forked from Jermmy/pytorch-quantization-demo
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
125 lines (106 loc) · 4.1 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
import torch
import torch.nn as nn
import torch.nn.functional as F
from module import *
class Net(nn.Module):
def __init__(self, num_channels=1):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(num_channels, 40, 3, 1)
self.conv2 = nn.Conv2d(40, 40, 3, 1, groups=20)
self.fc = nn.Linear(5*5*40, 10)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv2(x))
x = F.max_pool2d(x, 2, 2)
x = x.view(-1, 5*5*40)
x = self.fc(x)
return x
def quantize(self, num_bits=8):
self.qconv1 = QConv2d(self.conv1, qi=True, qo=True, num_bits=num_bits)
self.qrelu1 = QReLU()
self.qmaxpool2d_1 = QMaxPooling2d(kernel_size=2, stride=2, padding=0)
self.qconv2 = QConv2d(self.conv2, qi=False, qo=True, num_bits=num_bits)
self.qrelu2 = QReLU()
self.qmaxpool2d_2 = QMaxPooling2d(kernel_size=2, stride=2, padding=0)
self.qfc = QLinear(self.fc, qi=False, qo=True, num_bits=num_bits)
def quantize_forward(self, x):
x = self.qconv1(x)
x = self.qrelu1(x)
x = self.qmaxpool2d_1(x)
x = self.qconv2(x)
x = self.qrelu2(x)
x = self.qmaxpool2d_2(x)
x = x.view(-1, 5*5*40)
x = self.qfc(x)
return x
def freeze(self):
self.qconv1.freeze()
self.qrelu1.freeze(self.qconv1.qo)
self.qmaxpool2d_1.freeze(self.qconv1.qo)
self.qconv2.freeze(qi=self.qconv1.qo)
self.qrelu2.freeze(self.qconv2.qo)
self.qmaxpool2d_2.freeze(self.qconv2.qo)
self.qfc.freeze(qi=self.qconv2.qo)
def quantize_inference(self, x):
qx = self.qconv1.qi.quantize_tensor(x)
qx = self.qconv1.quantize_inference(qx)
qx = self.qrelu1.quantize_inference(qx)
qx = self.qmaxpool2d_1.quantize_inference(qx)
qx = self.qconv2.quantize_inference(qx)
qx = self.qrelu2.quantize_inference(qx)
qx = self.qmaxpool2d_2.quantize_inference(qx)
qx = qx.view(-1, 5*5*40)
qx = self.qfc.quantize_inference(qx)
out = self.qfc.qo.dequantize_tensor(qx)
return out
class NetBN(nn.Module):
def __init__(self, num_channels=1):
super(NetBN, self).__init__()
self.conv1 = nn.Conv2d(num_channels, 40, 3, 1)
self.bn1 = nn.BatchNorm2d(40)
self.conv2 = nn.Conv2d(40, 40, 3, 1)
self.bn2 = nn.BatchNorm2d(40)
self.fc = nn.Linear(5 * 5 * 40, 10)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = F.relu(x)
x = F.max_pool2d(x, 2, 2)
x = self.conv2(x)
x = self.bn2(x)
x = F.relu(x)
x = F.max_pool2d(x, 2, 2)
x = x.view(-1, 5 * 5 * 40)
x = self.fc(x)
return x
def quantize(self, num_bits=8):
self.qconv1 = QConvBNReLU(self.conv1, self.bn1, qi=True, qo=True, num_bits=num_bits)
self.qmaxpool2d_1 = QMaxPooling2d(kernel_size=2, stride=2, padding=0)
self.qconv2 = QConvBNReLU(self.conv2, self.bn2, qi=False, qo=True, num_bits=num_bits)
self.qmaxpool2d_2 = QMaxPooling2d(kernel_size=2, stride=2, padding=0)
self.qfc = QLinear(self.fc, qi=False, qo=True, num_bits=num_bits)
def quantize_forward(self, x):
x = self.qconv1(x)
x = self.qmaxpool2d_1(x)
x = self.qconv2(x)
x = self.qmaxpool2d_2(x)
x = x.view(-1, 5*5*40)
x = self.qfc(x)
return x
def freeze(self):
self.qconv1.freeze()
self.qmaxpool2d_1.freeze(self.qconv1.qo)
self.qconv2.freeze(qi=self.qconv1.qo)
self.qmaxpool2d_2.freeze(self.qconv2.qo)
self.qfc.freeze(qi=self.qconv2.qo)
def quantize_inference(self, x):
qx = self.qconv1.qi.quantize_tensor(x)
qx = self.qconv1.quantize_inference(qx)
qx = self.qmaxpool2d_1.quantize_inference(qx)
qx = self.qconv2.quantize_inference(qx)
qx = self.qmaxpool2d_2.quantize_inference(qx)
qx = qx.view(-1, 5*5*40)
qx = self.qfc.quantize_inference(qx)
out = self.qfc.qo.dequantize_tensor(qx)
return out