-
Notifications
You must be signed in to change notification settings - Fork 10
/
senformer.py
359 lines (313 loc) · 16.9 KB
/
senformer.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from mmcv.runner import force_fp32
from mmseg.ops import resize
from mmseg.models.builder import HEADS
from mmseg.models.decode_heads.decode_head import BaseDecodeHead
from mmseg.models.losses import accuracy
from timm.models.vision_transformer import Attention#, Mlp
from timm.models.layers import DropPath
from einops import rearrange
class Mlp(nn.Module):
""" Multilayer perceptron from timm library."""
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
class CrossAttention(nn.Module):
'''
Taken from timm library Attention module
with slight modifications to do Cross-Attention.
'''
def __init__(self, dim, num_heads=8, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim ** -0.5
self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.to_q = nn.Linear(dim, dim, bias=qkv_bias)
def forward(self,q_in, kv_in):
B, N, C = kv_in.shape
_, L, _ = q_in.shape
# Create key and value tokens
kv = self.kv(kv_in).reshape(B, N, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
k, v = kv[0], kv[1]
# Create query tokens
q = self.to_q(q_in)
q = q.reshape(B, L, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
# Compute attention
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, L, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
#--- Decoder Block ---
# │
# ├────────┐
# ┌───────┴───────┐│
# │ MLP ││
# └───────┬───────┘│
# ├────────┘
# ├────────┐
# ┌───────┴───────┐│
# │Self-Attention ││
# └───────┬───────┘│
# ├────────┘
# ├────────┐
# ┌───────┴───────┐│
# │Cross-Attention││
# └───────┬───────┘│
# ├────────┘
# │
class DecoderBlock(nn.Module):
def __init__(self, dim, num_heads,mlp_ratio=4.,
qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.norm1 = norm_layer(dim)
self.norm2 = norm_layer(dim)
self.norm3 = norm_layer(dim)
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.drop_path3 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.attn = Attention(dim=dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=0.)
self.cross_attn = CrossAttention(dim=dim, num_heads=num_heads,qkv_bias=qkv_bias, qk_scale=qk_scale,
attn_drop=attn_drop, proj_drop=0.)
self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), out_features=dim, act_layer=act_layer, drop=drop)
def forward_crossattn(self, queries, features):
##--------- Cross-Attention Block
out = queries + self.drop_path1(self.cross_attn(self.norm1(queries), features))
return out
def forward_attn(self, q):
##--------- Self-Attention Block
q = q + self.drop_path2(self.attn(self.norm2(q)))
return q
def forward_mlp(self, q):
##--------- MLP Block
cls_features = q + self.drop_path3(self.mlp(self.norm3(q)))
return cls_features
def forward(self, queries, features):
out = self.forward_crossattn(queries, features)
out = self.forward_attn(out)
out = self.forward_mlp(out)
return out
class TransformerLearner(nn.Module):
def __init__(self, dim, num_heads, num_queries, branch_depth, mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0.,
attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super(TransformerLearner, self).__init__()
dpr = [x.item() for x in torch.linspace(0, drop_path, branch_depth)]
self.layers = nn.ModuleList([
DecoderBlock(dim, num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop,
attn_drop=attn_drop, drop_path=dpr[i], act_layer=act_layer, norm_layer=norm_layer)
for i in range(branch_depth)
])
# Learnable class embeddings
self.queries = nn.Parameter(torch.zeros(1, num_queries, dim), requires_grad=True)
