G3学习笔记
- 🍨 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
准备工作
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
torch.manual_seed(1)
<torch._C.Generator at 0x23c3bf41230>
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
batch_size = 128
1. 导入数据
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=r'C:\Users\11054\Desktop\kLearning\G3\rps', transform=train_transform)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True,num_workers=6)
2. 数据可视化
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)
Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers). Got range [-0.92941177..1.0].
image_shape = (3, 128, 128)
image_dim = int(np.prod(image_shape))
latent_dim = 100
n_classes = 3
embedding_dim = 100
构建模型
# 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0.02torch.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0.02torch.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)
1. 构建生成器
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)
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())
)
from torchinfo import summarysummary(generator)
=================================================================
Layer (type:depth-idx) Param #
=================================================================
Generator --
├─Sequential: 1-1 --
│ └─Embedding: 2-1 300
│ └─Linear: 2-2 1,616
├─Sequential: 1-2 --
│ └─Linear: 2-3 827,392
│ └─LeakyReLU: 2-4 --
├─Sequential: 1-3 --
│ └─ConvTranspose2d: 2-5 4,202,496
│ └─BatchNorm2d: 2-6 1,024
│ └─ReLU: 2-7 --
│ └─ConvTranspose2d: 2-8 2,097,152
│ └─BatchNorm2d: 2-9 512
│ └─ReLU: 2-10 --
│ └─ConvTranspose2d: 2-11 524,288
│ └─BatchNorm2d: 2-12 256
│ └─ReLU: 2-13 --
│ └─ConvTranspose2d: 2-14 131,072
│ └─BatchNorm2d: 2-15 128
│ └─ReLU: 2-16 --
│ └─ConvTranspose2d: 2-17 3,072
│ └─Tanh: 2-18 --
=================================================================
Total params: 7,789,308
Trainable params: 7,789,308
Non-trainable params: 0
=================================================================
a = torch.ones(100)
b = torch.ones(1)
b = b.long()
a = a.to(device)
b = b.to(device)
# generator((a,b))
2. 构建鉴别器
import torch
import torch.nn as nnclass 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1nn.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2nn.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2nn.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2nn.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)
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())
)
summary(discriminator)
=================================================================
Layer (type:depth-idx) Param #
=================================================================
Discriminator --
├─Sequential: 1-1 --
│ └─Embedding: 2-1 300
│ └─Linear: 2-2 4,964,352
├─Sequential: 1-2 --
│ └─Conv2d: 2-3 6,144
│ └─LeakyReLU: 2-4 --
│ └─Conv2d: 2-5 131,072
│ └─BatchNorm2d: 2-6 256
│ └─LeakyReLU: 2-7 --
│ └─Conv2d: 2-8 524,288
│ └─BatchNorm2d: 2-9 512
│ └─LeakyReLU: 2-10 --
│ └─Conv2d: 2-11 2,097,152
│ └─BatchNorm2d: 2-12 1,024
│ └─LeakyReLU: 2-13 --
│ └─Flatten: 2-14 --
│ └─Dropout: 2-15 --
│ └─Linear: 2-16 4,609
│ └─Sigmoid: 2-17 --
=================================================================
Total params: 7,729,709
Trainable params: 7,729,709
Non-trainable params: 0
=================================================================
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()
torch.Size([2, 1])
训练模型
1. 定义损失函数
adversarial_loss = nn.BCELoss() def generator_loss(fake_output, label):gen_loss = adversarial_loss(fake_output, label)return gen_lossdef discriminator_loss(output, label):disc_loss = adversarial_loss(output, label)return disc_loss
2. 定义优化器
learning_rate = 0.0002G_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))
3. 训练模型
# 设置训练的总轮数
num_epochs = 20
# 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) / 2D_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], f'C:/Users/11054/Desktop/kLearning/G3/images/sample_{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(), f'C:/Users/11054/Desktop/kLearning/G3/generator_epoch_{epoch}.pth')torch.save(discriminator.state_dict(), f'C:/Users/11054/Desktop/kLearning/G3/discriminator_epoch_{epoch}.pth')
Epoch: [1/20]: D_loss: 0.328, G_loss: 1.420
Epoch: [2/20]: D_loss: 0.169, G_loss: 2.818
Epoch: [3/20]: D_loss: 0.261, G_loss: 2.547
Epoch: [4/20]: D_loss: 0.293, G_loss: 2.359
Epoch: [5/20]: D_loss: 0.301, G_loss: 2.415
Epoch: [6/20]: D_loss: 0.348, G_loss: 2.198
Epoch: [7/20]: D_loss: 0.346, G_loss: 2.292
Epoch: [8/20]: D_loss: 0.425, G_loss: 2.117
Epoch: [9/20]: D_loss: 0.470, G_loss: 1.596
模型分析
1. 加载模型
generator.load_state_dict(torch.load(f'C:/Users/11054/Desktop/kLearning/G3/generator_epoch_10.pth'), strict=False)
generator.eval()
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())
)
# 导入所需的库
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 * p2vectors.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) * labellabels = 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 = predelse:output = np.concatenate((output,pred))
torch.Size([10, 1])
torch.Size([10, 1])
torch.Size([10, 1])
output.shape
(30, 128, 128, 3)
nrow = 3
ncol = 10fig = 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 ) * 127.5pred = 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()
个人总结
CGAN 通过引入条件信息（Conditional Information） 来控制生成数据的类别或属性，从而增强生成数据的可控性和多样性，核心组件分析：
-
条件机制实现：
- 通过Embedding层将离散标签映射为连续向量（label_conditioned_generator/label_condition_disc）
- 使用Linear层调整维度后与图像特征拼接（forward中的concat操作）
-
生成器(Generator)关键设计：
- 输入：噪声向量( latent_dim ) + 类别标签( n_classes )
- 结构：5层转置卷积（ConvTranspose2d）实现上采样
- 输出：3通道RGB图像（Tanh激活归一化到[-1,1]）
- 创新点：将标签嵌入向量reshape为4x4特征图与噪声特征拼接
-
判别器(Discriminator)关键设计：
- 输入：真实/生成图像 + 对应标签特征
- 结构：4层卷积（Conv2d）实现下采样
- 输出：Sigmoid激活的单一概率值
- 特点：标签特征先转换为3x128x128伪图像再拼接
训练技巧观察：
-
参数配置：
- LeakyReLU(0.2)防止梯度消失
- BatchNorm momentum=0.1, eps=0.8 的非常规设置
- Dropout(0.4)防止判别器过拟合
-
维度变换：
- 生成器最终输出3x64x64图像（需注意与判别器输入尺寸匹配）
- 判别器Flatten后全连接层输入4608维（需根据输入尺寸计算）