G9 - ACGAN理论与实战

• 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
• 🍖 原作者：K同学啊

---

---

上周已经简单的了解了ACGAN的原理，并且不经实践的编写了部分代码，这周复现一下真正的ACGAN

环境

Pytorch： 2.3.1+cu121
Nvidia GTX 4090

步骤

环境设置

import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import datasets, transforms
from torchvision.utils import save_imagefrom torch.utils.data import DataLoader
from torch.autograd import Variable
import numpy as npdevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')# 全局参数
n_epochs = 200
batch_size = 64
lr = 0.0002
b1 = 0.5
b2 = 0.999
n_cpu = 8
latent_dim = 100
n_classes = 10
img_size = 32
channels = 1
sample_interval = 400

数据准备

# 创建中间采样图片的文件夹
import os
os.makedirs('images', exist_ok=True)
# 配置数据集
os.makedirs('data/mnist', exist_ok=True)
dataloader = DataLoader(datasets.MNIST('data/mnist',train=True,download=True,transform=transforms.Compose([transforms.Resize(img_size),transforms.ToTensor(),transforms.Normalize([0.5], [0.5])]),),batch_size=batch_size,shuffle=True,
)

工具方法

# 权重初始化函数
def weights_init_normal(m):classname = m.__class__.__name__if classname.find('Conv') != -1:torch.nn.init.normal_(m.weight.data, 0.0, 0.02)elif classname.find('BatchNorm2d') != -1:torch.nn.init.normal_(m.weight.data, 1.0, 0.02)torch.nn.init.constant_(m.bias.data, 0.0)# 日志函数 因为使用了jupyter notebook环境，长时间的任务日志无法直接查看，于是需要打印到文件
import logging
import sys
import datetimedef init_logger(filename, logger_name):'''@brief:initialize logger that redirect info to a file just in case we lost connection to the notebook@params:filename: to which file should we log all the infologger_name: an alias to the logger'''# get current timestamptimestamp = datetime.datetime.utcnow().strftime('%Y%m%d_%H-%M-%S')logging.basicConfig(level=logging.INFO, format='[%(asctime)s] %(name)s {%(filename)s:%(lineno)d} %(levelname)s - %(message)s',handlers=[logging.FileHandler(filename=filename),logging.StreamHandler(sys.stdout)])# Testlogger = logging.getLogger(logger_name)logger.info('### Init. Logger {} ###'.format(logger_name))return logger# Initialize
my_logger = init_logger("./ml_notebook.log", "ml_logger")# 生成函数的结果保存
def sample_image(n_row, batches_done):"""保存从0到n_classes的生成数字的图像风格"""# 采样噪声z = torch.randn((n_row**2, latent_dim), device=device)# 为n行生成标签从0到n_classeslabels = torch.tensor([num for _ in range(n_row) for num in range(n_row)], device=device)gen_imgs = generator(z, labels)save_image(gen_imgs.data.cpu(), 'images/%d.png' % batches_done, nrow=n_row, normalize=True)