# Norm features and class embeddings for training stability
self.norm_features = nn.LayerNorm(dim)
self.norm_embs = nn.LayerNorm(dim)
def forward(self, features):
B, _, H, W = features.shape
# Tokenize 2D feature maps
features = rearrange(features, 'b c h w -> b (h w) c')
features = self.norm_features(features)
# Expand to batch size
cls_embs = self.queries.expand(B, -1, -1)
# Decoder
for layer in self.layers:
cls_embs = layer(cls_embs, features)
# Norm class embeddings for stability
cls_embs = self.norm_embs(cls_embs)
# Prediction
pred = (features @ cls_embs.transpose(-2, -1))
# Reshape into 2D maps
pred = rearrange(pred, 'b (h w) c -> b c h w', h=H, w=W)
return pred
# ┌──────────────────┐
# ┌────────┼───────┐┌─────────┼──────┐
# │ ┌┐├┐ ││ ┌┐▼┐ │ ┌─────────┐
# │ ││││ ││ │││├─────┼──▶ Learner ────┐
# │ ││││ ││ ││││ │ └─────────┘ │
# │ └┘▲┘ ││ └┘├┘ │ │
# │ │ ──────┼┼────────▶│ │ │
# │ ┌┐┌┐├┐┌┐ ││ ┌┐┌┐▼┐┌┐ │ ┌─────────┐ │
# │ ││││││││ ││ │││││││├───┼──▶ Learner ──┐ │
# │ ││││││││ ││ ││││││││ │ └─────────┘ │ │
# │ └┘└┘▲┘└┘ ││ └┘└┘├┘└┘ │ ┌▼─▼┐ ┌─────┐
# │ │ ──────┼┼────────▶│ │ │ + ├─▶Pred │
# │ ┌┐┌┐┌┐├┐┌┐┌┐ ││ ┌┐┌┐┌┐▼┐┌┐┌┐ │ └▲─▲┘ └─────┘
# │ ││││││││││││ ││ ││││││││││││ │ ┌─────────┐ │ │
# │ ││││││││││││ ││ │││││││││││├─┼──▶ Learner ──┘ │
# │ └┘└┘└┘▲┘└┘└┘ ││ └┘└┘└┘├┘└┘└┘ │ └─────────┘ │
# │ │ ──────┼┼────────▶│ │ │
# │┌┐┌┐┌┐┌┐├┐┌┐┌┐┌┐││┌┐┌┐┌┐┌┐┌▼┌┐┌┐┌┐│ │
# ││││││││││││││││││││││││││││││││││││ ┌─────────┐ │
# ││││││││││││││││││││││││││││││││││├┼──┼▶Learner ────┘
# │└┘└┘└┘└┘└┘└┘└┘└┘││└┘└┘└┘└┘└┘└┘└┘└┘│ └─────────┘
# └───┬────────┬───┘└───┬────────┬───┘
# │Backbone│ │ FPNT │
# └────────┘ └────────┘
# ╔════════════════════════════════════════════════════════════╗
# ║ ____ ____ _ _ ____ ____ ____ _ _ ____ ____ ║
# ║ [__ |___ |\ | |___ | | |__/ |\/| |___ |__/ ║
# ║ ___] |___ | \| | |__| | \ | | |___ | \ ║
# ║ ║
# ╚════════════════════════════════════════════════════════════╝
@HEADS.register_module()
class SenFormer(BaseDecodeHead):
def __init__(self, feature_strides, num_heads, branch_depth, mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0.,
attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, eps=1.e-15, **kwargs):
super(SenFormer, self).__init__(
input_transform='multiple_select', **kwargs)
assert len(feature_strides) == len(self.in_channels)
assert min(feature_strides) == feature_strides[0]
self.feature_strides = feature_strides
self.eps = eps
self.learners = nn.ModuleList()
for i in range(len(feature_strides)):
self.learners.append(
TransformerLearner(dim=self.in_channels[i],
num_heads=num_heads,
num_queries=self.num_classes,
branch_depth=branch_depth,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path,
act_layer=act_layer,
norm_layer=norm_layer)
)
def forward(self, inputs):
x = self._transform_inputs(inputs)
# -------
prob_outputs = [] # save probabilty maps for ensemble prediction
logit_outputs = [] # save logits for learners' supervision
for i in range(0, len(self.feature_strides)):
# learner prediction
learner_pred = self.learners[i](x[i])
learner_pred = resize(learner_pred, size=x[0].shape[2:], mode='bilinear', align_corners=self.align_corners)
# -------
logit_outputs.append(learner_pred)
prob_outputs.append(F.softmax(learner_pred, dim=1))
# Ensemble prediction
ensemble_pred = torch.stack(prob_outputs, dim=0).sum(dim=0)
return logit_outputs, ensemble_pred
def forward_test(self, inputs, img_metas, test_cfg):
_, ensemble_pred = self.forward(inputs)
return ensemble_pred
@force_fp32(apply_to=('seg_logit',))
def losses(self, seg_logit, seg_label):
"""Compute segmentation loss."""
