深度学习-第G3周：CGAN｜生成手势图像

本周任务：

• 1.学习条件生成CGAN的基本原理
• 2.CGAN是如何实现条件控制的
• 3.学习本文CGAN的代码并跑通

一、CGAN的基本原理

cGAN的中心思想是希望 可以控制 GAN 生成的图片，而不 是单纯的随机生成图片。 具体来说，Conditional GAN 在生成器和判别器的输入中 增加了额外的 条件信息，生成器生成的图片只有足够真实 且与条件相符，才能够通过判别器。

      实际上 , 在无条件约束的生成模型中 , 没法控制数据生成的模式。然而，通过额外的信息对模型进行约束，有可能指导数据生成的过程。条件约束可以是类标签 , 可以是图像修补的部分数据, 甚至是来自不同模态的数据

cGAN将 无监督学习 转为 有监督学习 使得网络可以更好地在我们的掌控下进行学习！

从公式看，CGAN相当于在原始GAN的基础上对生成器部分 和判别器部分都加了一个条件
 

二、准备工作

# -*- coding:utf-8 -*-
import torch
import numpy as np
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
from torch.autograd import Variable
from torchvision.utils import save_image
from torchvision.utils import make_grid
from torch.utils.tensorboard import SummaryWriter
from torchsummary import summary
import matplotlib.pyplot as plt
import datetime
import os
os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
os.environ['CUBLAS_WORKSPACE_CONFIG']=':16:8'

manualSeed = 1 # 随机种子
print("Random Seed: ", manualSeed)
torch.manual_seed(manualSeed)
torch.use_deterministic_algorithms(True) # Needed for reproducible results

# 超参数
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
batch_size = 128

train_transform = transforms.Compose([
    transforms.Resize(128),
    transforms.ToTensor(),
    transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5])])

train_dataset = datasets.ImageFolder(root='./GAN-3-day-rps/rps/', transform=train_transform)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, 
                                           batch_size=batch_size, 
                                           shuffle=True,
                                           num_workers=6)


def show_images(images):
    fig, ax = plt.subplots(figsize=(20, 20))
    ax.set_xticks([]); ax.set_yticks([])
    ax.imshow(make_grid(images.detach(), nrow=22).permute(1, 2, 0))

def show_batch(dl):
    for images, _ in dl:
        show_images(images)
        break
    
# show_batch(train_loader)    

前期的数据准备如上所示，值得注意的是在

torch.use_deterministic_algorithms(True) 是会报错

RuntimeError: Deterministic behavior was enabled with either `torch.use_deterministic_algorithms(True)` or `at::Context::setDeterministicAlgorithms(true)`, but this operation is not deterministic because it uses CuBLAS and you have CUDA >= 10.2. To enable deterministic behavior in this case, you must set an environment variable before running your PyTorch application: CUBLAS_WORKSPACE_CONFIG=:4096:8 or CUBLAS_WORKSPACE_CONFIG=:16:8. For more information, go to https://docs.nvidia.com/cuda/cublas/index.html#cublasApi_reproducibility
 

根据官方的说明：如果CUDA版本为10.2或更高，则少数CUDA操作不确定，除非设置了环境变量 CUBLAS_WORKSPACE_CONFIG=:4096:8 或 CUBLAS_WORKSPACE_CONFIG=:16:8

但加os.environ['CUBLAS_WORKSPACE_CONFIG']=':16:8' 也不解决问题

后来添加os.environ['CUDA_LAUNCH_BLOCKING'] = '1'，完美解决

三、生成器与判别器

image_shape = (3, 128, 128)
image_dim = int(np.prod(image_shape))
latent_dim = 100

n_classes = 3
embedding_dim = 100

#
# 自定义权重初始化函数，用于初始化生成器和判别器的权重
def weights_init(m):
    # 获取当前层的类名
    classname = m.__class__.__name__

    # 如果当前层是卷积层（类名中包含 'Conv' ）
    if classname.find('Conv') != -1:
        # 使用正态分布随机初始化权重，均值为0，标准差为0.02
        torch.nn.init.normal_(m.weight, 0.0, 0.02)
    
