-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathmtl_loss_schemes.py
263 lines (200 loc) · 8.27 KB
/
mtl_loss_schemes.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
# --------------------------------------------------------
# MTLoRA
# GitHub: https://github.com/scale-lab/MTLoRA
#
# Original file:
# Licensed under the CC BY-NC 4.0 license (https://creativecommons.org/licenses/by-nc/4.0/)
# Written by Simon Vandenhende
#
# Modifications:
# Copyright (c) 2024 SCALE Lab, Brown University
# Licensed under the MIT License (see LICENSE for details)
# --------------------------------------------------------
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.modules.module import Module
import numpy as np
class SoftMaxwithLoss(Module):
"""
This function returns cross entropy loss for semantic segmentation
"""
def __init__(self, ignore_index=255):
super(SoftMaxwithLoss, self).__init__()
self.softmax = nn.LogSoftmax(dim=1)
self.criterion = nn.NLLLoss(ignore_index=ignore_index)
def forward(self, out, label):
assert not label.requires_grad
# out shape batch_size x channels x h x w
# label shape batch_size x 1 x h x w
label = label[:, 0, :, :].long()
loss = self.criterion(self.softmax(out), label)
return loss
class BalancedCrossEntropyLoss(Module):
"""
Balanced Cross Entropy Loss with optional ignore regions
"""
def __init__(self, size_average=True, batch_average=True, pos_weight=None):
super(BalancedCrossEntropyLoss, self).__init__()
self.size_average = size_average
self.batch_average = batch_average
self.pos_weight = pos_weight
def forward(self, output, label, void_pixels=None):
assert (output.size() == label.size())
labels = torch.ge(label, 0.5).float()
# Weighting of the loss, default is HED-style
if self.pos_weight is None:
num_labels_pos = torch.sum(labels)
num_labels_neg = torch.sum(1.0 - labels)
num_total = num_labels_pos + num_labels_neg
w = num_labels_neg / num_total
else:
w = self.pos_weight
output_gt_zero = torch.ge(output, 0).float()
loss_val = torch.mul(output, (labels - output_gt_zero)) - torch.log(
1 + torch.exp(output - 2 * torch.mul(output, output_gt_zero)))
loss_pos_pix = -torch.mul(labels, loss_val)
loss_neg_pix = -torch.mul(1.0 - labels, loss_val)
if void_pixels is not None and not self.pos_weight:
w_void = torch.le(void_pixels, 0.5).float()
loss_pos_pix = torch.mul(w_void, loss_pos_pix)
loss_neg_pix = torch.mul(w_void, loss_neg_pix)
num_total = num_total - torch.ge(void_pixels, 0.5).float().sum()
w = num_labels_neg / num_total
loss_pos = torch.sum(loss_pos_pix)
loss_neg = torch.sum(loss_neg_pix)
final_loss = w * loss_pos + (1 - w) * loss_neg
if self.size_average:
final_loss /= float(np.prod(label.size()))
elif self.batch_average:
final_loss /= label.size()[0]
return final_loss
class BinaryCrossEntropyLoss(Module):
"""
Binary Cross Entropy with ignore regions, not balanced.
"""
def __init__(self, size_average=True, batch_average=True):
super(BinaryCrossEntropyLoss, self).__init__()
self.size_average = size_average
self.batch_average = batch_average
def forward(self, output, label, void_pixels=None):
assert (output.size() == label.size())
labels = torch.ge(label, 0.5).float()
output_gt_zero = torch.ge(output, 0).float()
loss_val = torch.mul(output, (labels - output_gt_zero)) - torch.log(
1 + torch.exp(output - 2 * torch.mul(output, output_gt_zero)))
loss_pos_pix = -torch.mul(labels, loss_val)
loss_neg_pix = -torch.mul(1.0 - labels, loss_val)
if void_pixels is not None:
w_void = torch.le(void_pixels, 0.5).float()
loss_pos_pix = torch.mul(w_void, loss_pos_pix)
loss_neg_pix = torch.mul(w_void, loss_neg_pix)
loss_pos = torch.sum(loss_pos_pix)
loss_neg = torch.sum(loss_neg_pix)
final_loss = loss_pos + loss_neg
if self.size_average:
final_loss /= float(np.prod(label.size()))
elif self.batch_average:
final_loss /= label.size()[0]
return final_loss
class DepthLoss(nn.Module):
"""
Loss for depth prediction. By default L1 loss is used.