logit_outputs, ensemble_pred = seg_logit # unpack
# Upscale outputs to the ground truth size
## Ensemble predicition
ensemble_pred = torch.log(torch.clamp(ensemble_pred, min=self.eps))
ensemble_pred = resize(input=ensemble_pred, size=seg_label.shape[2:], mode='bilinear', align_corners=self.align_corners)
## Learners predicitions
seg_log_logit = [resize(input=logit_outputs[i], size=seg_label.shape[2:], mode='bilinear', align_corners=self.align_corners)
for i in range(len(logit_outputs))]
seg_label = seg_label.squeeze(1)
# Losses
loss = dict()
## Loss for the ensemble prediction
loss_classic = F.cross_entropy(ensemble_pred, seg_label, ignore_index=self.ignore_index)
## Loss for each learner
losses_seg = [F.cross_entropy(seg_log_logit[i], seg_label, ignore_index=self.ignore_index) for i in
range(len(seg_log_logit))]
loss_extra = torch.stack(losses_seg, dim=0).sum()
loss_extra = loss_extra / len(seg_log_logit)
loss['loss_classic'] = loss_classic
loss['loss_seg'] = loss_extra
loss['acc_seg'] = accuracy(ensemble_pred, seg_label)
return loss
@HEADS.register_module()
class SenFormerNWS(BaseDecodeHead):
'''
SenFormer with No Weight Sharing (NWS)
'''
def __init__(self, feature_strides, num_heads, branch_depth, mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0.,
attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, eps=1.e-15, **kwargs):
super(SenFormer, self).__init__(
input_transform='multiple_select', **kwargs)
assert len(feature_strides) == len(self.in_channels)
assert min(feature_strides) == feature_strides[0]
self.feature_strides = feature_strides
self.eps = eps
self.learners = nn.ModuleList()
for i in range(len(feature_strides)):
self.learners.append(
TransformerLearner(dim=self.in_channels[i],
num_heads=num_heads,
num_queries=self.num_classes,
branch_depth=branch_depth,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path,
act_layer=act_layer,
norm_layer=norm_layer)
)
def forward(self, inputs):
x = self._transform_inputs(inputs)
# -------
prob_outputs = [] # save probabilty maps for ensemble prediction
logit_outputs = [] # save logits for learners' supervision
for i in range(0, len(self.feature_strides)):
# learner prediction
learner_pred = self.learners[i](x[i])
learner_pred = resize(learner_pred, size=x[0].shape[2:], mode='bilinear', align_corners=self.align_corners)
# -------
logit_outputs.append(learner_pred)
prob_outputs.append(F.softmax(learner_pred, dim=1))
# Ensemble prediction
ensemble_pred = torch.stack(prob_outputs, dim=0).sum(dim=0)
return logit_outputs, ensemble_pred
def forward_test(self, inputs, img_metas, test_cfg):
_, ensemble_pred = self.forward(inputs)
return ensemble_pred
@force_fp32(apply_to=('seg_logit',))
def losses(self, seg_logit, seg_label):
"""Compute segmentation loss."""
logit_outputs, ensemble_pred = seg_logit # unpack
# Upscale outputs to the ground truth size
## Ensemble predicition
ensemble_pred = torch.log(torch.clamp(ensemble_pred, min=self.eps))
ensemble_pred = resize(input=ensemble_pred, size=seg_label.shape[2:], mode='bilinear', align_corners=self.align_corners)
## Learners predicitions
seg_log_logit = [resize(input=logit_outputs[i], size=seg_label.shape[2:], mode='bilinear', align_corners=self.align_corners)
for i in range(len(logit_outputs))]
seg_label = seg_label.squeeze(1)
# Losses
loss = dict()
## Loss for the ensemble prediction
loss_classic = F.cross_entropy(ensemble_pred, seg_label, ignore_index=self.ignore_index)
## Loss for each learner
losses_seg = [F.cross_entropy(seg_log_logit[i], seg_label, ignore_index=self.ignore_index) for i in
range(len(seg_log_logit))]
loss_extra = torch.stack(losses_seg, dim=0).sum()
loss_extra = loss_extra / len(seg_log_logit)
loss['loss_classic'] = loss_classic
loss['loss_seg'] = loss_extra
loss['acc_seg'] = accuracy(ensemble_pred, seg_label)
return loss