如何用GAN训练自己的数据生成新的图片

出于工作需要，学习了GAN，原理这块就不多讲了，主要讲怎么训练自己的数据生成新的图片，因为博客上大多是生成MNIST数据集,生成自己的图片时，有些小坑。

下面记录一下本人基于参考链接，将MNIST数据集的代码改成生成自己数据时遇到的坑。

一、读取数据问题

# MNIST dataset
mnist = datasets.MNIST(
   root='./data/', train=True, transform=img_transform, download=True)
# Data loader
dataloader = torch.utils.data.DataLoader(
   dataset=mnist, batch_size=batch_size, shuffle=True)

可以看到，datasets.MNIST这个肯定不能用于我们自己的数据。我借鉴了原来做二分类的datasets.ImageFolder。

发现老是报错：

RuntimeError: Found 0 files in subfolders of: E:\Projects\gan\battery\ng
Supported extensions are: .jpg,.jpeg,.png,.ppm,.bmp,.pgm,.tif,.tiff,.webp

后面单步调试，原来这个函数是需要文件夹下面有分类标签的，根据子文件夹名生成分类标签。

故放弃，只能自己写了。

下面是参考网上的，写了个读取数据的函数：

import numpy as np
import torch
import os
import random
from PIL import Image
from torch.utils.data import Dataset

class myDataset(Dataset):
   def __init__(self, data_dir, transform):

self.data_dir = data_dir
       self.transform = transform
       self.img_names = [name for name in list(filter(lambda x: x.endswith(".jpg"), os.listdir(self.data_dir)))]

def __getitem__(self, index):
       path_img = os.path.join(self.data_dir, self.img_names[index])
       img = Image.open(path_img).convert('RGB')

if self.transform is not None:
           img = self.transform(img)

return img

def __len__(self):
       if len(self.img_names) == 0:
           raise Exception("\ndata_dir:{} is a empty dir! Please checkout your path to images!".format(self.data_dir))
       return len(self.img_names)

二、维度不匹配问题

解决了读取数据之后，发现可以训练了，因为参考链接的MINIST数据都是单通道的，我们大部分图像都是3通道的，所以我将通道改为3后，发现判别器那块老是报错，标签和数据不匹配。

RuntimeError: mat1 dim 1 must match mat2 dim 0

后面一查，发现问题出在这句上面：

for i, (imgs, _) in enumerate(dataloader)

这样得到的imgs已经没有batch-size的信息了，需要改为这样：

for i, imgs in enumerate(dataloader):

下面是整个代码块，贴上去记录下来，以便过段时间万一忘了，还有个看的地方。

import argparse
import os
import numpy as np
import math

# import torchvision.transforms as transforms
from torchvision.utils import save_image

from torch.utils.data import DataLoader
from torchvision import datasets, models, transforms
from torch.autograd import Variable

import torch.nn as nn
import torch.nn.functional as F
from tools.my_dataset import myDataset
import torch

os.makedirs("images", exist_ok=True)

parser = argparse.ArgumentParser()
parser.add_argument("--n_epochs", type=int, default=50, help="number of epochs of training")
parser.add_argument("--batch_size", type=int, default=2, help="size of the batches")
parser.add_argument("--lr", type=float, default=0.0002, help="adam: learning rate")
parser.add_argument("--b1", type=float, default=0.5, help="adam: decay of first order momentum of gradient")
parser.add_argument("--b2", type=float, default=0.999, help="adam: decay of first order momentum of gradient")
parser.add_argument("--n_cpu", type=int, default=4, help="number of cpu threads to use during batch generation")
parser.add_argument("--latent_dim", type=int, default=100, help="dimensionality of the latent space")
parser.add_argument("--img_size", type=int, default=128, help="size of each image dimension")
parser.add_argument("--channels", type=int, default=3, help="number of image channels")
parser.add_argument("--sample_interval", type=int, default=400, help="interval betwen image samples")
opt = parser.parse_args()
print(opt)

img_shape = (opt.channels, opt.img_size, opt.img_size)

cuda = True if torch.cuda.is_available() else False
print('cuda is',cuda)

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

def block(in_feat, out_feat, normalize=True):
           layers = [nn.Linear(in_feat, out_feat)]
           if normalize:
               layers.append(nn.BatchNorm1d(out_feat, 0.8))
           layers.append(nn.LeakyReLU(0.2, inplace=True))
           return layers

self.model = nn.Sequential(
           *block(opt.latent_dim, 128, normalize=False),
           *block(128, 256),
           *block(256, 512),
           *block(512, 1024),
           nn.Linear(1024, int(np.prod(img_shape))),
           nn.Tanh()
       )

def forward(self, z):
       img = self.model(z)
       img = img.view(img.size(0), *img_shape)
       return img

