-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
186 lines (138 loc) · 7.1 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
import numpy as np
import torch
import torch.nn as nn
from backbone import resnet, mae
from loss import MaeContrastiveLoss, CenterLoss
def simple_transform(x, beta):
x = 1/torch.pow(torch.log(1/x+1), beta)
return x
def extended_simple_transform(x, beta):
zero_tensor = torch.zeros_like(x)
x_pos = torch.maximum(x, zero_tensor)
x_neg = torch.minimum(x, zero_tensor)
x_pos = 1/torch.pow(torch.log(1/(x_pos+1e-5)+1), beta)
x_neg = -1/torch.pow(torch.log(1/(-x_neg+1e-5)+1), beta)
return x_pos + x_neg
class CreateModel(nn.Module):
def __init__(self, args_config, arch="resnet50", class_num=10):
super(CreateModel, self).__init__()
self.args_config = args_config
self.arch = arch
self.class_num = class_num
if self.arch == "resnet50":
self.pool_dim = 2048
if args_config.pretrained:
print("=> using pre-trained model '{}'".format(args_config.arch))
self.main_net = resnet.resnet50(pretrained=args_config.pretrained)
self.layer0 = nn.Sequential(self.main_net.conv1, self.main_net.bn1, self.main_net.relu, self.main_net.maxpool)
self.layer1 = self.main_net.layer1
self.layer2 = self.main_net.layer2
self.layer3 = self.main_net.layer3
self.layer4 = self.main_net.layer4
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.classifier = nn.Linear(self.pool_dim, self.class_num)
if self.arch == "mae":
# self.pool_dim = 768
# self.main_net = mae.mae_vit_base_patch16(norm_pix_loss=False)
self.pool_dim = 1024
self.main_net = mae.mae_vit_large_patch16(norm_pix_loss=False)
# self.pool_dim = 1280
# self.main_net = mae.mae_vit_huge_patch14(norm_pix_loss=False)
if args_config.pretrained:
print("=> using pre-trained model '{}'".format(args_config.arch))
checkpoint = torch.load(args_config.pretrained_model_path)
self.main_net.load_state_dict(checkpoint["model"])
# model_dict = self.main_net.state_dict()
# pretrained_dict = {k: v for k, v in checkpoint["model"].items() if k in model_dict}
# model_dict.update(pretrained_dict)
# self.main_net.load_state_dict(model_dict)
self.classifier = nn.Linear(self.pool_dim, self.class_num)
if self.arch == "resnet50 + mae":
self.pool_dim = 768 + 2048
self.main_resnet50 = resnet.resnet50(pretrained=args_config.pretrained)
self.main_mae = mae.mae_vit_base_patch16(norm_pix_loss=False)
if args_config.pretrained:
print("=> using pre-trained model '{}'".format(args_config.arch))
checkpoint = torch.load(args_config.pretrained_mae_path)
self.main_mae.load_state_dict(checkpoint["model"])
self.layer0 = nn.Sequential(self.main_resnet50.conv1, self.main_resnet50.bn1, self.main_resnet50.relu, self.main_resnet50.maxpool)
self.layer1 = self.main_resnet50.layer1
self.layer2 = self.main_resnet50.layer2
self.layer3 = self.main_resnet50.layer3
self.layer4 = self.main_resnet50.layer4
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.classifier = nn.Linear(self.pool_dim, self.class_num)
def forward(self, x, target=None, target_rand=None, lam=None):
if self.arch == "resnet50":
x_0 = self.layer0(x)
x_1 = self.layer1(x_0)
x_2 = self.layer2(x_1)
x_3 = self.layer3(x_2)
x_4 = self.layer4(x_3)
x_pool_4 = self.avgpool(x_4)
x_pool_4 = x_pool_4.view(x_pool_4.size(0), x_pool_4.size(1))
p_4 = self.classifier(x_pool_4)
if self.training:
ce_loss = self.cal_loss(p_4, target, latent=None, target_rand=target_rand, lam=lam)
return p_4, ce_loss
else:
return p_4
if self.arch == "mae":
if self.training:
mae_loss, latent, pred, imgs, imgs_mask = self.main_net(x, mask_ratio=0.5, iscutmix=self.args_config.cutmix)
p = self.classifier(latent[:, 0, :])
ce_loss, mnce_loss = self.cal_loss(p, target, latent=latent, target_rand=target_rand, lam=lam)
return p, mae_loss, ce_loss, mnce_loss, pred, imgs, imgs_mask
else:
latent = self.main_net.forward_encoder_test(x)
p = self.classifier(latent[:, 0, :])
return p
if self.arch == "resnet50 + mae":
if self.training:
mae_loss, latent, pred, imgs, imgs_mask = self.main_mae(x, mask_ratio=0.5)
x_0 = self.layer0(pred)
x_1 = self.layer1(x_0)
x_2 = self.layer2(x_1)
x_3 = self.layer3(x_2)
x_4 = self.layer4(x_3)
x_pool_4 = self.avgpool(x_4)
x_pool_4 = x_pool_4.view(x_pool_4.size(0), x_pool_4.size(1))
p = self.classifier(torch.cat((latent[:, 0, :], x_pool_4), dim=1))
ce_loss, mnce_loss = self.cal_loss(p, target, latent=latent, target_rand=target_rand, lam=lam)
return p, mae_loss, ce_loss, mnce_loss, pred, imgs, imgs_mask
else:
latent = self.main_mae.forward_encoder_test(x)
x_0 = self.layer0(x)
x_1 = self.layer1(x_0)
x_2 = self.layer2(x_1)
x_3 = self.layer3(x_2)
x_4 = self.layer4(x_3)
x_pool_4 = self.avgpool(x_4)
x_pool_4 = x_pool_4.view(x_pool_4.size(0), x_pool_4.size(1))
p = self.classifier(torch.cat((latent[:, 0, :], x_pool_4), dim=1))
return p
def cal_loss(self, output, target, latent=None, target_rand=None, lam=None):
criterion_ce = nn.CrossEntropyLoss(label_smoothing=0.0)
criterion_mnce = MaeContrastiveLoss(t=1.0)
criterion_kl = nn.KLDivLoss(reduction="batchmean")
if self.args_config.cutmix:
ce_loss = criterion_ce(output, target) * lam + criterion_ce(output, target_rand) * (1. - lam)
# ce_loss = criterion_ce(output, target)
else:
ce_loss = criterion_ce(output, target)
if self.arch == "resnet50":
return ce_loss
if self.arch == "mae":
if self.args_config.cutmix:
mnce_loss = criterion_mnce(latent, target) * lam + criterion_mnce(latent, target_rand) * (1. - lam)
# mnce_loss = criterion_mnce(latent, target)
else:
mnce_loss = criterion_mnce(latent, target)
return ce_loss, mnce_loss
if self.arch == "resnet50 + mae":
if self.args_config.cutmix:
mnce_loss = criterion_mnce(latent, target) * lam + criterion_mnce(latent, target_rand) * (1. - lam)
# mnce_loss = criterion_mnce(latent, target)
else:
mnce_loss = criterion_mnce(latent, target)
return ce_loss, mnce_loss