非对称卷积加padding操作

最新推荐文章于 2024-01-28 21:05:24 发布

天空飘来五行码

最新推荐文章于 2024-01-28 21:05:24 发布

阅读量476

点赞数

本文链接：https://blog.csdn.net/qq_38827988/article/details/127027811

版权

import torch
import torch.nn as nn

import torch

a = torch.randn(1, 32, 1024, 2048)  # 创建一个1x32x1024x2048的输入，BNWH格式
conv3x1 = torch.nn.Conv2d(32, 32, (3, 1),padding=0)  # 3x1卷积
conv1x3 = torch.nn.Conv2d(32, 32, (1, 3),padding=1)  # 1x3卷积
conv3x3 = torch.nn.Conv2d(32, 32, (3, 3),padding=1)  # 1x3卷积

out = conv3x1(a)
print(out.shape)  # 进行了3x1后的output的形状
out = conv1x3(out)
print(out.shape)
out = conv3x3(a)
print(out.shape)

# torch.Size([1, 32, 1022, 2048])
# torch.Size([1, 32, 1024, 2048])
# torch.Size([1, 32, 1024, 2048])

注意非对称卷积只可以有一个卷积使用padding，不然与普通卷积输出的尺寸不一致