TreeFormer / losses /consistency_loss.py
franciszzj's picture
init
c964d4c
raw
history blame
No virus
12 kB
import numpy as np
import torch
import torch.nn.functional as F
import torch.nn as nn
from losses import ramps
class consistency_weight(object):
"""
ramp_types = ['sigmoid_rampup', 'linear_rampup', 'cosine_rampup', 'log_rampup', 'exp_rampup']
"""
def __init__(self, final_w, iters_per_epoch, rampup_starts=0, rampup_ends=7, ramp_type='sigmoid_rampup'):
self.final_w = final_w
self.iters_per_epoch = iters_per_epoch
self.rampup_starts = rampup_starts * iters_per_epoch
self.rampup_ends = rampup_ends * iters_per_epoch
self.rampup_length = (self.rampup_ends - self.rampup_starts)
self.rampup_func = getattr(ramps, ramp_type)
self.current_rampup = 0
def __call__(self, epoch, curr_iter):
cur_total_iter = self.iters_per_epoch * epoch + curr_iter
if cur_total_iter < self.rampup_starts:
return 0
self.current_rampup = self.rampup_func(cur_total_iter - self.rampup_starts, self.rampup_length)
return self.final_w * self.current_rampup
def CE_loss(input_logits, target_targets, ignore_index, temperature=1):
return F.cross_entropy(input_logits/temperature, target_targets, ignore_index=ignore_index)
# for FocalLoss
def softmax_helper(x):
# copy from: https://github.com/MIC-DKFZ/nnUNet/blob/master/nnunet/utilities/nd_softmax.py
rpt = [1 for _ in range(len(x.size()))]
rpt[1] = x.size(1)
x_max = x.max(1, keepdim=True)[0].repeat(*rpt)
e_x = torch.exp(x - x_max)
return e_x / e_x.sum(1, keepdim=True).repeat(*rpt)
def get_alpha(supervised_loader):
# get number of classes
num_labels = 0
for image_batch, label_batch in supervised_loader:
label_batch.data[label_batch.data==255] = 0 # pixels of ignore class added to background
l_unique = torch.unique(label_batch.data)
list_unique = [element.item() for element in l_unique.flatten()]
num_labels = max(max(list_unique),num_labels)
num_classes = num_labels + 1
# count class occurrences
alpha = [0 for i in range(num_classes)]
for image_batch, label_batch in supervised_loader:
label_batch.data[label_batch.data==255] = 0 # pixels of ignore class added to background
l_unique = torch.unique(label_batch.data)
list_unique = [element.item() for element in l_unique.flatten()]
l_unique_count = torch.stack([(label_batch.data==x_u).sum() for x_u in l_unique]) # tensor([65920, 36480])
list_count = [count.item() for count in l_unique_count.flatten()]
for index in list_unique:
alpha[index] += list_count[list_unique.index(index)]
return alpha
# for FocalLoss
def softmax_helper(x):
# copy from: https://github.com/MIC-DKFZ/nnUNet/blob/master/nnunet/utilities/nd_softmax.py
rpt = [1 for _ in range(len(x.size()))]
rpt[1] = x.size(1)
x_max = x.max(1, keepdim=True)[0].repeat(*rpt)
e_x = torch.exp(x - x_max)
return e_x / e_x.sum(1, keepdim=True).repeat(*rpt)
class FocalLoss(nn.Module):
"""
copy from: https://github.com/Hsuxu/Loss_ToolBox-PyTorch/blob/master/FocalLoss/FocalLoss.py
This is a implementation of Focal Loss with smooth label cross entropy supported which is proposed in
'Focal Loss for Dense Object Detection. (https://arxiv.org/abs/1708.02002)'
Focal_Loss= -1*alpha*(1-pt)*log(pt)
:param num_class:
:param alpha: (tensor) 3D or 4D the scalar factor for this criterion
:param gamma: (float,double) gamma > 0 reduces the relative loss for well-classified examples (p>0.5) putting more
focus on hard misclassified example
:param smooth: (float,double) smooth value when cross entropy
:param balance_index: (int) balance class index, should be specific when alpha is float
:param size_average: (bool, optional) By default, the losses are averaged over each loss element in the batch.
