Repository navigation
Expand file tree
/
Copy pathmodelAdjoint.py
More file actions
118 lines (94 loc) · 4.89 KB
/
Copy pathmodelAdjoint.py
File metadata and controls
118 lines (94 loc) · 4.89 KB
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
import torch.nn as nn
import torch.nn.init as init
from myconv import myconv2d
from mask import *
from adjointNetwork import *
class Lambda(nn.Module):
def __init__(self, func):
super().__init__()
self.func = func
def forward(self, x): return self.func(x)
def mnist_resize(x): return x.view(-1, 1, 28, 28)
def cifar_resize(x): return x.view(-1, 3, 32, 32)
def imagenet_resize(x): return x.view(-1, 3, 128, 128)
def noop(x): return x
class Flatten(nn.Module):
def forward(self, x): return x.view(x.size(0), -1)
act_func = nn.ReLU()
first = True
count = 0
def conv(ni, no, ks, s=1, bias=False, expansion=4, compression_factor=4, masking_factor=None):
if expansion==1:
# resnet 18
if no>64:
mask_layer = True
else:
mask_layer = False
else:
# resnet 50
if ni>64 and no>64:
mask_layer = True
else:
mask_layer = False
global first, count
count+=1
#count>=14
#print(ni,no,count)
if first:
first = False
return conv2dFirstLayer(ni, no, kernel_size=ks, stride=s, padding=ks//2, bias=bias)
return conv2dAdjoint(ni, no, kernel_size=ks, stride=s, padding=ks//2, bias=bias,mask_layer=mask_layer,compression_factor=compression_factor,masking_factor=masking_factor)
def conv_layer(ni, no, ks, s, zero_bn=False, act=True, expansion=1, compression_factor=4, masking_factor=None):
bn = batchNorm(no)
nn.init.constant_(bn.bn1.weight, 0. if zero_bn else 1.)
nn.init.constant_(bn.bn2.weight, 0. if zero_bn else 1.)
layers = [conv(ni, no, ks=ks, s=s, expansion=expansion, compression_factor=compression_factor, masking_factor=masking_factor), bn]
if act: layers += [act_func]
return nn.Sequential(*layers)
def init_cnn(m):
if isinstance(m, (nn.Conv2d, nn.Linear)):
init.kaiming_normal_(m.weight, a=1.0)
if getattr(m, 'bias', None) is not None: init.constant_(m.bias, 0.)
for l in m.children(): init_cnn(l)
class ResBlock(nn.Module):
def __init__(self, ni, no, expansion, compression_factor, masking_factor, s=1):
super().__init__()
#print('Res ni ',ni,'no ',no)
ni *= expansion
layers = [conv_layer(ni, no, 3, s, expansion=expansion, compression_factor=compression_factor, masking_factor=masking_factor),
conv_layer(no, no*expansion, 3, 1, zero_bn=True, act=False, expansion=expansion, compression_factor=compression_factor, masking_factor=masking_factor)
] if expansion == 1 else [
conv_layer(ni, no, 1, 1, expansion=expansion, compression_factor=compression_factor, masking_factor=masking_factor),
conv_layer(no, no, 3, s, expansion=expansion, compression_factor=compression_factor, masking_factor=masking_factor),
conv_layer(no, no*expansion, 1, 1, zero_bn=True, act=False, expansion=expansion, compression_factor=compression_factor, masking_factor=masking_factor)
]
self.convs = nn.Sequential(*layers)
self.idconv = noop if ni == no*expansion else conv_layer(ni, no*expansion, 1, 1, act=False, expansion=expansion, compression_factor=compression_factor, masking_factor=masking_factor)
self.pool = noop if s == 1 else nn.AvgPool2d(2,ceil_mode=True)
def forward(self, x):
return act_func(self.convs(x) + self.idconv(self.pool(x)))
class XResNet(nn.Sequential):
@classmethod
def create(cls, expansion, layers, c_in=3, c_out=10, resize=cifar_resize, compression_factor=4, masking_factor=None):
nbs = [c_in, 32,64,64]
stem = [conv_layer(nbs[i], nbs[i+1], 3, 2 if i==0 else 1,False,expansion=expansion)
for i in range(3)]
maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
nbs = [64//expansion, 64, 128, 256, 512]
res_layers = [cls._make_layer(nbs[i], nbs[i+1], expansion, 1 if i==0 else 2, l,compression_factor,masking_factor)
for i,l in enumerate(layers) ]
layers = [Lambda(resize), *stem, maxpool, *res_layers]
layers.extend([nn.AdaptiveAvgPool2d(1), Flatten(), linear(nbs[-1]*expansion, c_out) ])
resnet = cls(*layers)
init_cnn(resnet)
return resnet
@staticmethod
def _make_layer(ni, no, expansion, s, number_block,compression_factor,masking_factor):
return nn.Sequential(*[ResBlock(ni if i==0 else no, no, expansion, compression_factor, masking_factor,s=s if i==0 else 1)
for i in range(number_block)] )
def xresnet_fast18 (**kwargs):return XResNet.create(1, [2, 2, 2, 2], **kwargs)
def xresnet_fast34 (**kwargs): return XResNet.create(1, [3, 4, 6, 3], **kwargs)
def xresnet_fast50 (**kwargs): return XResNet.create(4, [3, 4, 6, 3], **kwargs)
def xresnet_fast50X2 (**kwargs): return XResNet.create(4, [6, 8, 12, 6], **kwargs)
def xresnet_fast101(**kwargs): return XResNet.create(4, [3, 4, 23, 3], **kwargs)
def resnet_fast152(**kwargs): return XResNet.create(4, [3, 8, 36, 3], **kwargs)