245 lines
8.4 KiB
Python
245 lines
8.4 KiB
Python
import os
|
||
import torch
|
||
import torch.nn as nn
|
||
from torchvision import transforms, utils
|
||
from torch.utils.data import DataLoader, Dataset
|
||
from tqdm import tqdm
|
||
import torch.multiprocessing
|
||
from PIL import Image # 自定义数据集需要 import PIL
|
||
import matplotlib.pyplot as plt
|
||
|
||
|
||
# 判别器
|
||
class Detector(nn.Module):
|
||
def __init__(self):
|
||
super(Detector, self).__init__()
|
||
self.model = nn.Sequential(
|
||
# 3 x 64 x 64 -> 64 x 32 x 32
|
||
nn.Conv2d(3, 64, 4, 2, 1, bias=False),
|
||
nn.LeakyReLU(0.2, inplace=True),
|
||
|
||
# 64 x 32 x 32 -> 128 x 16 x 16
|
||
nn.Conv2d(64, 128, 4, 2, 1, bias=False),
|
||
nn.BatchNorm2d(128),
|
||
nn.LeakyReLU(0.2, inplace=True),
|
||
|
||
# 128 x 16 x 16 -> 256 x 8 x 8
|
||
nn.Conv2d(128, 256, 4, 2, 1, bias=False),
|
||
nn.BatchNorm2d(256),
|
||
nn.LeakyReLU(0.2, inplace=True),
|
||
|
||
# 256 x 8 x 8 -> 512 x 4 x 4
|
||
nn.Conv2d(256, 512, 4, 2, 1, bias=False),
|
||
nn.BatchNorm2d(512),
|
||
nn.LeakyReLU(0.2, inplace=True),
|
||
|
||
# 512 x 4 x 4 -> 1024 x 2 x 2
|
||
nn.Conv2d(512, 1024, 4, 2, 1, bias=False),
|
||
nn.BatchNorm2d(1024),
|
||
nn.LeakyReLU(0.2, inplace=True),
|
||
|
||
# 1024 x 2 x 2 -> 1 x 1 x 1
|
||
nn.Conv2d(1024, 1, 2, 1, 0, bias=False),
|
||
nn.Sigmoid()
|
||
)
|
||
|
||
def forward(self, x):
|
||
return self.model(x)
|
||
|
||
|
||
# 生成器
|
||
class Generator(nn.Module):
|
||
def __init__(self, z_dim=100):
|
||
super(Generator, self).__init__()
|
||
self.model = nn.Sequential(
|
||
# z -> 1024 x 4 x 4
|
||
nn.ConvTranspose2d(z_dim, 1024, 4, 1, 0, bias=False),
|
||
nn.BatchNorm2d(1024),
|
||
nn.ReLU(True),
|
||
|
||
# 1024 x 4 x 4 -> 512 x 8 x 8
|
||
nn.ConvTranspose2d(1024, 512, 4, 2, 1, bias=False),
|
||
nn.BatchNorm2d(512),
|
||
nn.ReLU(True),
|
||
|
||
# 512 x 8 x 8 -> 256 x 16 x 16
|
||
nn.ConvTranspose2d(512, 256, 4, 2, 1, bias=False),
|
||
nn.BatchNorm2d(256),
|
||
nn.ReLU(True),
|
||
|
||
# 256 x 16 x 16 -> 128 x 32 x 32
|
||
nn.ConvTranspose2d(256, 128, 4, 2, 1, bias=False),
|
||
nn.BatchNorm2d(128),
|
||
nn.ReLU(True),
|
||
|
||
# 128 x 32 x 32 -> 3 x 64 x 64
|
||
nn.ConvTranspose2d(128, 3, 4, 2, 1, bias=False),
|
||
nn.Tanh()
|
||
)
|
||
|
||
def forward(self, z):
|
||
return self.model(z)
|
||
|
||
|
||
# 数据集加载（使用自定义路径 ./data/images）
|
||
class FlatImageDataset(Dataset):
|
||
def __init__(self, root_dir, transform=None):
|
||
self.root_dir = root_dir
|
||
self.transform = transform
|
||
if not os.path.exists(root_dir):
|
||
raise FileNotFoundError(f"Dataset directory '{root_dir}' not found. Please create it and add images.")
|
||
self.image_files = [f for f in os.listdir(root_dir) if f.lower().endswith(('.png', '.jpg', '.jpeg', '.bmp'))]
|
||
if len(self.image_files) == 0:
|
||
raise ValueError(f"No valid image files found in '{root_dir}'. Supported: .png, .jpg, .jpeg, .bmp")
|
||
self.image_paths = [os.path.join(root_dir, f) for f in self.image_files]
|
||
|
||
def __len__(self):
|
||
return len(self.image_paths)
|
||
|
||
def __getitem__(self, idx):
|
||
img_path = self.image_paths[idx]
|
||
image = Image.open(img_path).convert('RGB') # 确保转为 RGB（3 通道）
|
||
if self.transform:
|
||
image = self.transform(image)
|
||
return image, 0 # 返回图像和虚拟标签（GAN 不使用）
|
||
|
||
|
||
# 使用自定义数据集
|
||
transform = transforms.Compose([
|
||
transforms.Resize(64),
|
||
transforms.CenterCrop(64),
|
||
transforms.ToTensor(),
|
||
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
|
||
])
|
||
|
||
# 加载自定义数据集
|
||
dataset = FlatImageDataset(root_dir='./data/images', transform=transform)
|
||
print(f"Loaded {len(dataset)} images from ./data/images") # 调试：打印数据集大小
|
||
dataloader = DataLoader(dataset, batch_size=64, shuffle=True, num_workers=2) # 减小 num_workers 避免 Windows 问题
|
||
|
||
|
||
# 参数
|
||
z_dim = 100
|
||
num_epochs = 100 # 6.2w
|
||
lr_d = 0.001
|
||
lr_g = 0.002
|
||
n_g_steps = 2 # 标准 DCGAN是1步G
|
||
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
||
print(f"Using device: {device}")
|
||
|
||
# 模型与优化器
|
||
d = Detector().to(device)
|
||
g = Generator(z_dim=z_dim).to(device)
|
||
criterion = nn.BCELoss()
|
||
optimizer_d = torch.optim.Adam(d.parameters(), lr=lr_d, betas=(0.5, 0.999))
|
||
optimizer_g = torch.optim.Adam(g.parameters(), lr=lr_g, betas=(0.5, 0.999))
|
||
|
||
# 固定噪声用于观察训练过程
|
||
z_fixed = torch.randn(64, z_dim, 1, 1, device=device)
|
||
|
||
# 创建保存目录
|
||
os.makedirs("results", exist_ok=True)
|
||
|
||
|
||
# 根据保存的 G dict 生成图片的函数
|
||
def generate_from_g_dict(model_path, z_dim=100, num_images=64, output_path='generated.png'):
|
||
"""
|
||
从保存的生成器 state_dict 文件加载模型，并生成图片保存。
|
||
"""
|
||
if not os.path.exists(model_path):
|
||
print(f"Model path '{model_path}' not found. Skipping generation.")