    # 如果当前层是批归一化层（类名中包含 'BatchNorm' ）
    elif classname.find('BatchNorm') != -1:
        # 使用正态分布随机初始化权重，均值为1，标准差为0.02
        torch.nn.init.normal_(m.weight, 1.0, 0.02)
        # 将偏置项初始化为全零
        torch.nn.init.zeros_(m.bias)
        
class Generator(nn.Module):
    def __init__(self):
        super(Generator, self).__init__()

        # 定义条件标签的生成器部分，用于将标签映射到嵌入空间中
        # n_classes：条件标签的总数
        # embedding_dim：嵌入空间的维度
        self.label_conditioned_generator = nn.Sequential(
            nn.Embedding(n_classes, embedding_dim),  # 使用Embedding层将条件标签映射为稠密向量
            nn.Linear(embedding_dim, 16)             # 使用线性层将稠密向量转换为更高维度
        )

        # 定义潜在向量的生成器部分，用于将噪声向量映射到图像空间中
        # latent_dim：潜在向量的维度
        self.latent = nn.Sequential(
            nn.Linear(latent_dim, 4*4*512),  # 使用线性层将潜在向量转换为更高维度
            nn.LeakyReLU(0.2, inplace=True)  # 使用LeakyReLU激活函数进行非线性映射
        )

        # 定义生成器的主要结构，将条件标签和潜在向量合并成生成的图像
        self.model = nn.Sequential(
            # 反卷积层1：将合并后的向量映射为64x8x8的特征图
            nn.ConvTranspose2d(513, 64*8, 4, 2, 1, bias=False),
            nn.BatchNorm2d(64*8, momentum=0.1, eps=0.8),  # 批标准化
            nn.ReLU(True),  # ReLU激活函数
            # 反卷积层2：将64x8x8的特征图映射为64x4x4的特征图
            nn.ConvTranspose2d(64*8, 64*4, 4, 2, 1, bias=False),
            nn.BatchNorm2d(64*4, momentum=0.1, eps=0.8),
            nn.ReLU(True),
            # 反卷积层3：将64x4x4的特征图映射为64x2x2的特征图
            nn.ConvTranspose2d(64*4, 64*2, 4, 2, 1, bias=False),
            nn.BatchNorm2d(64*2, momentum=0.1, eps=0.8),
            nn.ReLU(True),
            # 反卷积层4：将64x2x2的特征图映射为64x1x1的特征图
            nn.ConvTranspose2d(64*2, 64*1, 4, 2, 1, bias=False),
            nn.BatchNorm2d(64*1, momentum=0.1, eps=0.8),
            nn.ReLU(True),
            # 反卷积层5：将64x1x1的特征图映射为3x64x64的RGB图像
            nn.ConvTranspose2d(64*1, 3, 4, 2, 1, bias=False),
            nn.Tanh()  # 使用Tanh激活函数将生成的图像像素值映射到[-1, 1]范围内
        )

    def forward(self, inputs):
        noise_vector, label = inputs
        # 通过条件标签生成器将标签映射为嵌入向量
        label_output = self.label_conditioned_generator(label)
        # 将嵌入向量的形状变为（batch_size, 1, 4, 4），以便与潜在向量进行合并
        label_output = label_output.view(-1, 1, 4, 4)
        # 通过潜在向量生成器将噪声向量映射为潜在向量
        latent_output = self.latent(noise_vector)
        # 将潜在向量的形状变为（batch_size, 512, 4, 4），以便与条件标签进行合并
        latent_output = latent_output.view(-1, 512, 4, 4)
        
        # 将条件标签和潜在向量在通道维度上进行合并，得到合并后的特征图
        concat = torch.cat((latent_output, label_output), dim=1)
        # 通过生成器的主要结构将合并后的特征图生成为RGB图像
        image = self.model(concat)
        return image        
        
generator = Generator().to(device)
generator.apply(weights_init)
print(generator)        
        
from torchinfo import summary

summary(generator)

# a = torch.ones(100)
# b = torch.ones(1)
# b = b.long()
# a = a.to(device)
# b = b.to(device)

import torch
import torch.nn as nn

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

        # 定义一个条件标签的嵌入层，用于将类别标签转换为特征向量
        self.label_condition_disc = nn.Sequential(
            nn.Embedding(n_classes, embedding_dim),     # 嵌入层将类别标签编码为固定长度的向量
            nn.Linear(embedding_dim, 3*128*128)         # 线性层将嵌入的向量转换为与图像尺寸相匹配的特征张量
        )
        
