-
Notifications
You must be signed in to change notification settings - Fork 43
/
Copy pathloss.py
394 lines (326 loc) · 15 KB
/
loss.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
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
"""
Loss.py
"""
import logging
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from config import cfg
def get_loss(args):
"""
Get the criterion based on the loss function
args: commandline arguments
return: criterion, criterion_val
"""
if args.img_wt_loss:
criterion = ImageBasedCrossEntropyLoss2d(
classes=args.dataset_cls.num_classes, size_average=True,
ignore_index=args.dataset_cls.ignore_label,
upper_bound=args.wt_bound).cuda()
elif args.jointwtborder:
if args.joint_edgeseg_loss:
# joint edge loss with boundary relax
criterion = JointEdgeSegLoss(classes=args.dataset_cls.num_classes,
ignore_index=args.dataset_cls.ignore_label,
edge_weight=args.edge_weight, seg_weight=args.seg_weight,
att_weight=args.att_weight).cuda()
else:
# add ohem option
criterion = ImgWtLossSoftNLL(classes=args.dataset_cls.num_classes,
ignore_index=args.dataset_cls.ignore_label,
upper_bound=args.wt_bound, ohem=args.ohem).cuda()
elif args.fpn_dsn_loss:
criterion = CriterionSFNet(ignore_index=args.dataset_cls.ignore_label)
else:
if args.ohem and args.aux:
criterion = OhemWithAux(ignore_index=args.dataset_cls.ignore_label).cuda()
elif args.ohem and not args.aux:
criterion = OhemCrossEntropy2dTensor(ignore_index=args.dataset_cls.ignore_label).cuda()
else:
criterion = CrossEntropyLoss2d(size_average=True,
ignore_index=args.dataset_cls.ignore_label).cuda()
criterion_val = CrossEntropyLoss2d(size_average=True,
weight=None,
ignore_index=args.dataset_cls.ignore_label).cuda()
return criterion, criterion_val
class ImageBasedCrossEntropyLoss2d(nn.Module):
"""
Image Weighted Cross Entropy Loss
"""
def __init__(self, classes, weight=None, size_average=True, ignore_index=255,
norm=False, upper_bound=1.0):
super(ImageBasedCrossEntropyLoss2d, self).__init__()
logging.info("Using Per Image based weighted loss")
self.num_classes = classes
self.nll_loss = nn.NLLLoss2d(weight, size_average, ignore_index)
self.norm = norm
self.upper_bound = upper_bound
self.batch_weights = cfg.BATCH_WEIGHTING
def calculate_weights(self, target):
"""
Calculate weights of classes based on the training crop
"""
hist = np.histogram(target.flatten(), range(
self.num_classes + 1), normed=True)[0]
if self.norm:
hist = ((hist != 0) * self.upper_bound * (1 / hist)) + 1
else:
hist = ((hist != 0) * self.upper_bound * (1 - hist)) + 1
return hist
def forward(self, inputs, targets):
target_cpu = targets.data.cpu().numpy()
if self.batch_weights:
weights = self.calculate_weights(target_cpu)
self.nll_loss.weight = torch.Tensor(weights).cuda()
loss = 0.0
for i in range(0, inputs.shape[0]):
if not self.batch_weights:
weights = self.calculate_weights(target_cpu[i])
self.nll_loss.weight = torch.Tensor(weights).cuda()
loss += self.nll_loss(F.log_softmax(inputs[i].unsqueeze(0)),
targets[i].unsqueeze(0))
return loss
class CrossEntropyLoss2d(nn.Module):
"""
Cross Entroply NLL Loss
"""
def __init__(self, weight=None, size_average=True, ignore_index=255):
super(CrossEntropyLoss2d, self).__init__()
logging.info("Using Cross Entropy Loss")
self.nll_loss = nn.NLLLoss2d(weight, size_average, ignore_index)
# self.weight = weight
def forward(self, inputs, targets):
return self.nll_loss(F.log_softmax(inputs), targets)
def customsoftmax(inp, multihotmask):
"""
Custom Softmax
"""
soft = F.softmax(inp)
# This takes the mask * softmax ( sums it up hence summing up the classes in border
# then takes of summed up version vs no summed version
return torch.log(
torch.max(soft, (multihotmask * (soft * multihotmask).sum(1, keepdim=True)))
)
class ImgWtLossSoftNLL(nn.Module):
"""
Relax Loss
"""
def __init__(self, classes, ignore_index=255, weights=None, upper_bound=1.0,
norm=False, ohem=False):
super(ImgWtLossSoftNLL, self).__init__()
self.weights = weights
self.num_classes = classes
self.ignore_index = ignore_index
