18. GAN
Durante sete anos, GAN era como se gerava uma imagem. Então, em cerca de dezoito meses, deixou de ser — a difusão tomou a tarefa de manchete e não a devolveu. Um capítulo que apresentasse GANs como estado da arte estaria ensinando 2019.
Então este capítulo faz algo mais útil. A arquitetura perdeu; a ideia, não. Uma perda adversarial — aprenda o que "realista" significa em vez de especificar — está rodando agora mesmo dentro do decodificador de todo Stable Diffusion, dentro do vocoder que transforma sua saída de texto-para-fala em áudio e dentro da destilação que torna possível um gerador de imagens de um passo. A GAN deixou de ser o modelo. Ela virou uma função de perda e está em toda parte.
A ideia: uma perda que você não consegue escrever
Suponha que você queira gerar rostos. Qual é a perda?
\(\ell_2\) contra um alvo não funciona — não há alvo e fazer média sobre rostos plausíveis dá o borrão do capítulo 17. Medidas de "realismo" projetadas à mão falham de imediato: toda uma que alguém já escreveu é trivialmente satisfeita por algo que não se parece nem um pouco com um rosto.
A resposta de Goodfellow1 é parar de escrever a perda e aprendê-la. Treine uma segunda rede — um discriminador — cujo único trabalho é distinguir dados reais de dados gerados. Aquilo que ele encontrar para distingui-los é a medida de realismo e ela se atualiza conforme o gerador melhora.
O gerador leva ruído a dados; o discriminador julga. Nenhum dos dois recebe uma descrição do que "realista" significa. Fonte: 4.
Dois jogadores, uma função de valor:
\[ \min_G \max_D \; V(D, G) = \mathbb{E}_{x \sim p_{\text{data}}}\big[\log D(x)\big] + \mathbb{E}_{z \sim p_z}\big[\log\big(1 - D(G(z))\big)\big] \]
O discriminador a maximiza — empurrando \(D(x) \to 1\) nos dados reais e \(D(G(z)) \to 0\) nos falsos. O gerador a minimiza. No ótimo teórico \(p_G = p_{\text{data}}\) e \(D \equiv 1/2\): o discriminador não consegue fazer melhor que cara ou coroa.
A perda do gerador que está no artigo não é a que se usa
Minimizar \(\log(1 - D(G(z)))\) tem gradiente proporcional a \(D(G(z))\), que some exatamente quando o discriminador está ganhando — ou seja, cedo no treino, quando o gerador mais precisa de sinal. A correção, no mesmo artigo: maximize \(\log D(G(z))\). Mesmo ponto fixo, gradiente forte quando você está perdendo. Essa é a perda não saturante e é o que toda implementação de fato contém.
Treinando uma e olhando o que não dá para ver
Abaixo há uma GAN de verdade — dois MLPs, retropropagação escrita à mão, Adam — treinada no seu navegador sobre oito gaussianas dispostas em anel.
O painel da esquerda já vale por si: a superfície de decisão do discriminador não está convergindo para nada, porque não existe alvo fixo. Ela se remodela em volta de onde quer que o gerador esteja colocando massa e o gerador a persegue. Isto é um jogo de dois jogadores, não uma otimização e jogos não têm um valor que desce.
Agora o terceiro painel, que é o ponto do exercício inteiro. A linha laranja é \(D(G(z))\) — a única grandeza que o laço de treino calcula. Ela fica perto de 0,5 e ali permanece. A linha verde é a cobertura de modos, medida à parte, por código que o laço de treino nunca roda.
Ponha a dimensão latente em 1 e retreine. A cobertura desaba para dois ou três modos de oito, a maior parte das amostras cai longe dos dados e a linha laranja não se mexe.
Este é o problema real das GANs e não é instabilidade
Uma GAN não tem perda que se possa ler. Uma perda que cai não significa nada. Não há critério principiado de early stopping, não há como comparar duas execuções pelas curvas, não há como saber se um checkpoint é melhor que o anterior sem gerar amostras e avaliá-las com uma métrica separada que você mesmo teve de escolher.