        # 定义主要的鉴别器模型
        self.model = nn.Sequential(
            nn.Conv2d(6, 64, 4, 2, 1, bias=False),       # 输入通道为6（包含图像和标签的通道数），输出通道为64，4x4的卷积核，步长为2，padding为1
            nn.LeakyReLU(0.2, inplace=True),             # LeakyReLU激活函数，带有负斜率，增加模型对输入中的负值的感知能力
            nn.Conv2d(64, 64*2, 4, 3, 2, bias=False),    # 输入通道为64，输出通道为64*2，4x4的卷积核，步长为3，padding为2
            nn.BatchNorm2d(64*2, momentum=0.1, eps=0.8),  # 批量归一化层，有利于训练稳定性和收敛速度
            nn.LeakyReLU(0.2, inplace=True),
            nn.Conv2d(64*2, 64*4, 4, 3, 2, bias=False),  # 输入通道为64*2，输出通道为64*4，4x4的卷积核，步长为3，padding为2
            nn.BatchNorm2d(64*4, momentum=0.1, eps=0.8),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Conv2d(64*4, 64*8, 4, 3, 2, bias=False),  # 输入通道为64*4，输出通道为64*8，4x4的卷积核，步长为3，padding为2
            nn.BatchNorm2d(64*8, momentum=0.1, eps=0.8),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Flatten(),                               # 将特征图展平为一维向量，用于后续全连接层处理
            nn.Dropout(0.4),                            # 随机失活层，用于减少过拟合风险
            nn.Linear(4608, 1),                         # 全连接层，将特征向量映射到输出维度为1的向量
            nn.Sigmoid()                                # Sigmoid激活函数，用于输出范围限制在0到1之间的概率值
        )

    def forward(self, inputs):
        img, label = inputs
        
        # 将类别标签转换为特征向量
        label_output = self.label_condition_disc(label)
        # 重塑特征向量为与图像尺寸相匹配的特征张量
        label_output = label_output.view(-1, 3, 128, 128)
        
        # 将图像特征和标签特征拼接在一起作为鉴别器的输入
        concat = torch.cat((img, label_output), dim=1)
        
        # 将拼接后的输入通过鉴别器模型进行前向传播，得到输出结果
        output = self.model(concat)
        return output

discriminator = Discriminator().to(device)
discriminator.apply(weights_init)
print(discriminator)

summary(discriminator)

# a = torch.ones(2,3,128,128)
# b = torch.ones(2,1)
# b = b.long()
# a = a.to(device)
# b = b.to(device)
# c = discriminator((a,b))
# c.size()

adversarial_loss = nn.BCELoss() 

def generator_loss(fake_output, label):
    gen_loss = adversarial_loss(fake_output, label)
    return gen_loss

def discriminator_loss(output, label):
    disc_loss = adversarial_loss(output, label)
    return disc_loss

learning_rate = 0.0002

G_optimizer = optim.Adam(generator.parameters(),     lr = learning_rate, betas=(0.5, 0.999))
D_optimizer = optim.Adam(discriminator.parameters(), lr = learning_rate, betas=(0.5, 0.999))