模型设计

# 生成器
class Generator(nn.Module):def __init__(self):super().__init__()# 标签嵌入self.label_emb = nn.Embedding(n_classes, latent_dim)# 计算上采样前的初始大小self.init_size = img_size // 4# 第一层线性层self.l1 = nn.Sequential(nn.Linear(latent_dim, 128*self.init_size**2))# 卷积层self.conv_blocks = nn.Sequential(nn.BatchNorm2d(128),nn.Upsample(scale_factor=2),nn.Conv2d(128, 128, 3, stride=1, padding=1),nn.BatchNorm2d(128, 0.8),nn.LeakyReLU(0.2, inplace=True),nn.Upsample(scale_factor=2),nn.Conv2d(128, 64, 3, stride=1, padding=1),nn.BatchNorm2d(64, 0.8),nn.LeakyReLU(0.2, inplace=True),nn.Conv2d(64, channels, 3, stride=1, padding=1),nn.Tanh(),)def forward(self, noise, labels):# 标签嵌入到噪声中gen_input = torch.mul(self.label_emb(labels), noise)# 通过第一层线性层out = self.l1(gen_input)# 整形out = out.view(out.shape[0], 128, self.init_size, self.init_size)# 卷积生成图像img = self.conv_blocks(out)return img
# 判别器
class Discriminator(nn.Module):def __init__(self):super().__init__()# 判别器块生成函数def discriminator_block(in_filters, out_filters, bn=True):"""返回每个判别器层"""block = [nn.Conv2d(in_filters, out_filters, 3, 2, 1), nn.LeakyReLU(0.2, inplace=True), nn.Dropout2d(0.25)]if bn:block.append(nn.BatchNorm2d(out_filters, 0.8))return block# 卷积层self.conv_blocks = nn.Sequential(*discriminator_block(channels, 16, bn=False),*discriminator_block(16, 32),*discriminator_block(32, 64),*discriminator_block(64, 128),)# 下采样后，图像的宽高ds_size = img_size // 2 ** 4# 输出层self.adv_layer = nn.Sequential(nn.Linear(128 * ds_size ** 2, 1), nn.Sigmoid())self.aux_layer = nn.Sequential(nn.Linear(128 * ds_size ** 2, n_classes), nn.Softmax())def forward(self, img):out = self.conv_blocks(img)out = out.view(out.shape[0], -1)validity = self.adv_layer(out)label = self.aux_layer(out)return validity, label# 模型初始化# 损失函数
adversarial_loss = nn.BCELoss()
auxiliary_loss = nn.CrossEntropyLoss()# 初始化生成器和判别器
generator = Generator().to(device)
discriminator = Discriminator().to(device)# 初始化权重
generator.apply(weights_init_normal)
discriminator.apply(weights_init_normal)

模型训练

# 训练# 优化器
optimizer_G = torch.optim.Adam(generator.parameters(), lr=lr, betas=(b1, b2))
optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=lr, betas=(b1, b2))for epoch in range(n_epochs):for i, (imgs, labels) in enumerate(dataloader):batch_size = imgs.shape[0]# 图像是 真实的 标签valid = torch.ones((batch_size, 1), requires_grad=False, device=device)# 图像是 生成的 标签fake = torch.zeros((batch_size, 1), requires_grad=False, device=device)real_imgs = imgs.to(device)labels = labels.to(device)# 训练生成器optimizer_G.zero_grad()# 采样噪声和标签作为生成器的输入z = torch.randn((batch_size, latent_dim), device=device)gen_labels = torch.randint(0, 1, (batch_size,), device=device)# 生成一批图像gen_imgs = generator(z, gen_labels)# 损失度量 生成器欺骗判别器的能力validity, pred_label = discriminator(gen_imgs)g_loss = 0.5 * (adversarial_loss(validity, valid) + auxiliary_loss(pred_label, gen_labels))g_loss.backward()optimizer_G.step()# 训练判别器optimizer_D.zero_grad()# 真实图像的损失real_pred, real_aux = discriminator(real_imgs)d_real_loss = 0.5 * (adversarial_loss(real_pred, valid) + auxiliary_loss(real_aux, labels))# 生成图像的损失fake_pred, fake_aux = discriminator(gen_imgs.detach())d_fake_loss = 0.5 * (adversarial_loss(fake_pred, fake) + auxiliary_loss(fake_aux, gen_labels))# 判别器的总损失d_loss = 0.5 * (d_real_loss + d_fake_loss)# 计算判别器的准确率pred = np.concatenate([real_aux.data.cpu().numpy(), fake_aux.data.cpu().numpy()], axis=0)gt = np.concatenate([labels.data.cpu().numpy(), gen_labels.data.cpu().numpy()], axis=0)d_acc = np.mean(np.argmax(pred, axis=1) == gt)d_loss.backward()optimizer_D.step()if i % 100 == 0:my_logger.info("[Epoch %d/%d] [Batch %d/%d] [D loss: %f, acc: %d%%] [G loss: %f]" % (epoch, n_epochs, i, len(dataloader), d_loss.item(), 100 * d_acc, g_loss.item()))batches_done = epoch * len(dataloader) + iif batches_done % sample_interval == 0:sample_image(n_row=10, batches_done=batches_done)

模型效果展示

总结与心得体会

通过对模型的复现，发现我之前对判别器的理解有偏差，如果在判别器的输入中插入分类信息，等于是将答案直接给了判别器，生成的结果反而不会太好。还有一个和我预想的不一样的地方，在生成器中，将标签嵌入到特征向量使用了矩阵乘法，而没有直接使用concatenate操作。