이전: TensorBoard 시각화 | 다음: 생성적 적대 신경망(GAN)

28. 생성 모델 - GAN (Generative Adversarial Networks)¶

학습 목표¶

GAN의 기본 원리와 적대적 학습 이해
Generator와 Discriminator 구조 설계
다양한 손실 함수 (Adversarial, Wasserstein, WGAN-GP)
DCGAN 아키텍처 구현
학습 안정화 기법 적용
StyleGAN 개념 이해

1. GAN 기초 이론¶

개념¶

GAN = 두 신경망의 경쟁적 학습

Generator (G): 가짜 데이터 생성
    랜덤 노이즈 z → 가짜 이미지

Discriminator (D): 진짜/가짜 판별
    이미지 → 진짜(1) / 가짜(0)

목표: G가 D를 속일 수 있을 만큼 좋은 가짜 생성

적대적 학습 (Adversarial Training)¶

┌────────────┐     ┌────────────┐
│  Noise z   │────▶│ Generator  │───┬──▶ 가짜 이미지
└────────────┘     └────────────┘   │
                                    │
┌────────────┐                      ▼
│ 진짜 이미지│────────────────────▶ Discriminator ──▶ 진짜/가짜
└────────────┘

D 학습: 진짜=1, 가짜=0 판별 정확도 최대화
G 학습: D가 가짜를 진짜로 판별하도록 유도

Min-Max 게임¶

# GAN 목적 함수 (min-max game)
# min_G max_D V(D, G) = E[log D(x)] + E[log(1 - D(G(z)))]

# D의 목표: V(D, G) 최대화
#   - D(x) → 1 (진짜를 진짜로)
#   - D(G(z)) → 0 (가짜를 가짜로)

# G의 목표: V(D, G) 최소화
#   - D(G(z)) → 1 (D가 가짜를 진짜로 판단하게)

2. 기본 GAN 구현¶

Generator¶

import torch
import torch.nn as nn

class Generator(nn.Module):
    """간단한 Generator (⭐⭐)"""
    def __init__(self, latent_dim=100, img_shape=(1, 28, 28)):
        super().__init__()
        self.img_shape = img_shape

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

        self.model = nn.Sequential(
            *block(latent_dim, 128, normalize=False),
            *block(128, 256),
            *block(256, 512),
            *block(512, 1024),
            nn.Linear(1024, int(torch.prod(torch.tensor(img_shape)))),
            nn.Tanh()  # 출력: [-1, 1]
        )

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

Discriminator¶

class Discriminator(nn.Module):
    """간단한 Discriminator (⭐⭐)"""
    def __init__(self, img_shape=(1, 28, 28)):
        super().__init__()

        self.model = nn.Sequential(
            nn.Flatten(),
            nn.Linear(int(torch.prod(torch.tensor(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()  # 출력: [0, 1] 확률
        )

    def forward(self, img):
        return self.model(img)

학습 루프¶

def train_gan(generator, discriminator, dataloader, epochs=100, latent_dim=100):
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    generator.to(device)
    discriminator.to(device)

    criterion = nn.BCELoss()

    optimizer_G = torch.optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999))
    optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=0.0002, betas=(0.5, 0.999))

    for epoch in range(epochs):
        for i, (real_imgs, _) in enumerate(dataloader):
            batch_size = real_imgs.size(0)
            real_imgs = real_imgs.to(device)

            # 레이블
            real_labels = torch.ones(batch_size, 1, device=device)
            fake_labels = torch.zeros(batch_size, 1, device=device)

            # ==================
            # Discriminator 학습
            # ==================
            optimizer_D.zero_grad()

            # 진짜 이미지
            real_output = discriminator(real_imgs)
            d_loss_real = criterion(real_output, real_labels)

            # 가짜 이미지
            z = torch.randn(batch_size, latent_dim, device=device)
            fake_imgs = generator(z)
            fake_output = discriminator(fake_imgs.detach())
            d_loss_fake = criterion(fake_output, fake_labels)

            d_loss = d_loss_real + d_loss_fake
            d_loss.backward()
            optimizer_D.step()

