pytorch nn.Conv1d and nn.Conv2d

最新推荐文章于 2025-04-28 16:07:48 发布

原创最新推荐文章于 2025-04-28 16:07:48 发布 · 719 阅读

1 ·

CC 4.0 BY-SA版权

Machine Learning 专栏收录该内容

8 篇文章

订阅专栏

本文介绍了一种将一维卷积层(nn.Conv1d)转换为二维卷积层(nn.Conv2d)的方法，通过增加维度并复制权重，使两者在特定输入下输出相同，适用于从一维信号处理过渡到二维图像处理的场景。

可以使用nn.Conv2d代替nn.Conv1d

import torch
import torch.nn as nn

conv1d = nn.Conv1d(in_channels=64,
                                  out_channels=1,
                                  kernel_size=2,
                                  stride=1,
                                  padding=1,
                                  dilation=1,
                                  bias=False)

conv2d = nn.Conv2d(in_channels=64,
                                  out_channels=1,
                                  kernel_size=(2, 1),
                                  stride=1,
                                  padding=(1, 0),
                                  dilation=1,
                                  bias=False)

with torch.no_grad():
    conv2d.weight.copy_(conv1d.weight.unsqueeze(3))   # copy conv1d weight to conv2d

input = torch.randn(1, 64, 114)

out1 = conv1d(input)
out2 = conv2d(input.unsqueeze(3))    #  add dim 1 to last dimension
out2 = out2.squeeze(3)

print((out1==out2).all())

# out print True, so out1 == out2

####