CNN卷积和反卷积输出的计算方法

月疯2026-07-10 23:28

卷积：

复制代码

# 输入通道1，输出通道2，卷积核3x3，步长2，填充1
conv_layer = nn.Conv2d(in_channels=1, out_channels=2, kernel_size=3, stride=2, padding=1)

# 前向传播
x_conv = conv_layer(x)
print(f"卷积后形状: {x_conv.shape}")  # torch.Size([1, 2, 3, 3])

# 验证输出尺寸: (5 - 3 + 2*1)/2 + 1 = (4)/2 + 1 = 3 
print(f"卷积后的特征图 (第一个通道):\n{x_conv[0, 0, :, :]}\n")

反卷积：

复制代码

# 输入通道2（对应卷积的输出），输出通道1，卷积核3x3，步长2，填充1
deconv_layer = nn.ConvTranspose2d(in_channels=2, out_channels=1, kernel_size=3, stride=2, padding=1, output_padding=0)

# 前向传播
x_deconv = deconv_layer(x_conv)
print(f"反卷积后形状: {x_deconv.shape}")  # torch.Size([1, 1, 5, 5])

# 验证输出尺寸: (3 - 1)*2 - 2*1 + 3 + 0 = 4 - 2 + 3 = 5 
print(f"反卷积恢复后的特征图:\n{x_deconv[0, 0, :, :]}")

完整demo：

复制代码

import torch
import torch.nn as nn

torch.manual_seed(42)

# 1. 输入
x = torch.randn(1, 1, 5, 5)
print(f"输入: {x.shape}\n{x[0,0]}\n")

# 2. 卷积 (5x5 -> 3x3)
conv = nn.Conv2d(1, 2, kernel_size=3, stride=2, padding=1)
x_conv = conv(x)
print(f"卷积后: {x_conv.shape}\n{x_conv[0,0]}\n")

# 3. 反卷积 (3x3 -> 5x5)
deconv = nn.ConvTranspose2d(2, 1, kernel_size=3, stride=2, padding=1, output_padding=0)
x_deconv = deconv(x_conv)
print(f"反卷积后: {x_deconv.shape}\n{x_deconv[0,0]}")