            # ==================
            # Generator 학습
            # ==================
            optimizer_G.zero_grad()

            # D가 가짜를 진짜로 판단하도록
            fake_output = discriminator(fake_imgs)
            g_loss = criterion(fake_output, real_labels)  # 진짜 레이블 사용

            g_loss.backward()
            optimizer_G.step()

        print(f"Epoch [{epoch+1}/{epochs}] D Loss: {d_loss.item():.4f}, G Loss: {g_loss.item():.4f}")

3. 손실 함수¶

Vanilla GAN Loss (BCE)¶

# Binary Cross Entropy
criterion = nn.BCELoss()

# D loss
d_loss = criterion(D(real), 1) + criterion(D(G(z)), 0)

# G loss (non-saturating)
g_loss = criterion(D(G(z)), 1)  # -log(D(G(z)))

# G loss (original, saturating)
# g_loss = -criterion(D(G(z)), 0)  # log(1 - D(G(z))) - 잘 안 씀

Wasserstein Loss (WGAN)¶

def wasserstein_loss(y_pred, y_true):
    """Wasserstein distance (Earth Mover's Distance)"""
    return torch.mean(y_pred * y_true)

# D (Critic) loss - 최대화
d_loss = torch.mean(D(real)) - torch.mean(D(G(z)))
# → 최소화하려면 부호 반전: -D(real) + D(G(z))

# G loss - 최소화
g_loss = -torch.mean(D(G(z)))

# Weight Clipping (WGAN)
for p in discriminator.parameters():
    p.data.clamp_(-0.01, 0.01)

WGAN-GP (Gradient Penalty)¶

def gradient_penalty(discriminator, real_imgs, fake_imgs, device):
    """Gradient Penalty for WGAN-GP (⭐⭐⭐)"""
    batch_size = real_imgs.size(0)

    # 랜덤 interpolation
    alpha = torch.rand(batch_size, 1, 1, 1, device=device)
    interpolated = alpha * real_imgs + (1 - alpha) * fake_imgs
    interpolated.requires_grad_(True)

    # Discriminator 출력
    d_interpolated = discriminator(interpolated)

    # Gradient 계산
    gradients = torch.autograd.grad(
        outputs=d_interpolated,
        inputs=interpolated,
        grad_outputs=torch.ones_like(d_interpolated),
        create_graph=True,
        retain_graph=True
    )[0]

    # Gradient norm
    gradients = gradients.view(batch_size, -1)
    gradient_norm = gradients.norm(2, dim=1)

    # Penalty: (||grad|| - 1)^2
    penalty = ((gradient_norm - 1) ** 2).mean()

    return penalty

# WGAN-GP Loss
lambda_gp = 10
gp = gradient_penalty(discriminator, real_imgs, fake_imgs, device)
d_loss = -torch.mean(D(real)) + torch.mean(D(G(z))) + lambda_gp * gp

Hinge Loss¶

# Discriminator loss
d_loss_real = torch.mean(torch.relu(1.0 - D(real)))
d_loss_fake = torch.mean(torch.relu(1.0 + D(G(z))))
d_loss = d_loss_real + d_loss_fake

# Generator loss
g_loss = -torch.mean(D(G(z)))

4. DCGAN 아키텍처¶

핵심 원칙¶

1. Pooling 제거 → Strided Conv (D), Transposed Conv (G)
2. BatchNorm 사용 (G의 출력층, D의 입력층 제외)
3. G에서 ReLU (출력층은 Tanh)
4. D에서 LeakyReLU

DCGAN Generator¶

class DCGANGenerator(nn.Module):
    """DCGAN Generator (⭐⭐⭐)

    z (100,) → (1024, 4, 4) → (512, 8, 8) → (256, 16, 16) → (128, 32, 32) → (3, 64, 64)
    """
    def __init__(self, latent_dim=100, ngf=64, nc=3):
        super().__init__()

        self.main = nn.Sequential(
            # 입력: z (latent_dim,)
            # 출력: (ngf*8, 4, 4)
            nn.ConvTranspose2d(latent_dim, ngf * 8, 4, 1, 0, bias=False),
            nn.BatchNorm2d(ngf * 8),
            nn.ReLU(True),

