-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathmodel.py
291 lines (228 loc) · 8.7 KB
/
model.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
class PositionEmbs(nn.Module):
def __init__(self, num_patches, emb_dim, dropout_rate=0.1):
super(PositionEmbs, self).__init__()
self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, emb_dim))
if dropout_rate > 0:
self.dropout = nn.Dropout(dropout_rate)
else:
self.dropout = None
def forward(self, x):
out = x + self.pos_embedding
if self.dropout:
out = self.dropout(out)
return out
class MlpBlock(nn.Module):
""" Transformer Feed-Forward Block """
def __init__(self, in_dim, mlp_dim, out_dim, dropout_rate=0.1):
super(MlpBlock, self).__init__()
# init layers
self.fc1 = nn.Linear(in_dim, mlp_dim)
self.fc2 = nn.Linear(mlp_dim, out_dim)
self.act = nn.GELU()
if dropout_rate > 0.0:
self.dropout1 = nn.Dropout(dropout_rate)
self.dropout2 = nn.Dropout(dropout_rate)
else:
self.dropout1 = None
self.dropout2 = None
def forward(self, x):
out = self.fc1(x)
out = self.act(out)
if self.dropout1:
out = self.dropout1(out)
out = self.fc2(out)
out = self.dropout2(out)
return out
class LinearGeneral(nn.Module):
def __init__(self, in_dim=(768,), feat_dim=(12, 64)):
super(LinearGeneral, self).__init__()
self.weight = nn.Parameter(torch.randn(*in_dim, *feat_dim))
self.bias = nn.Parameter(torch.zeros(*feat_dim))
def forward(self, x, dims):
a = torch.tensordot(x, self.weight, dims=dims) + self.bias
return a
class SelfAttention(nn.Module):
def __init__(self, in_dim, heads=8, dropout_rate=0.1):
super(SelfAttention, self).__init__()
self.heads = heads
self.head_dim = in_dim // heads
self.scale = self.head_dim ** 0.5
self.query = LinearGeneral((in_dim,), (self.heads, self.head_dim))
self.key = LinearGeneral((in_dim,), (self.heads, self.head_dim))
self.value = LinearGeneral((in_dim,), (self.heads, self.head_dim))
self.out = LinearGeneral((self.heads, self.head_dim), (in_dim,))
if dropout_rate > 0:
self.dropout = nn.Dropout(dropout_rate)
else:
self.dropout = None
def forward(self, x):
b, n, _ = x.shape
q = self.query(x, dims=([2], [0]))
k = self.key(x, dims=([2], [0]))
v = self.value(x, dims=([2], [0]))
q = q.permute(0, 2, 1, 3)
k = k.permute(0, 2, 1, 3)
v = v.permute(0, 2, 1, 3)
attn_weights = torch.matmul(q, k.transpose(-2, -1)) / self.scale
attn_weights = F.softmax(attn_weights, dim=-1)
out = torch.matmul(attn_weights, v)
out = out.permute(0, 2, 1, 3)
out = self.out(out, dims=([2, 3], [0, 1]))
return out
class EncoderBlock(nn.Module):
def __init__(self, in_dim, mlp_dim, num_heads, dropout_rate=0.1, attn_dropout_rate=0.1):
super(EncoderBlock, self).__init__()
self.norm1 = nn.LayerNorm(in_dim)
self.attn = SelfAttention(in_dim, heads=num_heads, dropout_rate=dropout_rate)
if dropout_rate > 0:
self.dropout = nn.Dropout(dropout_rate)
else:
self.dropout = None
self.norm2 = nn.LayerNorm(in_dim)
self.mlp = MlpBlock(in_dim, mlp_dim, in_dim, dropout_rate)
def forward(self, x):
residual = x
out = self.norm1(x)
out = self.attn(out)
if self.dropout:
out = self.dropout(out)
out += residual
residual = out
out = self.norm2(out)
out = self.mlp(out)
out += residual
return out
class Encoder(nn.Module):
def __init__(self, num_patches, emb_dim, mlp_dim, num_layers=12, num_heads=12, dropout_rate=0.1, attn_dropout_rate=0.0):
super(Encoder, self).__init__()
