-
Notifications
You must be signed in to change notification settings - Fork 1
/
target_44cnn_quantz.py
150 lines (109 loc) · 6.97 KB
/
target_44cnn_quantz.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
import brevitas.nn as qnn
import torch.nn as nn
from torch.nn.utils import prune
import numpy as np
class QTtrgChessNET(nn.Module):
def __init__(self, n_bits=4, w_bits=4, rqt=True, b_q = None, hidden_size=128):
super(QTtrgChessNET, self).__init__()
# define layers of CNN
# input >> chessboard (12,8,8)
self.quant_1 = qnn.QuantIdentity(bit_width=n_bits, return_quant_tensor=rqt)
self.qinp_chessboard = qnn.QuantConv2d(12, hidden_size, kernel_size=3, stride=1, padding=1, bias=True, bias_quant=b_q, weight_bit_width=w_bits)
#self.qbatchn1 = qnn.BatchNorm2dToQuantScaleBias(hidden_size) #nn.BatchNorm2d(hidden_size)
self.qrelu1 = qnn.QuantReLU(bit_width=n_bits, return_quant_tensor=rqt)
self.quant_2 = qnn.QuantIdentity(bit_width=n_bits, return_quant_tensor=rqt)
self.qconv2 = qnn.QuantConv2d(hidden_size, hidden_size, kernel_size=3, stride=1, padding=1, bias=True, bias_quant=b_q, weight_bit_width=w_bits)
#self.qbatchn2 = qnn.BatchNorm2dToQuantScaleBias(hidden_size) #nn.BatchNorm2d(hidden_size) #qnn.BatchNorm2dToQuantScaleBias(hidden_size)
self.qrelu2 = qnn.QuantReLU(bit_width=n_bits, return_quant_tensor=rqt)
self.quant_3 = qnn.QuantIdentity(bit_width=n_bits, return_quant_tensor=rqt)
self.qconv3 = qnn.QuantConv2d(hidden_size, hidden_size, kernel_size=3, stride=1, padding=1, bias=True, bias_quant=b_q, weight_bit_width=w_bits)
#self.qbatchn3 = qnn.BatchNorm2dToQuantScaleBias(hidden_size) #nn.BatchNorm2d(hidden_size) #qnn.BatchNorm2dToQuantScaleBias(hidden_size)
self.qrelu3 = qnn.QuantReLU(bit_width=n_bits, return_quant_tensor=rqt)
self.quant_4 = qnn.QuantIdentity(bit_width=n_bits, return_quant_tensor=rqt)
self.qconv4 = qnn.QuantConv2d(hidden_size, hidden_size, kernel_size=3, stride=1, padding=1, bias=True, bias_quant=b_q, weight_bit_width=w_bits)
#self.qbatchn4 = qnn.BatchNorm2dToQuantScaleBias(hidden_size) #nn.BatchNorm2d(hidden_size) #qnn.BatchNorm2dToQuantScaleBias(hidden_size)
self.qrelu4 = qnn.QuantReLU(bit_width=n_bits, return_quant_tensor=rqt)
self.quant_5 = qnn.QuantIdentity(bit_width=n_bits, return_quant_tensor=rqt)
self.qconv5 = qnn.QuantConv2d(hidden_size, hidden_size, kernel_size=3, stride=1, padding=1, bias=True, bias_quant=b_q, weight_bit_width=w_bits)
#self.qbatchn5 = qnn.BatchNorm2dToQuantScaleBias(hidden_size) #nn.BatchNorm2d(hidden_size) #qnn.BatchNorm2dToQuantScaleBias(hidden_size)
self.qrelu5 = qnn.QuantReLU(bit_width=n_bits, return_quant_tensor=rqt)
self.flatten = nn.Flatten()
self.qfc1 = qnn.QuantLinear(hidden_size * 64, 64, bias=True, bias_quant=b_q, weight_bit_width=w_bits)
#self.q_flat_chess_relu = qnn.QuantReLU(bit_width=n_bits, return_quant_tensor=rqt)
# input >> source (the selected squares (64,) array)
#self.quant_source1 = qnn.QuantIdentity(bit_width=n_bits, return_quant_tensor=rqt)
self.qinput_source = qnn.QuantLinear(64, 64, bias=True, bias_quant=b_q, weight_bit_width=w_bits)
