-
Notifications
You must be signed in to change notification settings - Fork 0
/
Quant.py
192 lines (162 loc) · 7.71 KB
/
Quant.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
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import torch
import torch.nn.functional as F
from torch.nn.modules.linear import Linear
import math
from torch.nn.parameter import Parameter
from _quan_base import _Conv2dQ, Qmodes, _LinearQ, _ActQ
__all__ = ['Conv2dQ', 'LinearQ', 'ActQ']
class FunQ(torch.autograd.Function):
@staticmethod
def forward(ctx, weight, alpha, g, Qn, Qp):
assert alpha > 0, 'alpha = {}'.format(alpha)
ctx.save_for_backward(weight, alpha)
ctx.other = g, Qn, Qp
q_w = (weight / alpha).round().clamp(Qn, Qp)
w_q = q_w * alpha
return w_q
@staticmethod
# return 了权重的梯度以及scale的梯度
# 中间在判断q_w的值来确定scale的梯度
def backward(ctx, grad_weight):
weight, alpha = ctx.saved_tensors
g, Qn, Qp = ctx.other
q_w = weight / alpha
indicate_small = (q_w < Qn).float()
indicate_big = (q_w > Qp).float()
# indicate_middle = torch.ones(indicate_small.shape).to(indicate_small.device) - indicate_small - indicate_big
indicate_middle = 1.0 - indicate_small - indicate_big # Thanks to @haolibai
grad_alpha = ((indicate_small * Qn + indicate_big * Qp + indicate_middle * (
-q_w + q_w.round())) * grad_weight * g).sum().unsqueeze(dim=0)
grad_weight = indicate_middle * grad_weight
# The following operation can make sure that alpha is always greater than zero in any case and can also
# suppress the update speed of alpha. (Personal understanding)
# grad_alpha.clamp_(-alpha.item(), alpha.item()) # FYI
return grad_weight, grad_alpha, None, None, None
def grad_scale(x, scale):
y = x
y_grad = x * scale
return y.detach() - y_grad.detach() + y_grad
def round_pass(x):
y = x.round()
y_grad = x
return y.detach() - y_grad.detach() + y_grad
class Conv2dQ(_Conv2dQ):
def __init__(self, in_channels, out_channels, kernel_size, stride=1,
padding=0, dilation=1, groups=1, bias=True, nbits_w=8, mode=Qmodes.kernel_wise, **kwargs):
super(Conv2dQ, self).__init__(
in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size,
stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias,
nbits=nbits_w, mode=mode)
self.act = ActQ(in_features=in_channels, nbits_a=nbits_w)
def forward(self, x):
if self.alpha is None:
return F.conv2d(x, self.weight, self.bias, self.stride,
self.padding, self.dilation, self.groups)
# w_reshape = self.weight.reshape([self.weight.shape[0], -1]).transpose(0, 1)
Qn = -2 ** (self.nbits - 1)
Qp = 2 ** (self.nbits - 1) - 1
if self.training and self.init_state == 0:
# self.alpha.data.copy_(self.weight.abs().max() / 2 ** (self.nbits - 1))
self.alpha.data.copy_(2 * self.weight.abs().mean() / math.sqrt(Qp))
# self.alpha.data.copy_(self.weight.abs().max() * 2)
self.init_state.fill_(1)
"""
Implementation according to paper.
Feels wrong ...
When we initialize the alpha as a big number (e.g., self.weight.abs().max() * 2),
the clamp function can be skipped.
Then we get w_q = w / alpha * alpha = w, and $\frac{\partial w_q}{\partial \alpha} = 0$
As a result, I don't think the pseudo-code in the paper echoes the formula.
Please see jupyter/STE_LSQ.ipynb fo detailed comparison.