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

self.model = nn.Sequential(
           nn.Linear(int(np.prod(img_shape)), 512),
           nn.LeakyReLU(0.2, inplace=True),
           nn.Linear(512, 256),
           nn.LeakyReLU(0.2, inplace=True),
           nn.Linear(256, 1),
           nn.Sigmoid(),
       )

def forward(self, img):
       img_flat = img.view(img.size(0), -1)
       validity = self.model(img_flat)

return validity

# Loss function
adversarial_loss = torch.nn.BCELoss()

# Initialize generator and discriminator
generator = Generator()
discriminator = Discriminator()

if cuda:
   generator.cuda()
   discriminator.cuda()
   adversarial_loss.cuda()

# # Configure data loader
# os.makedirs("./data/mnist", exist_ok=True)
# dataloader = torch.utils.data.DataLoader(
#     datasets.MNIST(
#         "./data/mnist",
#         train=True,
#         download=True,
#         transform=transforms.Compose(
#             [transforms.Resize(opt.img_size), transforms.ToTensor(), transforms.Normalize([0.5], [0.5])]
#         ),
#     ),
#     batch_size=opt.batch_size,
#     shuffle=True,
# )
dataset = r'E:\Projects\gan\battery'
ng_directory = os.path.join(dataset, 'ng')
ok_directory = os.path.join(dataset, 'ok')

image_transforms = {
   'ng': transforms.Compose([
       transforms.Resize([opt.img_size,opt.img_size]),
       transforms.ToTensor(),
       ]),

'ok': transforms.Compose([
       transforms.Resize([opt.img_size,opt.img_size]),
       transforms.ToTensor(),
      ])}

data = {
   'ng': myDataset(data_dir=ng_directory, transform=image_transforms['ng']),
   'ok': myDataset(data_dir=ok_directory, transform=image_transforms['ok'])
}

dataloader = DataLoader(data['ng'], batch_size=opt.batch_size, shuffle=True)
ng_data_size = len(data['ng'])
ok_data_size = len(data['ok'])
print('train_size: {:4d}  valid_size:{:4d}'.format(ng_data_size, ok_data_size))
# Optimizers
optimizer_G = torch.optim.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))

Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor

# ----------
#  Training
# ----------

for epoch in range(opt.n_epochs):
   # for i, (imgs, _) in enumerate(dataloader):
   for i, imgs in enumerate(dataloader):

# Adversarial ground truths
       valid = Variable(Tensor(imgs.size(0), 1).fill_(1.0), requires_grad=False)
       fake = Variable(Tensor(imgs.size(0), 1).fill_(0.0), requires_grad=False)

# Configure input
       real_imgs = Variable(imgs.type(Tensor))

# -----------------
       #  Train Generator
       # -----------------

optimizer_G.zero_grad()

# Sample noise as generator input
       z = Variable(Tensor(np.random.normal(0, 3, (imgs.shape[0], opt.latent_dim))))

# Generate a batch of images
       gen_imgs = generator(z)

# Loss measures generator's ability to fool the discriminator
       aa = discriminator(gen_imgs)
       g_loss = adversarial_loss(aa, valid)

g_loss.backward()
       optimizer_G.step()

# ---------------------
       #  Train Discriminator
       # ---------------------

optimizer_D.zero_grad()

# Measure discriminator's ability to classify real from generated samples
       bb = discriminator(real_imgs)
       real_loss = adversarial_loss(bb, valid)
       # 此处需要注意，detach()是为了截断梯度流，不计算生成网络的损失，
       # 因为d_loss包含了fake_loss，回传的时候如果不做处理，默认会计算generator的梯度，
       # 而这里只需要计算判别网络的梯度，更新其权重值，生成网络保持不变即可。
       fake_loss = adversarial_loss(discriminator(gen_imgs.detach()), fake)
       d_loss = (real_loss + fake_loss) / 2

d_loss.backward()
       optimizer_D.step()

print(
           "[Epoch ％d/％d] [Batch ％d/％d] [D loss: ％f] [G loss: ％f]"
           ％ (epoch, opt.n_epochs, i, len(dataloader), d_loss.item(), g_loss.item())
       )

batches_done = epoch * len(dataloader) + i
       if batches_done ％ opt.sample_interval == 0:
           save_image(gen_imgs.data[:25], "images/％d.png" ％ batches_done, nrow=5, normalize=True)

上面是原始图片，下面是生成的图片，从开始的噪声，到慢慢有点样子，还没训练完，由于我的显卡比较小，GTX1660Ti，6G显存，所以将原始图片从800x800压缩到了128x128，可能影响了效果，没关系，后面还可以优化，包括将全连接网络改为卷积的，图片设置大点，等等。

来源：https://blog.csdn.net/qq_36563273/article/details/119325318