Generator(
  (label_conditioned_generator): Sequential(
    (0): Embedding(3, 100)
    (1): Linear(in_features=100, out_features=16, bias=True)
  )
  (latent): Sequential(
    (0): Linear(in_features=100, out_features=8192, bias=True)
    (1): LeakyReLU(negative_slope=0.2, inplace=True)
  )
  (model): Sequential(
    (0): ConvTranspose2d(513, 512, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False)
    (1): BatchNorm2d(512, eps=0.8, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU(inplace=True)
    (3): ConvTranspose2d(512, 256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False)
    (4): BatchNorm2d(256, eps=0.8, momentum=0.1, affine=True, track_running_stats=True)
    (5): ReLU(inplace=True)
    (6): ConvTranspose2d(256, 128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False)
    (7): BatchNorm2d(128, eps=0.8, momentum=0.1, affine=True, track_running_stats=True)
    (8): ReLU(inplace=True)
    (9): ConvTranspose2d(128, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False)
    (10): BatchNorm2d(64, eps=0.8, momentum=0.1, affine=True, track_running_stats=True)
    (11): ReLU(inplace=True)
    (12): ConvTranspose2d(64, 3, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False)
    (13): Tanh()
  )
)
Discriminator(
  (label_condition_disc): Sequential(
    (0): Embedding(3, 100)
    (1): Linear(in_features=100, out_features=49152, bias=True)
  )
  (model): Sequential(
    (0): Conv2d(6, 64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), bias=False)
    (1): LeakyReLU(negative_slope=0.2, inplace=True)
    (2): Conv2d(64, 128, kernel_size=(4, 4), stride=(3, 3), padding=(2, 2), bias=False)
    (3): BatchNorm2d(128, eps=0.8, momentum=0.1, affine=True, track_running_stats=True)
    (4): LeakyReLU(negative_slope=0.2, inplace=True)
    (5): Conv2d(128, 256, kernel_size=(4, 4), stride=(3, 3), padding=(2, 2), bias=False)
    (6): BatchNorm2d(256, eps=0.8, momentum=0.1, affine=True, track_running_stats=True)
    (7): LeakyReLU(negative_slope=0.2, inplace=True)
    (8): Conv2d(256, 512, kernel_size=(4, 4), stride=(3, 3), padding=(2, 2), bias=False)
    (9): BatchNorm2d(512, eps=0.8, momentum=0.1, affine=True, track_running_stats=True)
    (10): LeakyReLU(negative_slope=0.2, inplace=True)
    (11): Flatten(start_dim=1, end_dim=-1)
    (12): Dropout(p=0.4, inplace=False)
    (13): Linear(in_features=4608, out_features=1, bias=True)
    (14): Sigmoid()
  )
)

四、训练模型

训练代码逻辑图

# 设置训练的总轮数
num_epochs = 100
# 初始化用于存储每轮训练中判别器和生成器损失的列表
D_loss_plot, G_loss_plot = [], []

# 循环进行训练
for epoch in range(1, num_epochs + 1):
    
    # 初始化每轮训练中判别器和生成器损失的临时列表
    D_loss_list, G_loss_list = [], []
    
    # 遍历训练数据加载器中的数据
    for index, (real_images, labels) in enumerate(train_loader):
        # 清空判别器的梯度缓存
        D_optimizer.zero_grad()
        # 将真实图像数据和标签转移到GPU（如果可用）
        real_images = real_images.to(device)
        labels      = labels.to(device)
        
        # 将标签的形状从一维向量转换为二维张量（用于后续计算）
        labels = labels.unsqueeze(1).long()
        # 创建真实目标和虚假目标的张量（用于判别器损失函数）
        real_target = Variable(torch.ones(real_images.size(0), 1).to(device))
        fake_target = Variable(torch.zeros(real_images.size(0), 1).to(device))

        # 计算判别器对真实图像的损失
        D_real_loss = discriminator_loss(discriminator((real_images, labels)), real_target)
        
        # 从噪声向量中生成假图像（生成器的输入）
        noise_vector = torch.randn(real_images.size(0), latent_dim, device=device)
        noise_vector = noise_vector.to(device)
        generated_image = generator((noise_vector, labels))
        
        # 计算判别器对假图像的损失（注意detach()函数用于分离生成器梯度计算图）
        output = discriminator((generated_image.detach(), labels))
        D_fake_loss = discriminator_loss(output, fake_target)

        # 计算判别器总体损失（真实图像损失和假图像损失的平均值）
        D_total_loss = (D_real_loss + D_fake_loss) / 2
        D_loss_list.append(D_total_loss)

        # 反向传播更新判别器的参数
        D_total_loss.backward()
        D_optimizer.step()

        # 清空生成器的梯度缓存
        G_optimizer.zero_grad()
        # 计算生成器的损失
        G_loss = generator_loss(discriminator((generated_image, labels)), real_target)
        G_loss_list.append(G_loss)