#self.q_source_relu1 = qnn.QuantReLU(bit_width=n_bits, return_quant_tensor=rqt)
self.qbatchn1d_merge = nn.BatchNorm1d(64)
self.quant_merge = qnn.QuantIdentity(bit_width=n_bits, return_quant_tensor=rqt)
self.q_merge_relu = qnn.QuantReLU(bit_width=n_bits, return_quant_tensor=rqt)
self.qSigmoid = qnn.QuantSigmoid(bit_width=n_bits, return_quant_tensor=False)
# output target (the targeted square)
self.qoutput_target = qnn.QuantLinear(64, 64, bias=True, bias_quant=b_q, weight_bit_width=w_bits)
# enable pruning
self.pruning_conv(True)
# from https://github.com/zama-ai/concrete-ml/blob/release/0.6.chessboard/docs/advanced_examples/ConvolutionalNeuralNetwork.ipynb
def pruning_conv(self, enable):
"""Enables or removes pruning."""
# Maximum number of active neurons (i.e. corresponding weight != 0)
n_active = 84
# Go through all the convolution layers
for layer in (self.qinp_chessboard, self.qconv2, self.qconv3, self.qconv4, self.qconv5):
s = layer.weight.shape
# Compute fan-in (number of inputs to a neuron)
# and fan-out (number of neurons in the layer)
st = [s[0], np.prod(s[1:])]
# The number of input neurons (fan-in) is the product of
# the kernel width x height x inChannels.
if st[1] > n_active:
if enable:
# This will create a forward hook to create a mask tensor that is multiplied
# with the weights during forward. The mask will contain 0s or 1s
prune.l1_unstructured(layer, "weight", (st[1] - n_active) * st[0])
else:
# When disabling pruning, the mask is multiplied with the weights
# and the result is stored in the weights member
prune.remove(layer, "weight")
def forward(self, chessboard, source):
# define forward behavior
# add sequence of convolutional
chessboard = self.quant_1(chessboard)
chessboard = self.qinp_chessboard(chessboard)
#chessboard = self.qbatchn1(chessboard)
chessboard = self.qrelu1(chessboard)
chessboard = self.quant_2(chessboard)
chessboard = self.qconv2(chessboard)
#chessboard = self.qbatchn2(chessboard)
chessboard = self.qrelu2(chessboard)
chessboard = self.quant_3(chessboard)
chessboard = self.qconv3(chessboard)
#chessboard = self.qbatchn3(chessboard)
chessboard = self.qrelu3(chessboard)
chessboard = self.quant_4(chessboard)
chessboard = self.qconv4(chessboard)
#chessboard = self.qbatchn4(chessboard)
chessboard = self.qrelu4(chessboard)
chessboard = self.quant_5(chessboard)
chessboard = self.qconv5(chessboard)
#chessboard = self.qbatchn5(chessboard)
chessboard = self.qrelu5(chessboard)
chessboard = self.flatten(chessboard)
chessboard = self.qfc1(chessboard)
#chessboard = self.q_flat_chess_relu(chessboard)
#source = self.quant_source1(source)
source = self.qinput_source(source)
#source = self.q_source_relu1(source)
# merging chessboard (context + selected source square)
chessboard = self.quant_merge(chessboard)
source = self.quant_merge(source)
#print("SCALE -->",chessboard.scale.item()-source.scale.item())
merge = chessboard + source
merge = self.qbatchn1d_merge(merge)
merge = self.q_merge_relu(merge)
## not good
#x = self.qSigmoid(merge)
#x_target = self.qoutput_target(x)
## better
x = self.qoutput_target(merge)
x_target = self.qSigmoid(x)
return x_target