# positional embedding
self.pos_embedding = PositionEmbs(num_patches, emb_dim, dropout_rate)
# encoder blocks
in_dim = emb_dim
self.encoder_layers = nn.ModuleList()
for i in range(num_layers):
layer = EncoderBlock(in_dim, mlp_dim, num_heads, dropout_rate, attn_dropout_rate)
self.encoder_layers.append(layer)
self.norm = nn.LayerNorm(in_dim)
def forward(self, x):
out = self.pos_embedding(x)
for layer in self.encoder_layers:
out = layer(out)
out = self.norm(out)
return out
class VisionTransformer(nn.Module):
""" Vision Transformer """
def __init__(self,
image_size=(256, 256),
patch_size=(16, 16),
emb_dim=768,
mlp_dim=3072,
num_heads=12,
num_layers=12,
num_classes=1000,
attn_dropout_rate=0.0,
dropout_rate=0.1):
super(VisionTransformer, self).__init__()
h, w = image_size
# embedding layer
fh, fw = patch_size
gh, gw = h // fh, w // fw
num_patches = gh * gw
self.embedding = nn.Conv2d(3, emb_dim, kernel_size=(fh, fw), stride=(fh, fw))
# class token
self.cls_token = nn.Parameter(torch.zeros(1, 1, emb_dim))
# transformer
self.transformer = Encoder(
num_patches=num_patches,
emb_dim=emb_dim,
mlp_dim=mlp_dim,
num_layers=num_layers,
num_heads=num_heads,
dropout_rate=dropout_rate,
attn_dropout_rate=attn_dropout_rate)
self.classifier = nn.Linear(emb_dim, num_classes)
def _init_weights(self, m):
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def forward(self, x, feat_cls=False):
emb = self.embedding(x) # (n, c, gh, gw)
emb = emb.permute(0, 2, 3, 1) # (n, gh, hw, c)
b, h, w, c = emb.shape
emb = emb.reshape(b, h * w, c)
# prepend class token
cls_token = self.cls_token.repeat(b, 1, 1)
emb = torch.cat([cls_token, emb], dim=1)
# transformer
feat = self.transformer(emb)
# classifier
if feat_cls:
return feat[:, 0], self.classifier(feat[:, 0])
else:
return self.classifier(feat[:, 0])
class OODTransformer(nn.Module):
def __init__(self,
image_size=(256, 256),
patch_size=(16, 16),
emb_dim=768,
mlp_dim=3072,
num_heads=12,
num_layers=12,
num_classes=1000,
attn_dropout_rate=0.0,
dropout_rate=0.1):
super(OODTransformer, self).__init__()
h, w = image_size
# embedding layer
fh, fw = patch_size
gh, gw = h // fh, w // fw
num_patches = gh * gw
self.embedding = nn.Conv2d(3, emb_dim, kernel_size=(fh, fw), stride=(fh, fw))
# class token
self.cls_token = nn.Parameter(torch.zeros(1, 1, emb_dim))
# transformer
self.transformer = Encoder(
num_patches=num_patches,
emb_dim=emb_dim,
mlp_dim=mlp_dim,
num_layers=num_layers,
num_heads=num_heads,
dropout_rate=dropout_rate,
attn_dropout_rate=attn_dropout_rate)
self.classifier = nn.Linear(emb_dim, num_classes)
#self.apply(self._init_weights)
self.svdd = nn.Linear(emb_dim, emb_dim)
def _init_weights(self, m):
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def forward(self, x, feat_cls=False):
emb = self.embedding(x) # (n, c, gh, gw)
emb = emb.permute(0, 2, 3, 1) # (n, gh, hw, c)
b, h, w, c = emb.shape
emb = emb.reshape(b, h * w, c)
# prepend class token
cls_token = self.cls_token.repeat(b, 1, 1)
emb = torch.cat([cls_token, emb], dim=1)
# transformer
feat = self.transformer(emb)
# classifier
if feat_cls:
return self.svdd(feat[:, 0]), self.classifier(feat[:, 0])
else:
return self.svdd(feat[:, 0])
if __name__ == '__main__':
model = VisionTransformer(num_layers=2)
x = torch.randn((2, 3, 256, 256))
out = model(x)
state_dict = model.state_dict()
for key, value in state_dict.items():
print("{}: {}".format(key, value.shape))