            # (ngf*8, 4, 4) → (ngf*4, 8, 8)
            nn.ConvTranspose2d(ngf * 8, ngf * 4, 4, 2, 1, bias=False),
            nn.BatchNorm2d(ngf * 4),
            nn.ReLU(True),

            # (ngf*4, 8, 8) → (ngf*2, 16, 16)
            nn.ConvTranspose2d(ngf * 4, ngf * 2, 4, 2, 1, bias=False),
            nn.BatchNorm2d(ngf * 2),
            nn.ReLU(True),

            # (ngf*2, 16, 16) → (ngf, 32, 32)
            nn.ConvTranspose2d(ngf * 2, ngf, 4, 2, 1, bias=False),
            nn.BatchNorm2d(ngf),
            nn.ReLU(True),

            # (ngf, 32, 32) → (nc, 64, 64)
            nn.ConvTranspose2d(ngf, nc, 4, 2, 1, bias=False),
            nn.Tanh()
        )

        self._init_weights()

    def _init_weights(self):
        for m in self.modules():
            if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
                nn.init.normal_(m.weight, 0.0, 0.02)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.normal_(m.weight, 1.0, 0.02)
                nn.init.constant_(m.bias, 0)

    def forward(self, z):
        # z: (batch, latent_dim) → (batch, latent_dim, 1, 1)
        z = z.view(z.size(0), z.size(1), 1, 1)
        return self.main(z)

DCGAN Discriminator¶

class DCGANDiscriminator(nn.Module):
    """DCGAN Discriminator (⭐⭐⭐)

    (3, 64, 64) → (64, 32, 32) → (128, 16, 16) → (256, 8, 8) → (512, 4, 4) → (1,)
    """
    def __init__(self, nc=3, ndf=64):
        super().__init__()

        self.main = nn.Sequential(
            # (nc, 64, 64) → (ndf, 32, 32)
            nn.Conv2d(nc, ndf, 4, 2, 1, bias=False),
            nn.LeakyReLU(0.2, inplace=True),

            # (ndf, 32, 32) → (ndf*2, 16, 16)
            nn.Conv2d(ndf, ndf * 2, 4, 2, 1, bias=False),
            nn.BatchNorm2d(ndf * 2),
            nn.LeakyReLU(0.2, inplace=True),

            # (ndf*2, 16, 16) → (ndf*4, 8, 8)
            nn.Conv2d(ndf * 2, ndf * 4, 4, 2, 1, bias=False),
            nn.BatchNorm2d(ndf * 4),
            nn.LeakyReLU(0.2, inplace=True),

            # (ndf*4, 8, 8) → (ndf*8, 4, 4)
            nn.Conv2d(ndf * 4, ndf * 8, 4, 2, 1, bias=False),
            nn.BatchNorm2d(ndf * 8),
            nn.LeakyReLU(0.2, inplace=True),

            # (ndf*8, 4, 4) → (1, 1, 1)
            nn.Conv2d(ndf * 8, 1, 4, 1, 0, bias=False),
            nn.Sigmoid()
        )

        self._init_weights()

    def _init_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.normal_(m.weight, 0.0, 0.02)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.normal_(m.weight, 1.0, 0.02)
                nn.init.constant_(m.bias, 0)

    def forward(self, img):
        return self.main(img).view(-1, 1)

5. 학습 안정화 기법¶

Spectral Normalization¶

from torch.nn.utils import spectral_norm

class SNDiscriminator(nn.Module):
    """Spectral Normalization 적용 Discriminator (⭐⭐⭐)"""
    def __init__(self, nc=3, ndf=64):
        super().__init__()

        self.main = nn.Sequential(
            spectral_norm(nn.Conv2d(nc, ndf, 4, 2, 1, bias=False)),
            nn.LeakyReLU(0.2, inplace=True),

            spectral_norm(nn.Conv2d(ndf, ndf * 2, 4, 2, 1, bias=False)),
            nn.LeakyReLU(0.2, inplace=True),

            spectral_norm(nn.Conv2d(ndf * 2, ndf * 4, 4, 2, 1, bias=False)),
            nn.LeakyReLU(0.2, inplace=True),

            spectral_norm(nn.Conv2d(ndf * 4, ndf * 8, 4, 2, 1, bias=False)),
            nn.LeakyReLU(0.2, inplace=True),

            spectral_norm(nn.Conv2d(ndf * 8, 1, 4, 1, 0, bias=False))
        )

    def forward(self, img):
        return self.main(img).view(-1, 1)