self.upper_bound = upper_bound
self.norm = norm
self.batch_weights = cfg.BATCH_WEIGHTING
self.fp16 = False
self.ohem = ohem
self.ohem_loss = OhemCrossEntropy2dTensor(self.ignore_index).cuda()
def calculate_weights(self, target):
"""
Calculate weights of the classes based on training crop
"""
if len(target.shape) == 3:
hist = np.sum(target, axis=(1, 2)) * 1.0 / target.sum()
else:
hist = np.sum(target, axis=(0, 2, 3)) * 1.0 / target.sum()
if self.norm:
hist = ((hist != 0) * self.upper_bound * (1 / hist)) + 1
else:
hist = ((hist != 0) * self.upper_bound * (1 - hist)) + 1
return hist[:-1]
def onehot2label(self, target):
# a bug here
label = torch.argmax(target[:, :-1, :, :], dim=1).long()
label[target[:, -1, :, :]] = self.ignore_index
return label
def custom_nll(self, inputs, target, class_weights, border_weights, mask):
"""
NLL Relaxed Loss Implementation
"""
if (cfg.REDUCE_BORDER_EPOCH != -1 and cfg.EPOCH > cfg.REDUCE_BORDER_EPOCH):
if self.ohem:
return self.ohem_loss(inputs, self.onehot2label(target))
border_weights = 1 / border_weights
target[target > 1] = 1
if self.fp16:
loss_matrix = (-1 / border_weights *
(target[:, :-1, :, :].half() *
class_weights.unsqueeze(0).unsqueeze(2).unsqueeze(3) *
customsoftmax(inputs, target[:, :-1, :, :].half())).sum(1)) * \
(1. - mask.half())
else:
loss_matrix = (-1 / border_weights *
(target[:, :-1, :, :].float() *
class_weights.unsqueeze(0).unsqueeze(2).unsqueeze(3) *
customsoftmax(inputs, target[:, :-1, :, :].float())).sum(1)) * \
(1. - mask.float())
# loss_matrix[border_weights > 1] = 0
loss = loss_matrix.sum()
# +1 to prevent division by 0
loss = loss / (target.shape[0] * target.shape[2] * target.shape[3] - mask.sum().item() + 1)
return loss
def forward(self, inputs, target):
# add ohem loss for the final stage
if (cfg.REDUCE_BORDER_EPOCH != -1 and cfg.EPOCH > cfg.REDUCE_BORDER_EPOCH) and self.ohem:
return self.ohem_loss(inputs, self.onehot2label(target[:,:-1,:,:]))
if self.fp16:
weights = target[:, :-1, :, :].sum(1).half()
else:
weights = target[:, :-1, :, :].sum(1).float()
ignore_mask = (weights == 0)
weights[ignore_mask] = 1
loss = 0
target_cpu = target.data.cpu().numpy()
if self.batch_weights:
class_weights = self.calculate_weights(target_cpu)
for i in range(0, inputs.shape[0]):
if not self.batch_weights:
class_weights = self.calculate_weights(target_cpu[i])
loss = loss + self.custom_nll(inputs[i].unsqueeze(0),
target[i].unsqueeze(0),
class_weights=torch.Tensor(class_weights).cuda(),
border_weights=weights, mask=ignore_mask[i])
return loss
class OhemWithAux(nn.Module):
def __init__(self, ignore_index=255, thresh=0.7, min_kept=10000, aux_weight=0.4):
super(OhemWithAux, self).__init__()
self.ignore_index = ignore_index
self.thresh = float(thresh)
self.min_kept = int(min_kept)
self.aux_weight = aux_weight
self.main_loss = OhemCrossEntropy2dTensor(ignore_index, thresh, min_kept)
self.aux_loss = OhemCrossEntropy2dTensor(ignore_index, thresh, min_kept)
def forward(self, pred, target):
x_main, x_aux = pred
return self.main_loss(x_main, target) + self.aux_weight * self.aux_loss(x_aux, target)
class OhemCrossEntropy2dTensor(nn.Module):
"""
Ohem Cross Entropy Tensor Version
"""
def __init__(self, ignore_index=255, thresh=0.7, min_kept=10000,
use_weight=False):
super(OhemCrossEntropy2dTensor, self).__init__()
self.ignore_index = ignore_index
self.thresh = float(thresh)
self.min_kept = int(min_kept)
if use_weight:
weight = torch.FloatTensor(
[0.8373, 0.918, 0.866, 1.0345, 1.0166, 0.9969, 0.9754, 1.0489,
0.8786, 1.0023, 0.9539, 0.9843, 1.1116, 0.9037, 1.0865, 1.0955,
1.0865, 1.1529, 1.0507])
self.criterion = torch.nn.CrossEntropyLoss(reduction="elementwise_mean",
weight=weight,
ignore_index=ignore_index)
else:
self.criterion = torch.nn.CrossEntropyLoss(reduction="elementwise_mean",
ignore_index=ignore_index)
def forward(self, pred, target):
b, c, h, w = pred.size()
target = target.view(-1)
valid_mask = target.ne(self.ignore_index)
target = target * valid_mask.long()
num_valid = valid_mask.sum()
prob = F.softmax(pred, dim=1)