        # 反向传播更新生成器的参数
        G_loss.backward() 
        G_optimizer.step()

    # 打印当前轮次的判别器和生成器的平均损失
    print('Epoch: [%d/%d]: D_loss: %.3f, G_loss: %.3f' % (
            (epoch), num_epochs, torch.mean(torch.FloatTensor(D_loss_list)), 
            torch.mean(torch.FloatTensor(G_loss_list))))
    
    # 将当前轮次的判别器和生成器的平均损失保存到列表中
    D_loss_plot.append(torch.mean(torch.FloatTensor(D_loss_list)))
    G_loss_plot.append(torch.mean(torch.FloatTensor(G_loss_list)))

    if epoch%10 == 0:
        # 将生成的假图像保存为图片文件
        save_image(generated_image.data[:50], './images/sample_%d' % epoch + '.png', nrow=5, normalize=True)
        # 将当前轮次的生成器和判别器的权重保存到文件
        torch.save(generator.state_dict(), './training_weights/generator_epoch_%d.pth' % (epoch))
        torch.save(discriminator.state_dict(), './training_weights/discriminator_epoch_%d.pth' % (epoch))


generator.load_state_dict(torch.load('./training_weights/generator_epoch_100.pth'), strict=False)
generator.eval()  

Epoch: [1/100]: D_loss: 0.332, G_loss: 1.457
Epoch: [2/100]: D_loss: 0.166, G_loss: 3.010
Epoch: [3/100]: D_loss: 0.222, G_loss: 2.803
Epoch: [4/100]: D_loss: 0.323, G_loss: 2.267
Epoch: [5/100]: D_loss: 0.245, G_loss: 2.481
...
Epoch: [96/100]: D_loss: 0.284, G_loss: 2.552
Epoch: [97/100]: D_loss: 0.280, G_loss: 2.528
Epoch: [98/100]: D_loss: 0.277, G_loss: 2.525
Epoch: [99/100]: D_loss: 0.299, G_loss: 2.532
Epoch: [100/100]: D_loss: 0.298, G_loss: 2.603

五、模型分析

from numpy import asarray
from numpy.random import randn
from numpy.random import randint
from numpy import linspace
from matplotlib import pyplot
from matplotlib import gridspec

# 生成潜在空间的点，作为生成器的输入
def generate_latent_points(latent_dim, n_samples, n_classes=3):
    # 从标准正态分布中生成潜在空间的点
    x_input = randn(latent_dim * n_samples)
    # 将生成的点整形成用于神经网络的输入的批量
    z_input = x_input.reshape(n_samples, latent_dim)
    return z_input

# 在两个潜在空间点之间进行均匀插值
def interpolate_points(p1, p2, n_steps=10):
    # 在两个点之间进行插值，生成插值比率
    ratios = linspace(0, 1, num=n_steps)
    # 线性插值向量
    vectors = list()
    for ratio in ratios:
        v = (1.0 - ratio) * p1 + ratio * p2
        vectors.append(v)
    return asarray(vectors)

# 生成两个潜在空间的点
pts = generate_latent_points(100, 2)
# 在两个潜在空间点之间进行插值
interpolated = interpolate_points(pts[0], pts[1])

# 将数据转换为torch张量并将其移至GPU（假设device已正确声明为GPU）
interpolated = torch.tensor(interpolated).to(device).type(torch.float32)

output = None
# 对于三个类别的循环，分别进行插值和生成图片
for label in range(3):
    # 创建包含相同类别标签的张量
    labels = torch.ones(10) * label
    labels = labels.to(device)
    labels = labels.unsqueeze(1).long()
    print(labels.size())
    # 使用生成器生成插值结果
    predictions = generator((interpolated, labels))
    predictions = predictions.permute(0,2,3,1)
    pred = predictions.detach().cpu()
    if output is None:
        output = pred
    else:
        output = np.concatenate((output,pred))

output.shape

nrow = 3
ncol = 10

fig = plt.figure(figsize=(15,4))
gs = gridspec.GridSpec(nrow, ncol) 

k = 0
for i in range(nrow):
    for j in range(ncol):
        pred = (output[k, :, :, :] + 1 ) * 127.5
        pred = np.array(pred)  
        ax= plt.subplot(gs[i,j])
        ax.imshow(pred.astype(np.uint8))
        ax.set_xticklabels([])
        ax.set_yticklabels([])
        ax.axis('off')
        k += 1   

plt.show()