forked from selinakhan/stylistic-MTL-ukiyoe
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
254 lines (195 loc) · 9.31 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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models as models
import numpy as np
from models.configs import get_b16_config, get_b32_config, get_testing
from models.modeling import VisionTransformer
vit = np.load('models/ViT-B_16.npz')
res = models.resnet50(pretrained=True)
''' Multi-task models for hard and soft sharing parameters. '''
class MultiTaskHS(nn.Module):
def __init__(self, img_size, hidden_features, class1, class2, model_config):
super(MultiTaskHS, self).__init__()
# ViT UNfrozen config
if model_config == 'vit':
self.backbone = VisionTransformer(get_b16_config(), img_size=img_size, vis=True)
self.backbone.load_from(vit)
del self.backbone.head
ct = 0
for child in self.backbone.transformer.encoder.layer.children():
ct += 1
if ct <= 6:
for param in child.parameters():
param.requires_grad = False
self.linear = nn.Linear(768, 768)
self.linear2 = nn.Linear(768, hidden_features)
# Vit frozen config
if model_config == 'fr_vit':
self.backbone = VisionTransformer(get_b16_config(), img_size=img_size, vis=True)
self.backbone.load_from(vit)
del self.backbone.head
for param in self.backbone.parameters():
param.requires_grad = False
self.linear = nn.Linear(768, 768)
self.linear2 = nn.Linear(768, hidden_features)
# Frozen ResNet50 config
if model_config == 'fr_res':
print('frozen resnet training')
self.backbone = res
for param in self.backbone.parameters():
param.requires_grad = False
self.linear = nn.Linear(1000, 768)
self.linear2 = nn.Linear(768, hidden_features)
# UNfrozen ResNet50 config
if model_config == 'res':
self.backbone = res
ct = 0
for child in self.backbone.children():
ct += 1
if ct < 6:
for param in child.parameters():
param.requires_grad = False
self.linear = nn.Linear(1000, 768)
self.linear2 = nn.Linear(768, hidden_features)
self.artist_net = nn.Sequential(nn.Linear(hidden_features, hidden_features // 2),
nn.ReLU(), nn.Linear(hidden_features // 2, class1))
self.date_net = nn.Sequential(nn.Linear(hidden_features, hidden_features // 4),
nn.ReLU(), nn.Linear(hidden_features // 4, 1))
self.era_net = nn.Sequential(nn.Linear(hidden_features, hidden_features // 4),
nn.ReLU(), nn.Linear(hidden_features // 4, class2))
def forward(self, x, model_config):
# ViT forward pass
if model_config == 'fr_vit' or model_config == 'vit':
x, attn = self.backbone.transformer(x)
x = F.relu(x[:, 0])
x = self.linear(x)
x = F.relu(x)
x = self.linear2(x)
x = F.relu(x)
return self.artist_net(x), self.date_net(x), self.era_net(x), attn
# ResNet50 forward pass
if model_config == 'fr_res' or model_config == 'res':
x = self.backbone(x)
x = self.linear(x)
x = F.relu(x)
x = self.linear2(x)
x = F.relu(x)
return self.artist_net(x), self.date_net(x), self.era_net(x), 0
''' Single-task models '''
class SingleTaskClassification(nn.Module):
def __init__(self, img_size, hidden_features, class1, divisor, model_config):
super(SingleTaskClassification, self).__init__()
# ViT UNfrozen config
if model_config == 'vit':
self.backbone = VisionTransformer(get_b16_config(), img_size=img_size, vis=True)
self.backbone.load_from(vit)
del self.backbone.head
ct = 0
for child in self.backbone.transformer.encoder.layer.children():
ct += 1
if ct <= 6:
for param in child.parameters():
param.requires_grad = False
self.net = nn.Sequential(nn.Linear(768, hidden_features // 2),
nn.ReLU(), nn.Linear(hidden_features // 2, hidden_features // divisor),
nn.ReLU(), nn.Linear(hidden_features // divisor, class1))