Retreine o painel algumas vezes com os mesmos ajustes e veja a linha verde parar em lugares diferentes a cada vez. Essa variância — invisível ao laço de treino — é o que tornou GANs caras de colocar em produção e é a maior razão isolada pela qual a difusão as substituiu. A difusão tem uma perda de regressão comum, que desce.
Os modos de falha, nomeados
| Falha | Como aparece | Por que acontece |
|---|---|---|
| Colapso de modos | O gerador produz poucas saídas, independentemente de \(z\) | Nada no objetivo premia diversidade. Se uma saída engana o discriminador atual, produzir só aquela saída é uma jogada perfeitamente boa. Comportamento tipo KL reversa (capítulo 16). |
| Gradiente do gerador some | O treino trava, \(D(G(z)) \to 0\) | O discriminador venceu completo demais. A perda não saturante ajuda; equilibrar as duas taxas de aprendizado ajuda mais. |
| Oscilação | A qualidade sobe, cai, sobe | Os dois jogadores entram em ciclo. Não há razão para um jogo convergir e muitas vezes ele não converge. |
| Overfitting do discriminador | Ótimas amostras no treino, generalização ruim | O discriminador decora o conjunto de treino. Foi isso que limitou GANs em conjuntos pequenos até a augmentation adaptativa. |
As mitigações que ficaram: WGAN-GP7 (troque a divergência JS por uma distância de Wasserstein e restrinja o gradiente do crítico), normalização espectral8 (limite a constante de Lipschitz do discriminador, uma linha de código), regularização R19 (penalize o gradiente do discriminador nos dados reais) e augmentation adaptativa do discriminador10 (aumente o que o discriminador vê para que ele não decore). Cada uma é, no fundo, o mesmo movimento: impedir o discriminador de ficar afiado demais, porque um discriminador perfeitamente confiante entrega ao gerador um gradiente que não aponta para lugar nenhum.
A linhagem, em resumo
| Modelo | A contribuição |
|---|---|
| DCGAN (2015) | Gerador e discriminador convolucionais e um conjunto de regras empíricas que tornaram o treino de GANs reprodutível. |
| GAN condicional | Alimente um rótulo nas duas redes. Ancestral de todo modelo generativo condicionado, este curso incluído. |
| pix2pix / CycleGAN (2017) | Tradução imagem-para-imagem e — com consistência de ciclo — sem dados pareados. |
| WGAN / WGAN-GP (2017) | Uma divergência diferente e uma perda que ao menos correlaciona com qualidade. |
| StyleGAN 1–3 (2019–2021)5 | Ainda os rostos mais bonitos produzidos por uma GAN e a origem do espaço latente \(\mathcal{W}\) desemaranhado que fez a edição semântica funcionar. |
| GigaGAN (2023)6 | Texto-para-imagem com qualidade competitiva com difusão e centenas de vezes mais rápido na inferência. Chegou depois de a área ter se mudado. |
O StyleGAN ainda vale a leitura
Não pela arquitetura, mas pelo espaço \(\mathcal{W}\): um latente em que direções correspondem a atributos interpretáveis — idade, pose, expressão — de modo que editar é aritmética de vetores. Nada em difusão é tão limpo assim e edição latente é área de pesquisa ativa precisamente porque essa propriedade se perdeu.
Onde a perda adversarial de fato vive hoje
Esta é a parte para lembrar, porque você vai usá-la sem nunca treinar uma GAN de ponta a ponta.
| Onde | O que o discriminador faz |
|---|---|
| O decodificador do VAE em todo modelo de difusão latente | Só perda de reconstrução dá borrão. Um discriminador sobre a saída do decodificador pune exatamente essa suavidade — é a receita do VQGAN11 e é por isso que o decodificador do SD produz textura nítida. É o uso mais amplamente implantado de perda adversarial hoje. |
| Super-resolução | Ampliar é um problema um-para-muitos, então o \(\ell_2\) devolve a média e você fica com papa. Real-ESRGAN e descendentes são treinados adversarialmente. |
| Vocoders neurais | HiFi-GAN e sucessores transformam espectrogramas em formas de onda em quase todo sistema de texto-para-fala em produção. |
| Destilação de um e dois passos | Comprimir um modelo de difusão de 50 passos em 1–4 exige uma perda que diga "isto parece real", não "isto casa com a média do professor". A destilação adversarial de difusão12 é como o SDXL-Turbo e parentes chegam a tempo real. |
| Adaptação de domínio | Um classificador adversarial que não consegue distinguir origem de destino força features invariantes ao domínio. |
O padrão, enunciado uma vez
Use uma perda adversarial quando o alvo é uma entre muitas respostas válidas e fazer média delas é errado. Isso descreve ampliar, decodificar, vocodificar e destilar em um passo. Não descreve classificação, onde média está tudo bem — e é por isso que a ideia adversarial sobreviveu exatamente onde o \(\ell_2\) falha.
