ThriftyNets/Quan_layer.py at master · hleblevec/ThriftyNets · GitHub

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
import torch
import torch.nn
import sys
from math import *
first_batch=0
train=0
class Quan_layer(torch.nn.Conv2d):

    def __init__(self, in_features, out_features, kernel_size=3,stride=1, padding = 0,dilation=1, groups=1, bias=False,bits=8):
        super(Quan_layer, self).__init__(in_features, out_features, stride = stride, bias = bias, padding = padding, kernel_size = kernel_size,groups=groups)
        self.in_features = in_features
        self.out_features = out_features
        self.kernel_width = kernel_size
        self.groups = groups
        self.max_activation = 0
        self.bits=bits

        #self.n_bit_w = torch.nn.parameter.Parameter(torch.FloatTensor([8]))
        #self.n_bit_w.data.fill_(8)
        #self.n_bit_a = torch.nn.parameter.Parameter(torch.FloatTensor([8]))
        #self.n_bit_a.data.fill_(8)
                #b=torch.max(self.attentionWeights,dim=2)
        #print(b.shape)
        #self.attentionWeights.data[:,:,4]=b[0][:,:]

    def forward(self,input):
        global train
        global first_batch

        if train==1:
            maxi = torch.max(torch.abs(input)).item()
            if first_batch==1 or  self.max_activation < maxi:
                self.max_activation = maxi
        else:
        #print(torch.max(torch.abs(input)))
        #print(torch.max(torch.abs(input)).item())
            input = torch.clamp(input,-self.max_activation,self.max_activation)
        return torch.nn.functional.conv2d(quantifier(input,self.bits,self.max_activation), quantifier(self.weight,self.bits), bias = self.bias, stride = self.stride, padding = self.padding,groups=self.groups)
        #return torch.nn.functional.conv2d(quantifier(input,self.bits,0), quantifier(self.weight,self.bits), bias = self.bias, stride = self.stride, padding = self.padding,groups=self.groups)


class Quan_layer_fixed(torch.nn.Conv2d):

    def __init__(self, in_features, out_features, kernel_size=3,stride=1, padding = 0,dilation=1, groups=1, bias=False,bits=8):
        super(Quan_layer_fixed, self).__init__(in_features, out_features, stride = stride, bias = bias, padding = padding, kernel_size = kernel_size,groups=groups)
        self.in_features = in_features
        self.out_features = out_features
        self.kernel_width = kernel_size
        self.groups = groups
        self.max_activation = 19.6690
        self.max_weight = 0.7586
        self.bits=bits

        #self.n_bit_w = torch.nn.parameter.Parameter(torch.FloatTensor([8]))
        #self.n_bit_w.data.fill_(8)
        #self.n_bit_a = torch.nn.parameter.Parameter(torch.FloatTensor([8]))
        #self.n_bit_a.data.fill_(8)
                #b=torch.max(self.attentionWeights,dim=2)
        #print(b.shape)
        #self.attentionWeights.data[:,:,4]=b[0][:,:]

    def forward(self,input):
        global train
        global first_batch

        if train==1:
            maxi = torch.max(torch.abs(input)).item()
            if first_batch==1 or  self.max_activation < maxi:
                self.max_activation = maxi
        else:
            input = torch.clamp(input,-self.max_activation,self.max_activation)
        return torch.nn.functional.conv2d(fixed_quantifier(input,self.bits,self.max_activation), fixed_quantifier(self.weight,self.bits, self.max_weight), bias = self.bias, stride = self.stride, padding = self.padding,groups=self.groups)


class IntNoGradient(torch.autograd.Function):

    @staticmethod
    def forward(ctx, x):
        return x.int().float()

    @staticmethod
    def backward(ctx, g):
        return g

class IntNoGradientFloor(torch.autograd.Function):

    @staticmethod
    def forward(ctx, x):
        return x.floor()

    @staticmethod
    def backward(ctx, g):
        return g

def maxWeight(weight,maxi):
    #liste_max = []
    #index=0
    #maxi = 0
    global train
    w = weight
    v = w.view(-1)
    if(maxi==0 or train==1):
        maxi=torch.max(torch.abs(v))
    #maxi = torch.max(torch.abs(v)) #.cpu().data.numpy()
    n=0
    if(maxi<1):
        while(maxi<1):
            maxi*=2
            n+=1
        return n-1
    elif(maxi>=2):
        while(maxi>=2):
           maxi/=2
           n-=1
        return n-1
    else:
        return n-1

def quantifier(w, n_bit,maxi=0):

    maxi=maxWeight(w,maxi)

    #w = weight.clone().cuda()
    a = w.shape
    v = torch.zeros(a)
    v = v + pow(2, n_bit-1 + maxi)
    v = v.float() #FloatNoGradient.apply(v)
    v = v.cuda()
    w = w*v
    w = IntNoGradient.apply(w)
    #w = FloatNoGradient.apply(w)
    w = w/v
    return w

def fixed_quantifier(w, n_bit,maxi):
    if abs(maxi) < 1:
        int_w = 1
    else:
        int_w = ceil(log2(maxi)) + 1
    frac_w = n_bit - int_w
    a = w.shape
    v = torch.zeros(a)
    v = v + pow(2, frac_w)
    v = v.float()
    v = v.cuda()
    w = w*v
    w = IntNoGradientFloor.apply(w)
    w = w/v
    return w