"""
def __init__(self, loss='l1'):
super(DepthLoss, self).__init__()
if loss == 'l1':
self.loss = nn.L1Loss()
else:
raise NotImplementedError(
'Loss {} currently not supported in DepthLoss'.format(loss))
def forward(self, out, label):
mask = (label != 255)
return self.loss(torch.masked_select(out, mask), torch.masked_select(label, mask))
class Normalize(nn.Module):
def __init__(self):
super(Normalize, self).__init__()
def forward(self, bottom):
qn = torch.norm(bottom, p=2, dim=1).unsqueeze(dim=1) + 1e-12
top = bottom.div(qn)
return top
class NormalsLoss(Module):
"""
L1 loss with ignore labels
normalize: normalization for surface normals
"""
def __init__(self, size_average=True, normalize=False, norm=1):
super(NormalsLoss, self).__init__()
self.size_average = size_average
if normalize:
self.normalize = Normalize()
else:
self.normalize = None
if norm == 1:
# print('Using L1 loss for surface normals')
self.loss_func = F.l1_loss
elif norm == 2:
# print('Using L2 loss for surface normals')
self.loss_func = F.mse_loss
else:
raise NotImplementedError
def forward(self, out, label, ignore_label=255):
assert not label.requires_grad
mask = (label != ignore_label)
n_valid = torch.sum(mask).item()
if self.normalize is not None:
out_norm = self.normalize(out)
loss = self.loss_func(torch.masked_select(
out_norm, mask), torch.masked_select(label, mask), reduction='sum')
else:
loss = self.loss_func(torch.masked_select(
out, mask), torch.masked_select(label, mask), reduction='sum')
if self.size_average:
if ignore_label:
ret_loss = torch.div(loss, max(n_valid, 1e-6))
return ret_loss
else:
ret_loss = torch.div(loss, float(np.prod(label.size())))
return ret_loss
return loss
class SingleTaskLoss(nn.Module):
def __init__(self, loss_ft, task):
super(SingleTaskLoss, self).__init__()
self.loss_ft = loss_ft
self.task = task
def forward(self, pred, gt):
out = {self.task: self.loss_ft(pred[self.task], gt[self.task])}
out['total'] = out[self.task]
return out
class MultiTaskLoss(nn.Module):
def __init__(self, tasks: list, loss_ft: nn.ModuleDict, loss_weights: dict):
super(MultiTaskLoss, self).__init__()
assert (set(tasks) == set(loss_ft.keys()))
assert (set(tasks) == set(loss_weights.keys()))
self.tasks = tasks
self.loss_ft = loss_ft
self.loss_weights = loss_weights
def forward(self, pred, gt):
out = {
task: self.loss_ft[task](pred[task], gt[task]) for task in self.tasks
}
out['total'] = torch.sum(torch.stack(
[self.loss_weights[t] * out[t] for t in self.tasks]))
return out['total'], out
def get_loss(task_cfg, task=None, config={"DATA": {}}):
""" Return loss function for a specific task """
if task == 'edge':
criterion = BalancedCrossEntropyLoss(
size_average=True, pos_weight=task_cfg.get('edge_w', 0.95))
elif task == 'semseg' or task == 'human_parts':
criterion = SoftMaxwithLoss(ignore_index=255)
elif task == 'normals':
criterion = NormalsLoss(normalize=True, size_average=True, norm=1)
elif task == 'sal':
criterion = BalancedCrossEntropyLoss(size_average=True)
elif task == 'depth':
criterion = DepthLoss('l1')
else:
raise NotImplementedError('Undefined Loss: Choose a task among '
'edge, semseg, human_parts, sal, depth, or normals')
return criterion