GAN PyTorch：在批次中生成相同的图像

Question

所以看下图就可以看出问题

Image generated per epoch screenshot 每批次的图像相同但不完全相同，只有 5% 的 pizel 略有不同

在每个epoch结束时发送到保存图像生成器的随机噪声的潜在大小是torch.randn(Batch_size,Latent_size,1,1)

但似乎生成模型正在强迫自己，而不管每个图像的随机噪声输入是否不同， 相反，随机噪声在整个 diff 时期是恒定的，但图像生成不断变化。

代码：

#Data Preparation
import torchvision.transforms as tt
from torchvision.datasets import ImageFolder
from torch.utils.data import DataLoader

#mean = (0.6528, 0.4833, 0.4052)
#deviation = (0.2362, 0.2079, 0.1988)

ROOT = "./data/Gender/train"
BATCH_SIZE = 120
LATENT_SIZE = 100

train_ds = ImageFolder(ROOT, tt.Compose([tt.RandomHorizontalFlip(),tt.Resize(64),tt.ToTensor(),tt.Normalize((0.6528, 0.4833, 0.4052),(0.2362, 0.2079, 0.1988))]))
train_dl = DataLoader(train_ds, BATCH_SIZE, shuffle=1, num_workers=2, pin_memory=1)

#Generator

G = nn.Sequential(
    nn.ConvTranspose2d(LATENT_SIZE,800,kernel_size=4,stride=2,padding=1,bias=False),#1000 2 2 
    nn.BatchNorm2d(800),
    nn.ReLU(),

    nn.ConvTranspose2d(800,400,kernel_size=4,stride=2,padding=1,bias=False),#600 4 4
    nn.BatchNorm2d(400),
    nn.ReLU(),
    
    nn.ConvTranspose2d(400,200,kernel_size=4,stride=2,padding=1,bias=False),#300 8 8
    nn.BatchNorm2d(200),
    nn.ReLU(),
    
    nn.ConvTranspose2d(200,100,kernel_size=4,stride=2,padding=1,bias=False),#150 16
    nn.BatchNorm2d(100),
    nn.ReLU(),
    
    nn.ConvTranspose2d(100,50,kernel_size=4,stride=2,padding=1,bias=False),#50 32
    nn.BatchNorm2d(50),
    nn.ReLU(),
    
    nn.ConvTranspose2d(50,3,kernel_size=4,stride=2,padding=1,bias=False),#3 64
    nn.Tanh()
)

G = G.to(device)

#Generator Traing Function

def g_fit():
    g_opt.zero_grad()
    
    rand_g = torch.randn(BATCH_SIZE, LATENT_SIZE,1,1,dtype=torch.float32).to(device)
    fake_images = G(rand_g)
    g_loss = loss_fn(D(fake_images), real_labels)
    

    g_loss.backward()
    
    g_opt.step()
    
    return g_loss

#Saving Images

from torchvision.utils import save_image
from PIL import Image

gen = torch.randn(24,LATENT_SIZE,1,1).to(device)

def save_fake_images(epoch):

    with torch.no_grad():
        fake_images = G(gen)

        name = "./drive/MyDrive/gans_data/fake_n_images_"+str(epoch)+".png"
        
        #Denorming Images
        for i in range(len(fake_images)):
            fake_images[i][0]=(fake_images[i][0]*0.2362) +0.6528
            fake_images[i][1]=(fake_images[i][1]*0.2079) +0.4833
            fake_images[i][2]=(fake_images[i][2]*0.1988) +0.4052

        save_image(fake_images, name, nrow=6)

#this function is being used at the end of every training loop to save a small batch of images

Answer 1

这是 GAN 训练中的一个常见问题，称为 mode collapse。请查看有关如何解决此问题的一些信息：

https://developers.google.com/machine-learning/gan/problems#mode-collapse https://machinelearningmastery.com/practical-guide-to-gan-failure-modes/

根据个人经验，我发现 GAN 的训练变化无常，有时某些东西会工作一次，然后在下一次训练运行时就无法工作，或者某些参数的微小变化会对生成器的性能产生很大影响.不断尝试新事物，看看效果如何。