorical(logits: torch.Size([16, 2]))] def det_action(self, dist): return torch.stack([d.probs.argmax(dim=-1) for d in dist], dim=1) def sample_action(self, dist): return torch.stack([d.sample() for d in dist], dim=1) def is_raw_action_speaking(self, action): is_speaking = action[:, 1:][:, self.talk_switch_subhead] == 1 # talking heads are [1:] return is_speaking def no_speak_to_speak_action(self, action): action[:, 1] = 1 # set speaking action to speak (1) assert all(self.is_raw_action_speaking(action)) return action def raw_action_to_act_speak_mask(self, action): """ Defines how the final action to be sent to the environment is computed Does NOT define how gradients are propagated, see calculate_action_gradient_masks() for that """ assert action.shape[-1] == 4 assert self.model_raw_action_space.shape[0] == action.shape[-1] act_mask = action[:, 0] != self.move_switch_action # acting head is [0] # speak_mask = action[:, 1:][:, self.talk_switch_subhead] == 1 # talking heads are [1:] speak_mask = self.is_raw_action_speaking(action) return act_mask, speak_mask def construct_final_action(self, action): act_mask, speak_mask = self.raw_action_to_act_speak_mask(action) nan_mask = np.stack((act_mask, speak_mask, speak_mask), axis=1).astype(float) nan_mask[nan_mask == 0] = np.nan assert self.talk_switch_subhead == 0 final_action = action[:, [True, False, True, True]] # we drop the talk_switch_subhead final_action = nan_mask*final_action assert self.env_action_space.shape[0] == final_action.shape[-1] return final_action # add calculate log probs to fit our interaction pipeline def calculate_log_probs(self, dist, action): return torch.stack([d.log_prob(action[:, i]) for i, d in enumerate(dist)], dim=1) # add calculate action masks to fit our interaction pipeline def calculate_action_gradient_masks(self, action): """ Defines how the gradients are propagated. Moving head is always trained. Speak switch is always trained. Grammar heads are trained only when speak switch is ON """ _, speak_mask = self.raw_action_to_act_speak_mask(action) mask = torch.stack( ( torch.ones_like(speak_mask), # always train torch.ones_like(speak_mask), # always train speak_mask, # train only when speaking speak_mask, # train only when speaking ), dim=1).detach() assert action.shape == mask.shape return mask def get_config_dict(self): del self.config['__class__'] self.config['self'] = str(self.config['self']) self.config['action_space'] = self.config['action_space'].nvec.tolist() return self.config