自动摘要: 参考链接： 1.[https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.htmltorch.nn.Conv1d](https://p ……..
参考链接：
- https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html#torch.nn.Conv1d
- https://github.com/pytorch/tutorials/blob/082c8b1bddb48b75f59860db3679d8c439238f10/advanced_source/static_quantization_tutorial.rst#L373
- 简单来说如下：
- 每conv层是 w1x+b1 ，两层就是 w2(w1x+b1)+b2 , 合并成一层就是 w2w1x+w2b1+b2 ;
- 新的权重就是 w2w1, 新的偏置就是 w2*b1+b2
- 注意偏置计算!!!
- 是矩阵乘上一层的权重，写的是否注意顺序；
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
66import torch
import torch.nn as nn
import time
class MyModel(nn.Module):
def __init__(self, feature_size=512, hidden_dim=512):
super(MyModel, self).__init__()
self.hidden_dim = hidden_dim
# 定义多个 nn.Conv1d 层
self.fc_0 = nn.Conv1d(feature_size, hidden_dim * 4, 1)
self.fc_1 = nn.Conv1d(hidden_dim * 4, hidden_dim, 1)
self.fc_2 = nn.Conv1d(hidden_dim, hidden_dim // 2, 1)
w0, w1, w2 = self.fc_0.weight.reshape(hidden_dim * 4, -1), self.fc_1.weight.reshape(hidden_dim , -1), self.fc_2.weight.reshape(hidden_dim // 2, -1)
b0, b1, b2 = self.fc_0.bias.reshape(hidden_dim * 4, -1), self.fc_1.bias.reshape(hidden_dim , -1), self.fc_2.bias.reshape(hidden_dim // 2, -1)
# 将权重和偏置参数展开并合并
combined_weight = torch.mm(w2, torch.mm(w1, w0))
combined_bias = b2 + torch.mm(w2, b1) + torch.mm(torch.mm(w2, w1),b0)
# 创建新的 nn.Conv1d 层
self.fc_combined = nn.Conv1d(feature_size, hidden_dim // 2, 1)
self.fc_combined.weight = nn.Parameter(combined_weight.unsqueeze(-1))
self.fc_combined.bias = nn.Parameter(combined_bias.squeeze(-1))
def forward(self, features):
net = self.fc_combined(features)
return net
def forward2(self, features):
net2 = self.fc_2(self.fc_1(self.fc_0(features)))
return net2
def check(self, features):
net = self.fc_combined(features)
net2 = self.fc_2(self.fc_1(self.fc_0(features)))
print((net - net2).sum())
def t(fun):
def inner():
s= time.time()
fun()
print("time cost: ", time.time() - s)
return inner
# 创建模型
model = MyModel()
def f():
model.check(torch.randn(1, 512, 1024))
def f2():
model(torch.randn(1, 512, 1024))
def f3():
model.forward2(torch.randn(1, 512, 1024))
print("合并前后误差：")
f()
print("合并后：")
f2()
print("合并前：")
f3()
- 是矩阵乘上一层的权重，写的是否注意顺序；