"""
def __init__(self, apply_nonlin=None, ignore_index = None, alpha=None, gamma=2, balance_index=0, smooth=1e-5, size_average=True):
super(FocalLoss, self).__init__()
self.apply_nonlin = apply_nonlin
self.alpha = alpha
self.gamma = gamma
self.balance_index = balance_index
self.smooth = smooth
self.size_average = size_average
if self.smooth is not None:
if self.smooth < 0 or self.smooth > 1.0:
raise ValueError('smooth value should be in [0,1]')
def forward(self, logit, target):
if self.apply_nonlin is not None:
logit = self.apply_nonlin(logit)
num_class = logit.shape[1]
if logit.dim() > 2:
# N,C,d1,d2 -> N,C,m (m=d1*d2*...)
logit = logit.view(logit.size(0), logit.size(1), -1)
logit = logit.permute(0, 2, 1).contiguous()
logit = logit.view(-1, logit.size(-1))
target = torch.squeeze(target, 1)
target = target.view(-1, 1)
valid_mask = None
if self.ignore_index is not None:
valid_mask = target != self.ignore_index
target = target * valid_mask
alpha = self.alpha
if alpha is None:
alpha = torch.ones(num_class, 1)
elif isinstance(alpha, (list, np.ndarray)):
assert len(alpha) == num_class
alpha = torch.FloatTensor(alpha).view(num_class, 1)
alpha = alpha / alpha.sum()
alpha = 1/alpha # inverse of class frequency
elif isinstance(alpha, float):
alpha = torch.ones(num_class, 1)
alpha = alpha * (1 - self.alpha)
alpha[self.balance_index] = self.alpha
else:
raise TypeError('Not support alpha type')
if alpha.device != logit.device:
alpha = alpha.to(logit.device)
idx = target.cpu().long()
one_hot_key = torch.FloatTensor(target.size(0), num_class).zero_()
# to resolve error in idx in scatter_
idx[idx==225]=0
one_hot_key = one_hot_key.scatter_(1, idx, 1)
if one_hot_key.device != logit.device:
one_hot_key = one_hot_key.to(logit.device)
if self.smooth:
one_hot_key = torch.clamp(
one_hot_key, self.smooth/(num_class-1), 1.0 - self.smooth)
pt = (one_hot_key * logit).sum(1) + self.smooth
logpt = pt.log()
gamma = self.gamma
alpha = alpha[idx]
alpha = torch.squeeze(alpha)
loss = -1 * alpha * torch.pow((1 - pt), gamma) * logpt
if valid_mask is not None:
loss = loss * valid_mask.squeeze()
if self.size_average:
loss = loss.mean()
else:
loss = loss.sum()
return loss
class abCE_loss(nn.Module):
"""
Annealed-Bootstrapped cross-entropy loss
"""
def __init__(self, iters_per_epoch, epochs, num_classes, weight=None,
reduction='mean', thresh=0.7, min_kept=1, ramp_type='log_rampup'):
super(abCE_loss, self).__init__()
self.weight = torch.FloatTensor(weight) if weight is not None else weight
self.reduction = reduction
self.thresh = thresh
self.min_kept = min_kept
self.ramp_type = ramp_type
if ramp_type is not None:
self.rampup_func = getattr(ramps, ramp_type)
self.iters_per_epoch = iters_per_epoch
self.num_classes = num_classes
self.start = 1/num_classes
self.end = 0.9
self.total_num_iters = (epochs - (0.6 * epochs)) * iters_per_epoch
def threshold(self, curr_iter, epoch):
cur_total_iter = self.iters_per_epoch * epoch + curr_iter
current_rampup = self.rampup_func(cur_total_iter, self.total_num_iters)
return current_rampup * (self.end - self.start) + self.start
def forward(self, predict, target, ignore_index, curr_iter, epoch):
batch_kept = self.min_kept * target.size(0)
prob_out = F.softmax(predict, dim=1)
tmp_target = target.clone()
tmp_target[tmp_target == ignore_index] = 0
prob = prob_out.gather(1, tmp_target.unsqueeze(1))
mask = target.contiguous().view(-1, ) != ignore_index
sort_prob, sort_indices = prob.contiguous().view(-1, )[mask].contiguous().sort()
if self.ramp_type is not None:
thresh = self.threshold(curr_iter=curr_iter, epoch=epoch)
else:
thresh = self.thresh
min_threshold = sort_prob[min(batch_kept, sort_prob.numel() - 1)] if sort_prob.numel() > 0 else 0.0
threshold = max(min_threshold, thresh)
loss_matrix = F.cross_entropy(predict, target,
weight=self.weight.to(predict.device) if self.weight is not None else None,
ignore_index=ignore_index, reduction='none')
loss_matirx = loss_matrix.contiguous().view(-1, )
sort_loss_matirx = loss_matirx[mask][sort_indices]
select_loss_matrix = sort_loss_matirx[sort_prob < threshold]
if self.reduction == 'sum' or select_loss_matrix.numel() == 0:
return select_loss_matrix.sum()
elif self.reduction == 'mean':
return select_loss_matrix.mean()
else:
raise NotImplementedError('Reduction Error!')