prob = (prob.transpose(0, 1)).reshape(c, -1)
if self.min_kept > num_valid:
print('Labels: {}'.format(num_valid))
elif num_valid > 0:
prob = prob.masked_fill_(~valid_mask, 1)
mask_prob = prob[
target, torch.arange(len(target), dtype=torch.long)]
threshold = self.thresh
if self.min_kept > 0:
_, index = mask_prob.sort()
threshold_index = index[min(len(index), self.min_kept) - 1]
if mask_prob[threshold_index] > self.thresh:
threshold = mask_prob[threshold_index]
kept_mask = mask_prob.le(threshold)
target = target * kept_mask.long()
valid_mask = valid_mask * kept_mask
target = target.masked_fill_(~valid_mask, self.ignore_index)
target = target.view(b, h, w)
return self.criterion(pred, target)
class CriterionSFNet(nn.Module):
def __init__(self, aux_weight=1.0, thresh=0.7, min_kept=100000, ignore_index=255):
super(CriterionSFNet, self).__init__()
self._aux_weight = aux_weight
self._criterion1 = OhemCrossEntropy2dTensor(ignore_index, thresh, min_kept)
def forward(self, preds, target):
h, w = target.size(1), target.size(2)
main_pred, aux_preds = preds
main_pred = F.upsample(input=main_pred, size=(h, w), mode='bilinear', align_corners=True)
main_loss = self._criterion1(main_pred, target)
for aux_p in aux_preds:
aux_p = F.upsample(input=aux_p, size=(h, w), mode='bilinear', align_corners=True)
main_loss += self._aux_weight * self._criterion1(aux_p, target)
return main_loss
class JointEdgeSegLoss(nn.Module):
def __init__(self, classes, ignore_index=255,mode='train',
edge_weight=1, seg_weight=1, seg_body_weight=1, att_weight=1):
super(JointEdgeSegLoss, self).__init__()
self.num_classes = classes
if mode == 'train':
self.seg_loss = OhemCrossEntropy2dTensor(ignore_index=ignore_index).cuda()
elif mode == 'val':
self.seg_loss = CrossEntropyLoss2d(size_average=True,
ignore_index=ignore_index).cuda()
self.seg_body_loss = ImgWtLossSoftNLL(classes=classes,
ignore_index=ignore_index,
upper_bound=1.0, ohem=False).cuda()
self.edge_ohem_loss = OhemCrossEntropy2dTensor(ignore_index=ignore_index, min_kept=5000).cuda()
self.ignore_index = ignore_index
self.edge_weight = edge_weight
self.seg_weight = seg_weight
self.att_weight = att_weight
self.seg_body_weight = seg_body_weight
def bce2d(self, input, target):
n, c, h, w = input.size()
log_p = input.transpose(1, 2).transpose(2, 3).contiguous().view(1, -1)
target_t = target.transpose(1, 2).transpose(2, 3).contiguous().view(1, -1)
target_trans = target_t.clone()
pos_index = (target_t == 1)
neg_index = (target_t == 0)
ignore_index = (target_t > 1)
target_trans[pos_index] = 1
target_trans[neg_index] = 0
pos_index = pos_index.data.cpu().numpy().astype(bool)
neg_index = neg_index.data.cpu().numpy().astype(bool)
ignore_index = ignore_index.data.cpu().numpy().astype(bool)
weight = torch.Tensor(log_p.size()).fill_(0)
weight = weight.numpy()
pos_num = pos_index.sum()
neg_num = neg_index.sum()
sum_num = pos_num + neg_num
weight[pos_index] = neg_num * 1.0 / sum_num
weight[neg_index] = pos_num * 1.0 / sum_num
weight[ignore_index] = 0
weight = torch.from_numpy(weight).cuda()
log_p = log_p.cuda()
target_t = target_t.cuda()
loss = F.binary_cross_entropy_with_logits(log_p, target_t, weight, size_average=True)
return loss
def edge_attention(self, input, target, edge):
filler = torch.ones_like(target) * 255
return self.edge_ohem_loss(input, torch.where(edge.max(1)[0] > 0.8, target, filler))
def forward(self, inputs, targets):
seg_in, seg_body_in, edge_in = inputs
seg_bord_mask, edgemask = targets
segmask = self.onehot2label(seg_bord_mask)
losses = {}
losses['seg_loss'] = self.seg_weight * self.seg_loss(seg_in, segmask)
losses['seg_body'] = self.seg_body_weight * self.seg_body_loss(seg_body_in, seg_bord_mask)
losses['edge_loss'] = self.edge_weight * 20 * self.bce2d(edge_in, edgemask)
losses['edge_ohem_loss'] = self.att_weight * self.edge_attention(seg_in, segmask, edge_in)
return losses
def onehot2label(self, target):
"""
Args:
target:
Returns:
"""
label = torch.argmax(target[:, :-1, :, :], dim=1).long()
label[target[:, -1, :, :]] = self.ignore_index
return label