-
Notifications
You must be signed in to change notification settings - Fork 14
/
models.py
141 lines (110 loc) · 5.17 KB
/
models.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
import os
import numpy as np
from sklearn.metrics import confusion_matrix
from sklearn.metrics.ranking import roc_auc_score
import torch
import torch.nn as nn
import torchvision.models as models
import resnet_wider
import densenet
def ClassificationNet(arch_name, num_class, conv=None, weight=None, activation=None):
if weight is None:
weight = "none"
if conv is None:
try:
model = resnet_wider.__dict__[arch_name](sobel=False)
except:
model = models.__dict__[arch_name](pretrained=False)
else:
if arch_name.lower().startswith("resnet"):
model = resnet_wider.__dict__[arch_name + "_layerwise"](conv, sobel=False)
elif arch_name.lower().startswith("densenet"):
model = densenet.__dict__[arch_name + "_layerwise"](conv)
if arch_name.lower().startswith("resnet"):
kernelCount = model.fc.in_features
if activation is None:
model.fc = nn.Linear(kernelCount, num_class)
elif activation == "Sigmoid":
model.fc = nn.Sequential(nn.Linear(kernelCount, num_class), nn.Sigmoid())
# init the fc layer
if activation is None:
model.fc.weight.data.normal_(mean=0.0, std=0.01)
model.fc.bias.data.zero_()
else:
model.fc[0].weight.data.normal_(mean=0.0, std=0.01)
model.fc[0].bias.data.zero_()
elif arch_name.lower().startswith("densenet"):
kernelCount = model.classifier.in_features
if activation is None:
model.classifier = nn.Linear(kernelCount, num_class)
elif activation == "Sigmoid":
model.classifier = nn.Sequential(nn.Linear(kernelCount, num_class), nn.Sigmoid())
# init the classifier layer
if activation is None:
model.classifier.weight.data.normal_(mean=0.0, std=0.01)
model.classifier.bias.data.zero_()
else:
model.classifier[0].weight.data.normal_(mean=0.0, std=0.01)
model.classifier[0].bias.data.zero_()
def _weight_loading_check(_arch_name, _activation, _msg):
if len(_msg.missing_keys) != 0:
if _arch_name.lower().startswith("resnet"):
if _activation is None:
assert set(_msg.missing_keys) == {"fc.weight", "fc.bias"}
else:
assert set(_msg.missing_keys) == {"fc.0.weight", "fc.0.bias"}
elif _arch_name.lower().startswith("densenet"):
if _activation is None:
assert set(_msg.missing_keys) == {"classifier.weight", "classifier.bias"}
else:
assert set(_msg.missing_keys) == {"classifier.0.weight", "classifier.0.bias"}
if weight.lower() == "imagenet":
pretrained_model = models.__dict__[arch_name](pretrained=True)
state_dict = pretrained_model.state_dict()
# delete fc layer
for k in list(state_dict.keys()):
if k.startswith('fc') or k.startswith('classifier'):
del state_dict[k]
msg = model.load_state_dict(state_dict, strict=False)
_weight_loading_check(arch_name, activation, msg)
print("=> loaded supervised ImageNet pre-trained model")
elif os.path.isfile(weight):
checkpoint = torch.load(weight, map_location="cpu")
if "state_dict" in checkpoint:
state_dict = checkpoint["state_dict"]
else:
state_dict = checkpoint
state_dict = {k.replace("module.", ""): v for k, v in state_dict.items()}
state_dict = {k.replace("module.encoder_q.", ""): v for k, v in state_dict.items()}
for k in list(state_dict.keys()):
if k.startswith('fc') or k.startswith('classifier') or k.startswith('projection_head') or k.startswith('prototypes'):
del state_dict[k]
msg = model.load_state_dict(state_dict, strict=False)
_weight_loading_check(arch_name, activation, msg)
print("=> loaded pre-trained model '{}'".format(weight))
print("missing keys:", msg.missing_keys)
# reinitialize fc layer again
if arch_name.lower().startswith("resnet"):
if activation is None:
model.fc.weight.data.normal_(mean=0.0, std=0.01)
model.fc.bias.data.zero_()
else:
model.fc[0].weight.data.normal_(mean=0.0, std=0.01)
model.fc[0].bias.data.zero_()
elif arch_name.lower().startswith("densenet"):
if activation is None:
model.classifier.weight.data.normal_(mean=0.0, std=0.01)
model.classifier.bias.data.zero_()
else:
model.classifier[0].weight.data.normal_(mean=0.0, std=0.01)
return model
def build_classification_model(args):
if args.init.lower() =="random" or args.init.lower() =="imagenet":
model = ClassificationNet(args.model_name.lower(), args.num_class, weight=args.init,
activation=args.activate)
else:
model = ClassificationNet(args.model_name.lower(), args.num_class, weight=args.proxy_dir,
activation=args.activate)
return model
def save_checkpoint(state,filename='model'):
torch.save( state,filename + '.pth.tar')