"""
g = 1.0 / math.sqrt(self.weight.numel() * Qp)
# Method1: 31GB GPU memory (AlexNet w4a4 bs 2048) 17min/epoch
alpha = grad_scale(self.alpha, g)
# print(alpha.shape)
# print(self.weight.shape)B
alpha = alpha.unsqueeze(1).unsqueeze(2).unsqueeze(3)
# 伪量化操作
w_q = round_pass((self.weight / alpha).clamp(Qn, Qp)) * alpha
x = self.act(x)
# w = w.clamp(Qn, Qp)
# q_w = round_pass(w)
# w_q = q_w * alpha
# Method2: 25GB GPU memory (AlexNet w4a4 bs 2048) 32min/epoch
# w_q = FunLSQ.apply(self.weight, self.alpha, g, Qn, Qp)
# wq = y.transpose(0, 1).reshape(self.weight.shape).detach() + self.weight - self.weight.detach()
return F.conv2d(x, w_q, self.bias, self.stride,
self.padding, self.dilation, self.groups)
class LinearQ(_LinearQ):
def __init__(self, in_features, out_features, bias=True, nbits_w=4, **kwargs):
super(LinearQ, self).__init__(in_features=in_features,
out_features=out_features, bias=bias, nbits=nbits_w, mode=Qmodes.kernel_wise)
self.act = ActQ(in_features=in_features, nbits_a=nbits_w)
def forward(self, x):
if self.alpha is None:
return F.linear(x, self.weight, self.bias)
Qn = -2 ** (self.nbits - 1)
Qp = 2 ** (self.nbits - 1) - 1
if self.training and self.init_state == 0:
self.alpha.data.copy_(2 * self.weight.abs().mean() / math.sqrt(Qp))
# self.alpha.data.copy_(self.weight.abs().max() / 2 ** (self.nbits - 1))
self.init_state.fill_(1)
g = 1.0 / math.sqrt(self.weight.numel() * Qp)
# Method1:
alpha = grad_scale(self.alpha, g)
alpha = alpha.unsqueeze(1)
w_q = round_pass((self.weight / alpha).clamp(Qn, Qp)) * alpha
x = self.act(x)
# w = self.weight / alpha
# w = w.clamp(Qn, Qp)
# q_w = round_pass(w)
# w_q = q_w * alpha
# Method2:
# w_q = FunLSQ.apply(self.weight, self.alpha, g, Qn, Qp)
return F.linear(x, w_q, self.bias)
class ActQ(_ActQ):
def __init__(self, in_features, nbits_a=4, mode=Qmodes.kernel_wise, **kwargs):
super(ActQ, self).__init__(in_features=in_features, nbits=nbits_a, mode=mode)
# print(self.alpha.shape, self.zero_point.shape)
def forward(self, x):
if self.alpha is None:
return x
if self.training and self.init_state == 0:
# The init alpha for activation is very very important as the experimental results shows.
# Please select a init_rate for activation.
# self.alpha.data.copy_(x.max() / 2 ** (self.nbits - 1) * self.init_rate)
if x.min() < -1e-5:
self.signed.data.fill_(1)
if self.signed == 1:
Qn = -2 ** (self.nbits - 1)
Qp = 2 ** (self.nbits - 1) - 1
else:
Qn = 0
Qp = 2 ** self.nbits - 1
self.alpha.data.copy_(2 * x.abs().mean() / math.sqrt(Qp))
self.zero_point.data.copy_(self.zero_point.data * 0.9 + 0.1 * (torch.min(x.detach()) - self.alpha.data * Qn))
self.init_state.fill_(1)
if self.signed == 1:
Qn = -2 ** (self.nbits - 1)
Qp = 2 ** (self.nbits - 1) - 1
else:
Qn = 0
Qp = 2 ** self.nbits - 1
g = 1.0 / math.sqrt(x.numel() * Qp)
# Method1:
# print(self.zero_point.size())
zero_point = (self.zero_point.round() - self.zero_point).detach() + self.zero_point
alpha = grad_scale(self.alpha, g)
zero_point = grad_scale(zero_point, g)
# x = round_pass((x / alpha).clamp(Qn, Qp)) * alpha
if len(x.shape)==2:
alpha = alpha.unsqueeze(0)
zero_point = zero_point.unsqueeze(0)
elif len(x.shape)==4 and x.shape[1]!=x.shape[2]:
alpha = alpha.unsqueeze(0).unsqueeze(2).unsqueeze(3)
zero_point = zero_point.unsqueeze(0).unsqueeze(2).unsqueeze(3)
# print("x_size:",x.size())
# print("zero_point.size()",zero_point.size())
# print()
x = round_pass((x / alpha + zero_point).clamp(Qn, Qp))
x = (x - zero_point) * alpha
return x