如何将GAN的输入图像尺寸从32×32扩展至128×128？

原GAN针对32×32图像设计，扩展到128×128的核心是调整生成器的上采样层数和判别器的下采样层数，同时匹配特征通道数的变化，保证空间尺寸与特征维度的连续性。以下是完整重构代码及关键说明：

1. 生成器(Generator)重构 #

原生成器通过4次转置卷积从1×1噪声生成32×32图像，要得到128×128需增加1次上采样操作，同时调整通道数递减节奏（保持每次转置卷积后特征图尺寸翻倍、通道数减半）：

class Generator(Module):
    def __init__(self):
        super().__init__()
        self.gen = Sequential(
            # 输入: (batch_size, 100, 1, 1)
            ConvTranspose2d(in_channels=100, out_channels=1024, kernel_size=4, stride=1, padding=0, bias=False),
            # 输出: (batch_size, 1024, 4, 4)
            BatchNorm2d(1024),
            ReLU(inplace=True),

            ConvTranspose2d(in_channels=1024, out_channels=512, kernel_size=4, stride=2, padding=1, bias=False),
            # 输出: (batch_size, 512, 8, 8)
            BatchNorm2d(512),
            ReLU(inplace=True),

            ConvTranspose2d(in_channels=512, out_channels=256, kernel_size=4, stride=2, padding=1, bias=False),
            # 输出: (batch_size, 256, 16, 16)
            BatchNorm2d(256),
            ReLU(inplace=True),

            ConvTranspose2d(in_channels=256, out_channels=128, kernel_size=4, stride=2, padding=1, bias=False),
            # 输出: (batch_size, 128, 32, 32)
            BatchNorm2d(128),
            ReLU(inplace=True),

            ConvTranspose2d(in_channels=128, out_channels=3, kernel_size=4, stride=2, padding=1, bias=False),
            # 输出: (batch_size, 3, 128, 128)
            Tanh()
        )

    def forward(self, input):
        return self.gen(input)

关键调整：

• 新增1层转置卷积，将32×32特征图上采样至128×128
• 初始通道数从512提升至1024，保证深层特征表达能力，后续每层通道数减半
• 统一使用kernel_size=4, stride=2, padding=1的上采样参数，确保特征图尺寸精确翻倍

2. 判别器(Discriminator)重构 #

判别器需与生成器对称，增加1次下采样操作，从128×128逐步压缩至1×1，通道数逐层翻倍：

class Discriminator(Module):
    def __init__(self):
        super().__init__()
        self.dis = Sequential(
            # 输入: (batch_size, 3, 128, 128)
            Conv2d(in_channels=3, out_channels=64, kernel_size=4, stride=2, padding=1, bias=False),
            # 输出: (batch_size, 64, 64, 64)
            LeakyReLU(0.2, inplace=True),

            Conv2d(in_channels=64, out_channels=128, kernel_size=4, stride=2, padding=1, bias=False),
            # 输出: (batch_size, 128, 32, 32)
            BatchNorm2d(128),
            LeakyReLU(0.2, inplace=True),

            Conv2d(in_channels=128, out_channels=256, kernel_size=4, stride=2, padding=1, bias=False),
            # 输出: (batch_size, 256, 16, 16)
            BatchNorm2d(256),
            LeakyReLU(0.2, inplace=True),

            Conv2d(in_channels=256, out_channels=512, kernel_size=4, stride=2, padding=1, bias=False),
            # 输出: (batch_size, 512, 8, 8)
            BatchNorm2d(512),
            LeakyReLU(0.2, inplace=True),

            Conv2d(in_channels=512, out_channels=1024, kernel_size=4, stride=2, padding=1, bias=False),
            # 输出: (batch_size, 1024, 4, 4)
            BatchNorm2d(1024),
            LeakyReLU(0.2, inplace=True),

            Conv2d(in_channels=1024, out_channels=1, kernel_size=4, stride=1, padding=0, bias=False),
            # 输出: (batch_size, 1, 1, 1)
            Sigmoid()
        )

    def forward(self, input):
        return self.dis(input)

关键调整：

• 新增1层卷积下采样，将128×128图像压缩至64×64，后续每层尺寸减半
• 初始通道数从32提升至64，逐层翻倍，最后通过4×4卷积将4×4特征图压缩为1×1
• 最后一层使用kernel_size=4, stride=1, padding=0，刚好匹配4×4特征图的尺寸转换

3. 数据与训练适配 #

• 确保输入npz文件包含形状为(N, 128, 128, 3)的图像数据，数据转置逻辑保持不变
• 训练流程完全沿用原代码，仅需注意显存占用：128×128图像显存消耗更高，建议将batch_size调整为16或8
• 权重初始化函数、优化器参数可保持不变，若训练不稳定可将学习率降至0.0001

内容的提问来源于stack exchange，提问作者Jett t