def softmax_mse_loss(inputs, targets, conf_mask=False, threshold=None, use_softmax=False):
assert inputs.requires_grad == True and targets.requires_grad == False
assert inputs.size() == targets.size() # (batch_size * num_classes * H * W)
inputs = F.softmax(inputs, dim=1)
if use_softmax:
targets = F.softmax(targets, dim=1)
if conf_mask:
loss_mat = F.mse_loss(inputs, targets, reduction='none')
mask = (targets.max(1)[0] > threshold)
loss_mat = loss_mat[mask.unsqueeze(1).expand_as(loss_mat)]
if loss_mat.shape.numel() == 0: loss_mat = torch.tensor([0.]).to(inputs.device)
return loss_mat.mean()
else:
return F.mse_loss(inputs, targets, reduction='mean') # take the mean over the batch_size
def softmax_kl_loss(inputs, targets, conf_mask=False, threshold=None, use_softmax=False):
assert inputs.requires_grad == True and targets.requires_grad == False
assert inputs.size() == targets.size()
if use_softmax:
targets = F.softmax(targets, dim=1)
if conf_mask:
loss_mat = F.kl_div(input_log_softmax, targets, reduction='none')
mask = (targets.max(1)[0] > threshold)
loss_mat = loss_mat[mask.unsqueeze(1).expand_as(loss_mat)]
if loss_mat.shape.numel() == 0: loss_mat = torch.tensor([0.]).to(inputs.device)
return loss_mat.sum() / mask.shape.numel()
else:
return F.kl_div(inputs, targets, reduction='mean')
def softmax_js_loss(inputs, targets, **_):
assert inputs.requires_grad == True and targets.requires_grad == False
assert inputs.size() == targets.size()
epsilon = 1e-5
M = (F.softmax(inputs, dim=1) + targets) * 0.5
kl1 = F.kl_div(F.log_softmax(inputs, dim=1), M, reduction='mean')
kl2 = F.kl_div(torch.log(targets+epsilon), M, reduction='mean')
return (kl1 + kl2) * 0.5
def pair_wise_loss(unsup_outputs, size_average=True, nbr_of_pairs=8):
"""
Pair-wise loss in the sup. mat.
"""
if isinstance(unsup_outputs, list):
unsup_outputs = torch.stack(unsup_outputs)
# Only for a subset of the aux outputs to reduce computation and memory
unsup_outputs = unsup_outputs[torch.randperm(unsup_outputs.size(0))]
unsup_outputs = unsup_outputs[:nbr_of_pairs]
temp = torch.zeros_like(unsup_outputs) # For grad purposes
for i, u in enumerate(unsup_outputs):
temp[i] = F.softmax(u, dim=1)
mean_prediction = temp.mean(0).unsqueeze(0) # Mean over the auxiliary outputs
pw_loss = ((temp - mean_prediction)**2).mean(0) # Variance
pw_loss = pw_loss.sum(1) # Sum over classes
if size_average:
return pw_loss.mean()
return pw_loss.sum()