Label Smoothing¶

# One-sided label smoothing
real_labels = torch.ones(batch_size, 1, device=device) * 0.9  # 1.0 → 0.9
fake_labels = torch.zeros(batch_size, 1, device=device)

# 또는 noisy labels
real_labels = 0.7 + 0.3 * torch.rand(batch_size, 1, device=device)  # [0.7, 1.0]

Two Time-Scale Update Rule (TTUR)¶

# D는 더 높은 학습률
optimizer_D = torch.optim.Adam(D.parameters(), lr=0.0004, betas=(0.0, 0.9))
optimizer_G = torch.optim.Adam(G.parameters(), lr=0.0001, betas=(0.0, 0.9))

Progressive Growing¶

# 작은 해상도에서 시작해서 점진적으로 키움
# ProGAN (Progressive GAN)

resolutions = [4, 8, 16, 32, 64, 128, 256, 512, 1024]

# 각 해상도에서 일정 epoch 학습 후 다음 해상도로
# Fade-in: 새 레이어를 점진적으로 추가

6. StyleGAN 개요¶

핵심 아이디어¶

기존 GAN: z → G → 이미지
StyleGAN: z → Mapping Network → w → Synthesis Network → 이미지

Mapping Network: 8층 MLP, z를 "disentangled" w로 변환
AdaIN: w를 사용해 각 레이어의 스타일 주입

Mapping Network¶

class MappingNetwork(nn.Module):
    """StyleGAN Mapping Network (⭐⭐⭐⭐)"""
    def __init__(self, latent_dim=512, w_dim=512, num_layers=8):
        super().__init__()

        layers = []
        for i in range(num_layers):
            layers.append(nn.Linear(latent_dim if i == 0 else w_dim, w_dim))
            layers.append(nn.LeakyReLU(0.2))

        self.mapping = nn.Sequential(*layers)

    def forward(self, z):
        return self.mapping(z)

AdaIN (Adaptive Instance Normalization)¶

class AdaIN(nn.Module):
    """Adaptive Instance Normalization (⭐⭐⭐⭐)"""
    def __init__(self, num_features, w_dim):
        super().__init__()
        self.norm = nn.InstanceNorm2d(num_features)
        self.style = nn.Linear(w_dim, num_features * 2)  # scale + bias

    def forward(self, x, w):
        # x: (batch, channels, H, W)
        # w: (batch, w_dim)

        normalized = self.norm(x)

        style = self.style(w)  # (batch, channels*2)
        gamma, beta = style.chunk(2, dim=1)
        gamma = gamma.unsqueeze(-1).unsqueeze(-1)
        beta = beta.unsqueeze(-1).unsqueeze(-1)

        return gamma * normalized + beta

Style Mixing¶

# 서로 다른 z에서 w 생성
z1, z2 = torch.randn(2, latent_dim)
w1, w2 = mapping(z1), mapping(z2)

# 특정 레이어까지는 w1, 그 이후는 w2 사용
# → 스타일 혼합 (coarse: w1, fine: w2)

7. 이미지 생성 및 평가¶

샘플 이미지 생성¶

def generate_samples(generator, latent_dim, num_samples=64):
    generator.eval()
    with torch.no_grad():
        z = torch.randn(num_samples, latent_dim, device=device)
        fake_imgs = generator(z)
    return fake_imgs

# 시각화
import matplotlib.pyplot as plt
import torchvision.utils as vutils

def show_generated_images(images, nrow=8):
    """생성된 이미지 그리드 표시"""
    # [-1, 1] → [0, 1]
    images = (images + 1) / 2
    grid = vutils.make_grid(images.cpu(), nrow=nrow, normalize=False)
    plt.figure(figsize=(12, 12))
    plt.imshow(grid.permute(1, 2, 0))
    plt.axis('off')
    plt.savefig('generated_samples.png')
    plt.close()

