运行Deep Convolutional Autoencoder时出现RuntimeError问题求助

问题场景 #

尝试构建解码器架构与DCGAN生成器一致的深度卷积自编码器，运行时触发维度不匹配错误。核心模型代码如下：

class Autoencoder(nn.Module):
def __init__(self):
    super(Autoencoder, self).__init__()

    self.encoder = nn.Sequential(
        nn.Conv2d(1, 16, 3, 2, 1, bias=False),
        nn.LeakyReLU(0.2, inplace=True),
        nn.Conv2d(16, 16 * 2, 3, 2, 1, bias=False),
        nn.BatchNorm2d(16 * 2),
        nn.LeakyReLU(0.2, inplace=True),
        nn.Conv2d(16 * 2, 16 * 4, 3, 2, 1, bias=False),
        nn.BatchNorm2d(16 * 4),
        nn.LeakyReLU(0.2, inplace=True),
        nn.Conv2d(16 * 4, 16 * 8, 3, 2, 1, bias=False),
        nn.BatchNorm2d(16 * 8),
        nn.LeakyReLU(0.2, inplace=True),
        nn.Conv2d(16 * 8, 16 * 16, 3),
        nn.Sigmoid()

    )
    self.decoder = nn.Sequential(
        nn.ConvTranspose2d(     16 * 16, 16 * 8, 3),
        nn.BatchNorm2d(64 * 8),
        nn.ReLU(True),
        nn.ConvTranspose2d(16 * 8, 16 * 4, 3, 2, 1, output_padding=1),
        nn.BatchNorm2d(16 * 4),
        nn.ReLU(True),
        nn.ConvTranspose2d(16 * 4, 16 * 2, 3, 2, 1, output_padding=1),
        nn.BatchNorm2d(16 * 2),
        nn.ReLU(True),
        nn.ConvTranspose2d(16 * 2,     16, 3, 2, 1, output_padding=1),
        nn.BatchNorm2d(16),
        nn.ReLU(True),
        nn.ConvTranspose2d(    16,      1, 3, 2, 1, output_padding=1),
        nn.Tanh()
    )

def forward(self, x):
    x = self.encoder(x)
    x = self.decoder(x)
    return x

错误信息 #

RuntimeError: Calculated padded input size per channel: (2 x 2). Kernel size: (3 x 3). Kernel size can't be greater than actual input size

问题分析 #

1. 编码器最后一层卷积尺寸不匹配：以MNIST 28x28输入为例，经过前4次步长为2的卷积后，特征图尺寸变为2x2。此时最后一层nn.Conv2d(16 * 8, 16 * 16, 3)使用3x3卷积核且无padding，输入2x2的特征图无法容纳3x3的卷积核，直接触发报错。
2. 解码器BatchNorm通道数错误：nn.BatchNorm2d(64 * 8)的通道数应为168=128，而非648=512，会导致后续维度不匹配。

修复方案 #

1. 修正编码器最后一层卷积 #

给最后一层卷积添加padding=1，保证输入2x2的特征图经过padding后可适配3x3卷积核，输出尺寸保持2x2：

nn.Conv2d(16 * 8, 16 * 16, 3, padding=1),

2. 修正解码器BatchNorm通道数 #

将错误的通道数改为16*8：

nn.BatchNorm2d(16 * 8),

修复后的完整模型代码 #

class Autoencoder(nn.Module):
    def __init__(self):
        super(Autoencoder, self).__init__()

        self.encoder = nn.Sequential(
            nn.Conv2d(1, 16, 3, 2, 1, bias=False),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Conv2d(16, 16 * 2, 3, 2, 1, bias=False),
            nn.BatchNorm2d(16 * 2),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Conv2d(16 * 2, 16 * 4, 3, 2, 1, bias=False),
            nn.BatchNorm2d(16 * 4),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Conv2d(16 * 4, 16 * 8, 3, 2, 1, bias=False),
            nn.BatchNorm2d(16 * 8),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Conv2d(16 * 8, 16 * 16, 3, padding=1),
            nn.Sigmoid()
        )
        self.decoder = nn.Sequential(
            nn.ConvTranspose2d(16 * 16, 16 * 8, 3),
            nn.BatchNorm2d(16 * 8),
            nn.ReLU(True),
            nn.ConvTranspose2d(16 * 8, 16 * 4, 3, 2, 1, output_padding=1),
            nn.BatchNorm2d(16 * 4),
            nn.ReLU(True),
            nn.ConvTranspose2d(16 * 4, 16 * 2, 3, 2, 1, output_padding=1),
            nn.BatchNorm2d(16 * 2),
            nn.ReLU(True),
            nn.ConvTranspose2d(16 * 2, 16, 3, 2, 1, output_padding=1),
            nn.BatchNorm2d(16),
            nn.ReLU(True),
            nn.ConvTranspose2d(16, 1, 3, 2, 1, output_padding=1),
            nn.Tanh()
        )

    def forward(self, x):
        x = self.encoder(x)
        x = self.decoder(x)
        return x

内容的提问来源于stack exchange，提问作者al. ekrami