如何让GAN模型生成512×512及更大尺寸图像？卷积层原理答疑

一、Conv2D与Conv2DTranspose的尺寸调整逻辑 #

Conv2D（下采样，缩小图像） #

输入经过Conv2D后，输出尺寸由公式计算：
输出尺寸 = floor((输入尺寸 - 核尺寸 + 2×填充)/步长) + 1

• 当padding设为same时，填充值为(核尺寸-1)//2，保证输入输出尺寸比例为1/步长（步长为2时，尺寸减半）。
  比如原网络中，64×64输入经过Conv2D(64,4,2,'same')，计算过程：(64-4+2×1)/2 +1 = 32，输出32×32，三次下采样后得到8×8特征图。

Conv2DTranspose（上采样，放大图像） #

输入经过Conv2DTranspose后，输出尺寸公式：
输出尺寸 = (输入尺寸 -1)×步长 + 核尺寸 - 2×填充

• 同样当padding为same时，填充值为(核尺寸-1)//2，输入输出尺寸比例为步长×1（步长为2时，尺寸翻倍）。
  原网络中8×8输入经过Conv2DTranspose(128,4,2,'same')，计算：(8-1)×2 +4 -2×1=16，输出16×16，三次上采样后得到64×64图像。

二、修改为生成512×512图像的具体方案 #

1. 数据集适配 #

修改image_dataset_from_directory的image_size参数为(512,512)，同时调小batch_size避免显存溢出：

dataset = keras.preprocessing.image_dataset_from_directory(
    directory = 'Humans', label_mode = None, image_size = (512,512), batch_size = 8,
    shuffle = True
).map(lambda x: x/255.0)

2. 生成器重构 #

从4×4的初始特征图开始，通过6次上采样将尺寸翻倍至512×512，同时增加批量归一化稳定训练：

latent_dim = 128
generator = keras.models.Sequential(
    [
        layers.Input(shape = (latent_dim,)),
        layers.Dense(4*4*256),
        layers.Reshape((4,4,256)),
        layers.BatchNormalization(),
        layers.LeakyReLU(0.2),
        
        layers.Conv2DTranspose(128, kernel_size = 4, strides = 2, padding = 'same'),
        layers.BatchNormalization(),
        layers.LeakyReLU(0.2),
        layers.Conv2DTranspose(128, kernel_size = 4, strides = 2, padding = 'same'),
        layers.BatchNormalization(),
        layers.LeakyReLU(0.2),
        layers.Conv2DTranspose(64, kernel_size = 4, strides = 2, padding = 'same'),
        layers.BatchNormalization(),
        layers.LeakyReLU(0.2),
        layers.Conv2DTranspose(64, kernel_size = 4, strides = 2, padding = 'same'),
        layers.BatchNormalization(),
        layers.LeakyReLU(0.2),
        layers.Conv2DTranspose(32, kernel_size = 4, strides = 2, padding = 'same'),
        layers.BatchNormalization(),
        layers.LeakyReLU(0.2),
        layers.Conv2DTranspose(32, kernel_size = 4, strides = 2, padding = 'same'),
        layers.BatchNormalization(),
        layers.LeakyReLU(0.2),
        
        layers.Conv2D(3, kernel_size = 5,padding = 'same',activation = 'sigmoid')
    ]
)

3. 判别器重构 #

对应处理512×512输入，增加6次下采样层，逐步提升通道数捕捉细节：

discriminator = keras.models.Sequential(
    [
        keras.Input(shape = (512,512,3)),
        layers.Conv2D(64, kernel_size = 4, strides = 2, padding = 'same'),
        layers.LeakyReLU(0.2),
        layers.Conv2D(128, kernel_size = 4, strides = 2, padding = 'same'),
        layers.LeakyReLU(0.2),
        layers.Conv2D(256, kernel_size = 4, strides = 2, padding = 'same'),
        layers.LeakyReLU(0.2),
        layers.Conv2D(512, kernel_size = 4, strides = 2, padding = 'same'),
        layers.LeakyReLU(0.2),
        layers.Conv2D(512, kernel_size = 4, strides = 2, padding = 'same'),
        layers.LeakyReLU(0.2),
        layers.Conv2D(512, kernel_size = 4, strides = 2, padding = 'same'),
        layers.LeakyReLU(0.2),
        layers.Flatten(),
        layers.Dropout(0.3),
        layers.Dense(1,activation = 'sigmoid')
    ]
)

4. 训练参数优化 #

调低学习率避免大模型训练震荡，增加训练轮数保证收敛：

opt_gen = keras.optimizers.Adam(5e-5)
opt_disc = keras.optimizers.Adam(5e-5)
loss_fn = keras.losses.BinaryCrossentropy()

for epoch in range(1000):
  for idx, real in enumerate(tqdm(dataset)):
    batch_size = real.shape[0]
    random_latent_vectors = tf.random.normal(shape = (batch_size,latent_dim))
    fake = generator(random_latent_vectors)

    if idx % 20 == 0:
      img = keras.preprocessing.image.array_to_img(fake[0])
      img.save(f'gen_images/generated_img{epoch}_{idx}_.png')

    with tf.GradientTape() as disc_tape:
      loss_disc_real = loss_fn(tf.ones((batch_size,1)), discriminator(real))
      loss_disc_fake = loss_fn(tf.zeros((batch_size,1)), discriminator(fake))
      loss_disc = (loss_disc_real+loss_disc_fake)/2

    grads = disc_tape.gradient(loss_disc, discriminator.trainable_weights)
    opt_disc.apply_gradients(zip(grads, discriminator.trainable_weights))

    with tf.GradientTape() as gen_tape:
      fake = generator(random_latent_vectors)
      output = discriminator(fake)
      loss_gen = loss_fn(tf.ones((batch_size,1)),output)

    grads = gen_tape.gradient(loss_gen, generator.trainable_weights)
    opt_gen.apply_gradients(zip(grads, generator.trainable_weights))

注意事项 #

• 若出现显存溢出，可进一步调小batch_size至4，或开启混合精度训练；
• 大尺寸GAN训练初期图像会模糊，需训练500轮以上才能看到清晰效果；
• 数据集图像质量直接影响生成效果，建议提前预处理高清图像。

内容的提问来源于stack exchange，提问作者Andrei Arseni