# Vit frozen config
if model_config == 'fr_vit':
self.backbone = VisionTransformer(get_b16_config(), img_size=img_size, vis=True)
self.backbone.load_from(vit)
del self.backbone.head
for param in self.backbone.parameters():
param.requires_grad = False
self.net = nn.Sequential(nn.Linear(768, hidden_features // 2),
nn.ReLU(), nn.Linear(hidden_features // 2, hidden_features // divisor),
nn.ReLU(), nn.Linear(hidden_features // divisor, class1))
# Frozen ResNet50 config
if model_config == 'fr_res':
self.backbone = res
for param in self.backbone.parameters():
param.requires_grad = False
self.net = nn.Sequential(nn.Linear(1000, hidden_features // 2),
nn.ReLU(), nn.Linear(hidden_features // 2, hidden_features // divisor),
nn.ReLU(), nn.Linear(hidden_features // divisor, class1))
# UNfrozen ResNet50 config
if model_config == 'res':
self.backbone = res
ct = 0
for child in self.backbone.children():
ct += 1
if ct < 6:
for param in child.parameters():
param.requires_grad = False
self.net = nn.Sequential(nn.Linear(1000, hidden_features // 2),
nn.ReLU(), nn.Linear(hidden_features // 2, hidden_features // divisor),
nn.ReLU(), nn.Linear(hidden_features // divisor, class1))
def forward(self, x, model_config):
# ViT forward pass
if model_config == 'fr_vit' or model_config == 'vit':
x, attn = self.backbone.transformer(x)
x = F.relu(x[:, 0])
return self.net(x), attn
# ResNet50 forward pass
if model_config == 'fr_res' or model_config == 'res':
x = self.backbone(x)
x = F.relu(x)
return self.net(x), 0
class SingleTaskRegression(nn.Module):
def __init__(self, img_size, hidden_features, model_config):
super(SingleTaskRegression, self).__init__()
# ViT UNfrozen config
if model_config == 'vit':
self.backbone = VisionTransformer(get_b16_config(), img_size=img_size, vis=True)
self.backbone.load_from(vit)
del self.backbone.head
ct = 0
for child in self.backbone.transformer.encoder.layer.children():
ct += 1
if ct <= 6:
for param in child.parameters():
param.requires_grad = False
self.net = nn.Sequential(nn.Linear(768, hidden_features // 2),
nn.ReLU(), nn.Linear(hidden_features // 2, hidden_features // 4),
nn.ReLU(), nn.Linear(hidden_features // 4, 1))
# Vit frozen config
if model_config == 'fr_vit':
self.backbone = VisionTransformer(get_b16_config(), img_size=img_size, vis=True)
self.backbone.load_from(vit)
del self.backbone.head
for param in self.backbone.parameters():
param.requires_grad = False
self.net = nn.Sequential(nn.Linear(768, hidden_features // 2),
nn.ReLU(), nn.Linear(hidden_features // 2, hidden_features // 4),
nn.ReLU(), nn.Linear(hidden_features // 4, 1))
# Frozen ResNet50 config
if model_config == 'fr_res':
self.backbone = res
for param in self.backbone.parameters():
param.requires_grad = False
self.net = nn.Sequential(nn.Linear(1000, hidden_features // 2),
nn.ReLU(), nn.Linear(hidden_features // 2, hidden_features // 4),
nn.ReLU(), nn.Linear(hidden_features // 4, 1))
# UNfrozen ResNet50 config
if model_config == 'res':
self.backbone = res
ct = 0
for child in self.backbone.children():
ct += 1
if ct < 6:
for param in child.parameters():
param.requires_grad = False
self.net = nn.Sequential(nn.Linear(1000, hidden_features // 2),
nn.ReLU(), nn.Linear(hidden_features // 2, hidden_features // 4),
nn.ReLU(), nn.Linear(hidden_features // 4, 1))
def forward(self, x, model_config):
# ViT forward pass
if model_config == 'fr_vit' or model_config == 'vit':
x, attn = self.backbone.transformer(x)
x = F.relu(x[:, 0])
return self.net(x), attn
# ResNet50 forward pass
if model_config == 'fr_res' or model_config == 'res':
x = self.backbone(x)
x = F.relu(x)
return self.net(x), 0