Interpolation¶

def interpolate_latent(generator, z1, z2, steps=10):
    """잠재 공간 보간 (⭐⭐)"""
    generator.eval()
    images = []

    with torch.no_grad():
        for alpha in torch.linspace(0, 1, steps):
            z = (1 - alpha) * z1 + alpha * z2
            img = generator(z.unsqueeze(0))
            images.append(img)

    return torch.cat(images, dim=0)

# Spherical interpolation (slerp) - 더 나은 결과
def slerp(z1, z2, alpha):
    """구면 선형 보간"""
    omega = torch.acos((z1 * z2).sum() / (z1.norm() * z2.norm()))
    return (torch.sin((1 - alpha) * omega) / torch.sin(omega)) * z1 + \
           (torch.sin(alpha * omega) / torch.sin(omega)) * z2

FID (Frechet Inception Distance)¶

# FID 계산 (pytorch-fid 라이브러리 사용)
# pip install pytorch-fid

# from pytorch_fid import fid_score
# fid = fid_score.calculate_fid_given_paths(
#     [real_images_path, fake_images_path],
#     batch_size=50,
#     device=device,
#     dims=2048
# )
# 낮을수록 좋음 (0이 완벽)

8. MNIST GAN 완전 예제¶

import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torchvision import datasets, transforms
import matplotlib.pyplot as plt

# 하이퍼파라미터
latent_dim = 100
lr = 0.0002
batch_size = 64
epochs = 50

# 데이터
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize([0.5], [0.5])  # [-1, 1]
])

mnist = datasets.MNIST('data', train=True, download=True, transform=transform)
dataloader = DataLoader(mnist, batch_size=batch_size, shuffle=True)

# 간단한 모델
G = Generator(latent_dim=latent_dim, img_shape=(1, 28, 28)).to(device)
D = Discriminator(img_shape=(1, 28, 28)).to(device)

# 옵티마이저
optimizer_G = torch.optim.Adam(G.parameters(), lr=lr, betas=(0.5, 0.999))
optimizer_D = torch.optim.Adam(D.parameters(), lr=lr, betas=(0.5, 0.999))

criterion = nn.BCELoss()

# 학습
for epoch in range(epochs):
    for real_imgs, _ in dataloader:
        batch_size = real_imgs.size(0)
        real_imgs = real_imgs.to(device)

        # 레이블
        real = torch.ones(batch_size, 1, device=device)
        fake = torch.zeros(batch_size, 1, device=device)

        # D 학습
        optimizer_D.zero_grad()
        z = torch.randn(batch_size, latent_dim, device=device)
        fake_imgs = G(z)

        d_loss = criterion(D(real_imgs), real) + criterion(D(fake_imgs.detach()), fake)
        d_loss.backward()
        optimizer_D.step()

        # G 학습
        optimizer_G.zero_grad()
        g_loss = criterion(D(fake_imgs), real)
        g_loss.backward()
        optimizer_G.step()

    if (epoch + 1) % 10 == 0:
        print(f"Epoch {epoch+1}: D={d_loss.item():.4f}, G={g_loss.item():.4f}")

        # 샘플 저장
        with torch.no_grad():
            sample = G(torch.randn(16, latent_dim, device=device))
            # 저장 코드...

정리¶

핵심 개념¶

GAN: Generator와 Discriminator의 적대적 학습
손실 함수: BCE, Wasserstein, Gradient Penalty
DCGAN: Transposed Conv + BatchNorm + LeakyReLU
안정화: Spectral Norm, TTUR, Progressive Growing
StyleGAN: Mapping Network + AdaIN으로 스타일 제어

GAN 학습 팁¶

1. D와 G 균형 유지 (D가 너무 강하면 G가 학습 불가)
2. BatchNorm은 minibatch 전체에 적용 (진짜/가짜 분리하지 않음)
3. Adam beta1=0.5 사용 (momentum 줄임)
4. 학습률: 보통 0.0001 ~ 0.0002
5. 생성 이미지 주기적으로 확인

손실 함수 비교¶

손실 함수	장점	단점
BCE	간단함	Mode collapse
WGAN	안정적 학습	Weight clipping
WGAN-GP	매우 안정적	계산 비용
Hinge	간단하고 효과적	-

다음 단계¶

30_Generative_Models_VAE.md에서 VAE (Variational Autoencoder)를 학습합니다.