forked from RolandGao/RegSeg
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
156 lines (147 loc) · 6.42 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
from blocks import *
from competitor_blocks import BiseNetDecoder,SFNetDecoder,FaPNDecoder
from benchmark import benchmark_eval,benchmark_train,benchmark_memory
class RegSeg(nn.Module):
# exp48_decoder26 is what we call RegSeg in our paper
# exp53_decoder29 is a larger version of exp48_decoder26
# all the other models are for ablation studies
def __init__(self, name, num_classes, pretrained="", ablate_decoder=False,change_num_classes=False):
super().__init__()
self.stem=ConvBnAct(3,32,3,2,1)
body_name, decoder_name=name.split("_")
if "exp30" == body_name:
self.body=RegSegBody(5*[[1,4]]+8*[[1,10]])
elif "exp43"==body_name:
self.body=RegSegBody([[1],[1,2],[1,4],[1,6],[1,8],[1,10]]+7*[[1,12]])
elif "exp46"==body_name:
self.body=RegSegBody([[1],[1,2],[1,4],[1,6],[1,8]]+8*[[1,10]])
elif "exp47"==body_name:
self.body=RegSegBody([[1],[1,2],[1,4],[1,6],[1,8],[1,10],[1,12]]+6*[[1,14]])
elif "exp48"==body_name:
self.body=RegSegBody([[1],[1,2]]+4*[[1,4]]+7*[[1,14]])
elif "exp49"==body_name:
self.body=RegSegBody([[1],[1,2]]+6*[[1,4]]+5*[[1,6,12,18]])
elif "exp50"==body_name:
self.body=RegSegBody([[1],[1,2],[1,4],[1,6],[1,8],[1,10]]+7*[[1,3,6,12]])
elif "exp51"==body_name:
self.body=RegSegBody([[1],[1,2],[1,4],[1,6],[1,8],[1,10]]+7*[[1,4,8,12]])
elif "exp52"==body_name:
self.body=RegSegBody([[1],[1,2],[1,4]]+10*[[1,6]])
elif "exp53"==body_name:
self.body=RegSegBody2([[1],[1,2]]+4*[[1,4]]+7*[[1,14]])
elif "regnety600mf"==body_name:
self.body=RegNetY600MF()
else:
raise NotImplementedError()
if "decoder4" ==decoder_name:
self.decoder=Exp2_Decoder4(num_classes,self.body.channels())
elif "decoder10" ==decoder_name:
self.decoder=Exp2_Decoder10(num_classes,self.body.channels())
elif "decoder12" ==decoder_name:
self.decoder=Exp2_Decoder12(num_classes,self.body.channels())
elif "decoder14"==decoder_name:
self.decoder=Exp2_Decoder14(num_classes,self.body.channels())
elif "decoder26"==decoder_name:
self.decoder=Exp2_Decoder26(num_classes,self.body.channels())
elif "decoder29"==decoder_name:
self.decoder=Exp2_Decoder29(num_classes,self.body.channels())
elif "BisenetDecoder"==decoder_name:
self.decoder=BiseNetDecoder(num_classes,self.body.channels())
elif "SFNetDecoder"==decoder_name:
self.decoder=SFNetDecoder(num_classes,self.body.channels())
elif "FaPNDecoder"==decoder_name:
self.decoder=FaPNDecoder(num_classes,self.body.channels())
else:
raise NotImplementedError()
if pretrained != "" and not ablate_decoder:
dic = torch.load(pretrained, map_location='cpu')
print(type(dic))
if type(dic)==dict and "model" in dic:
dic=dic['model']
print(type(dic))
if change_num_classes:
current_model=self.state_dict()
new_state_dict={}
print("change_num_classes: True")
for k in current_model:
if dic[k].size()==current_model[k].size():
new_state_dict[k]=dic[k]
else:
print(k)
new_state_dict[k]=current_model[k]
self.load_state_dict(new_state_dict,strict=True)
else:
self.load_state_dict(dic,strict=True)
def forward(self,x):
input_shape=x.shape[-2:]
x=self.stem(x)
x=self.body(x)
x=self.decoder(x)
x = F.interpolate(x, size=input_shape, mode='bilinear', align_corners=False)
return x
def num_classes_speed_test():
# We find that the training speed highly correlates with the number of classes
# while the eval speed does not depend much on the number of classes
v=[10,20,30,40,50,60,70,80]
models=[]
for num_classes in v:
model=RegSeg("exp48_decoder26",num_classes=num_classes)
models.append(model)
benchmark_train([model],8,512,True,num_classes)
x=torch.randn(1,3,1024,2048)
benchmark_eval(models,x,True)
def dilation_speed_test():
group_width=16
w=256
x=torch.randn(1,256,64,128)
ts=[]
for d in range(1,19):
model=nn.Conv2d(w,w,3,1,padding=d,dilation=d,groups=w//group_width,bias=False)
ts.extend(benchmark_eval([model],x,True))
print(ts)
def block_speed_test():
print("block speed test")
model1=DBlock(256,256,[1],16,1,"se")
model2=DBlock(256,256,[1,1],16,1,"se")
model3=DBlock(256,256,[1,4],16,1,"se")
model4=DBlock(256,256,[1,10],16,1,"se")
x=torch.randn(1,256,64,128) # 1/16 original resolution
ts=benchmark_eval([model1,model2,model3,model4],x,True)
print(ts)
def calculate_flops():
from fvcore.nn import FlopCountAnalysis, flop_count_table,ActivationCountAnalysis
model1=RegSeg("exp48_decoder26",19).eval()
from competitors_models.DDRNet_Reimplementation import get_ddrnet_23,get_ddrnet_23slim
x=torch.randn(1,3,1024,2048)
model2=get_ddrnet_23().eval()
for model in [model1,model2]:
flops = FlopCountAnalysis(model, x)
print(flop_count_table(flops))
def calculate_params(model):
#https://discuss.pytorch.org/t/how-do-i-check-the-number-of-parameters-of-a-model/4325/6
import numpy as np
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
model_parameters = model.parameters()
params2 = sum([np.prod(p.size()) for p in model_parameters])
return params,params2
def cityscapes_speed_test():
print("cityscapes speed test")
from competitors_models.DDRNet_Reimplementation import get_ddrnet_23
regseg=RegSeg("exp48_decoder26",19)
ddrnet23=get_ddrnet_23()
x=torch.randn(1,3,1024,2048)
ts=[]
ts.extend(benchmark_eval([regseg,ddrnet23],x,True))
print(ts)
def camvid_speed_test():
print("camvid speed test")
from competitors_models.DDRNet_Reimplementation import get_ddrnet_23
regseg=RegSeg("exp48_decoder26",19)
ddrnet23=get_ddrnet_23()
x=torch.randn(1,3,720,960)
ts=[]
ts.extend(benchmark_eval([regseg,ddrnet23],x,True))
print(ts)
if __name__ == "__main__":
cityscapes_speed_test()