|
||
return
|
||
# 加载生成器并恢复权重
|
||
g_loaded = Generator(z_dim=z_dim).to(device)
|
||
g_loaded.load_state_dict(torch.load(model_path, map_location=device))
|
||
g_loaded.eval()
|
||
|
||
# 生成假图像
|
||
with torch.no_grad():
|
||
z = torch.randn(num_images, z_dim, 1, 1, device=device)
|
||
fake_images = g_loaded(z).detach().cpu()
|
||
|
||
# 保存图片
|
||
utils.save_image(fake_images, output_path, normalize=True, nrow=8)
|
||
print(f"Generated images saved to {output_path}")
|
||
|
||
|
||
# 训练循环
|
||
def train():
|
||
for epoch in range(1, num_epochs + 1):
|
||
loss_d_total, loss_g_total = 0, 0
|
||
for real_images, _ in tqdm(dataloader, desc=f"Epoch {epoch}/{num_epochs}", leave=False):
|
||
real_images = real_images.to(device)
|
||
B = real_images.size(0)
|
||
|
||
# 标签平滑（真实 = 0.9, 假 = 0.0）
|
||
real_labels = torch.full((B, 1, 1, 1), 0.9, device=device)
|
||
fake_labels = torch.full((B, 1, 1, 1), 0.0, device=device)
|
||
|
||
# 生成假图像
|
||
z = torch.randn(B, z_dim, 1, 1, device=device)
|
||
fake_images = g(z)
|
||
|
||
# 判别器训练
|
||
output_real = d(real_images)
|
||
output_fake = d(fake_images.detach())
|
||
|
||
loss_real = criterion(output_real, real_labels)
|
||
loss_fake = criterion(output_fake, fake_labels)
|
||
loss_d = loss_real + loss_fake
|
||
|
||
optimizer_d.zero_grad()
|
||
loss_d.backward()
|
||
optimizer_d.step()
|
||
|
||
# 生成器训练（标准 BCE）
|
||
for _ in range(n_g_steps):
|
||
#每次生成器更新前重新生成假图像
|
||
z = torch.randn(B, z_dim, 1, 1, device=device)
|
||
fake_images = g(z)
|
||
|
||
output = d(fake_images)
|
||
loss_g = criterion(output, real_labels) # 欺骗 D：希望 D 输出真
|
||
|
||
optimizer_g.zero_grad()
|
||
loss_g.backward()
|
||
optimizer_g.step()
|
||
loss_g_total += loss_g.item()
|
||
|
||
loss_d_total += loss_d.item()
|
||
|
||
# 平均损失（G 损失已累加 n_g_steps 次）
|
||
avg_loss_d = loss_d_total / len(dataloader)
|
||
avg_loss_g = loss_g_total / (len(dataloader) * n_g_steps) # 修复：除以总 G 步数
|
||
print(f"Epoch [{epoch}/{num_epochs}] Loss_D: {avg_loss_d:.4f} Loss_G: {avg_loss_g:.4f}")
|
||
|
||
loss_history = {"D": [], "G": []}
|
||
|
||
# 每轮结束时：
|
||
loss_history["D"].append(avg_loss_d)
|
||
loss_history["G"].append(avg_loss_g)
|
||
|
||
# 最后画图：
|
||
plt.plot(loss_history["D"], label="Loss_D")
|
||
plt.plot(loss_history["G"], label="Loss_G")
|
||
plt.legend()
|
||
plt.savefig("results/loss_curve.png")
|
||
|
||
# 每2轮保存一次生成图像和 G 的 state_dict
|
||
if epoch % 2 == 0:
|
||
with torch.no_grad():
|
||
fake = g(z_fixed).detach().cpu()
|
||
utils.save_image(fake, f"results/epoch_{epoch}.png", normalize=True, nrow=8)
|
||
|
||
# 保存 G 的 state_dict（dict 形式）
|
||
g_state_dict_path = f"results/g_epoch_{epoch}.pth"
|
||
torch.save(g.state_dict(), g_state_dict_path)
|
||
print(f"Generator state_dict saved to {g_state_dict_path}")
|
||
|
||
|
||
if __name__ == "__main__":
|
||
torch.multiprocessing.freeze_support()
|
||
#train()
|
||
for i in range(100):
|
||
text = 'results/generated_after_train' + str(i) + '.png'
|
||
generate_from_g_dict('results/g_epoch_85.pth', output_path=text) |