Pontos principais
- Uma GAN aprende a própria perda. O trabalho do discriminador é descobrir o que "realista" significa para que você não precise especificar.
- A perda do gerador do artigo original satura; todo mundo usa a forma não saturante \(-\log D(G(z))\).
- O problema real não é instabilidade, é impossibilidade de medir: \(D(G(z))\) fica perto de 0,5 tanto com um modelo excelente quanto com um colapsado. Não há perda em que fazer early stopping.
- Colapso de modos é o que o objetivo permite, não um acidente — nada nele premia diversidade.
- Toda estabilização que ficou faz a mesma coisa: impedir o discriminador de ficar afiado demais (WGAN-GP, norma espectral, R1, ADA).
- A difusão venceu porque tem uma perda de regressão comum, que desce, é estável e cobre modos. As GANs ainda vencem em velocidade de inferência — uma passagem direta.
- A perda adversarial está em toda parte: em decodificadores de autoencoders latentes, super-resolução, vocoders e destilação de poucos passos. Use-a quando a resposta é uma entre muitas e fazer média é errado.
Implementação
CIFAR-10
CIFAR-10
Abaixo, um exemplo básico de como criar e treinar uma GAN no conjunto CIFAR-10. Leia como história e como mecânica — o laço de treino, os dois otimizadores, as convenções de rótulos — e não como algo sobre o que construir um produto.
Example extracted from https://www.geeksforgeeks.org/deep-learning/generative-adversarial-network-gan/
1. Importing Libraries
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
from torchvision import datasets, transforms
import matplotlib.pyplot as plt
import numpy as np
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
2. Defining Image Transformations
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
3. Loaging the CIFAR-10 Dataset
train_dataset = datasets.CIFAR10(root='./data',\
train=True, download=True, transform=transform)
dataloader = torch.utils.data.DataLoader(train_dataset, \
batch_size=32, shuffle=True)
100%|██████████| 170M/170M [00:16<00:00, 10.5MB/s]
4. Defining GAN Hyperparameters
Set important training parameters:
- latent_dim: Dimensionality of the noise vector.
- lr: Learning rate of the optimizer.
- beta1, beta2: Beta parameters for Adam optimizer (e.g 0.5, 0.999)
- num_epochs: Number of times the entire dataset will be processed (e.g 10)
5. Building the Generator
Create a neural network that converts random noise into images. Use transpose convolutional layers, batch normalization and ReLU activations. The final layer uses Tanh activation to scale outputs to the range [-1, 1].
- nn.Linear(latent_dim, 128 * 8 * 8): Defines a fully connected layer that projects the noise vector into a higher dimensional feature space.
- nn.Upsample(scale_factor=2): Doubles the spatial resolution of the feature maps by upsampling.
- nn.Conv2d(128, 128, kernel_size=3, padding=1): Applies a convolutional layer keeping the number of channels the same to refine features.
class Generator(nn.Module):
def __init__(self, latent_dim):
super(Generator, self).__init__()
self.model = nn.Sequential(
nn.Linear(latent_dim, 128 * 8 * 8),
nn.ReLU(),
nn.Unflatten(1, (128, 8, 8)),
nn.Upsample(scale_factor=2),
nn.Conv2d(128, 128, kernel_size=3, padding=1),
nn.BatchNorm2d(128, momentum=0.78),
nn.ReLU(),
nn.Upsample(scale_factor=2),
nn.Conv2d(128, 64, kernel_size=3, padding=1),
nn.BatchNorm2d(64, momentum=0.78),
nn.ReLU(),
nn.Conv2d(64, 3, kernel_size=3, padding=1),
nn.Tanh()
)
def forward(self, z):
img = self.model(z)
return img
6. Building the Discriminator
Create a binary classifier network that distinguishes real from fake images. Use convolutional layers, batch normalization, dropout, LeakyReLU activation and a Sigmoid output layer to give a probability between 0 and 1.
- nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1): Second convolutional layer increasing channels to 64, downsampling further.
- nn.BatchNorm2d(256, momentum=0.8): Batch normalization for 256 feature maps with momentum 0.8.
class Discriminator(nn.Module):
def __init__(self):
super(Discriminator, self).__init__()
self.model = nn.Sequential(
nn.Conv2d(3, 32, kernel_size=3, stride=2, padding=1),
nn.LeakyReLU(0.2),
nn.Dropout(0.25),
nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1),
nn.ZeroPad2d((0, 1, 0, 1)),
nn.BatchNorm2d(64, momentum=0.82),
nn.LeakyReLU(0.25),
nn.Dropout(0.25),
nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1),
nn.BatchNorm2d(128, momentum=0.82),
nn.LeakyReLU(0.2),
nn.Dropout(0.25),
nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(256, momentum=0.8),
nn.LeakyReLU(0.25),
nn.Dropout(0.25),
nn.Flatten(),
nn.Linear(256 * 5 * 5, 1),
nn.Sigmoid()
)
def forward(self, img):
validity = self.model(img)
return validity
7. Initializing
- Generator and Discriminator are initialized on the available device (GPU or CPU).
- Binary Cross-Entropy (BCE) Loss is chosen as the loss function.
- Adam optimizers are defined separately for the generator and discriminator with specified learning rates and betas.
generator = Generator(latent_dim).to(device)
discriminator = Discriminator().to(device)
adversarial_loss = nn.BCELoss()
optimizer_G = optim.Adam(generator.parameters()\
, lr=lr, betas=(beta1, beta2))
optimizer_D = optim.Adam(discriminator.parameters()\
, lr=lr, betas=(beta1, beta2))
8. Training
Train the discriminator on real and fake images, then update the generator to improve its fake image quality. Track losses and visualize generated images after each epoch.
- valid = torch.ones(real_images.size(0), 1, device=device): Create a tensor of ones representing real labels for the discriminator.
- fake = torch.zeros(real_images.size(0), 1, device=device): Create a tensor of zeros representing fake labels for the discriminator.
- z = torch.randn(real_images.size(0), latent_dim, device=device): Generate random noise vectors as input for the generator.
- g_loss = adversarial_loss(discriminator(gen_images), valid): Calculate generator loss based on the discriminator classifying fake images as real.
- grid = torchvision.utils.make_grid(generated, nrow=4, normalize=True): Arrange generated images into a grid for display, normalizing pixel values.
for epoch in range(num_epochs):
for i, batch in enumerate(dataloader):
real_images = batch[0].to(device)
# Create a tensor of ones representing real labels for the discriminator.
valid = torch.ones(real_images.size(0), 1, device=device)
# Create a tensor of zeros representing fake labels for the discriminator.
fake = torch.zeros(real_images.size(0), 1, device=device)
real_images = real_images.to(device)
optimizer_D.zero_grad()
# Generate random noise vectors as input for the generator.
z = torch.randn(real_images.size(0), latent_dim, device=device)
fake_images = generator(z)
real_loss = adversarial_loss(discriminator\
(real_images), valid)
fake_loss = adversarial_loss(discriminator\
(fake_images.detach()), fake)
d_loss = (real_loss + fake_loss) / 2
d_loss.backward()
optimizer_D.step()
optimizer_G.zero_grad()
gen_images = generator(z)
# Calculate generator loss based on the discriminator classifying fake images as real.
g_loss = adversarial_loss(discriminator(gen_images), valid)
g_loss.backward()
optimizer_G.step()
if (i + 1) % 100 == 0:
print(
f"Epoch [{epoch+1}/{num_epochs}]\
Batch {i+1}/{len(dataloader)} "
f"Discriminator Loss: {d_loss.item():.4f} "
f"Generator Loss: {g_loss.item():.4f}"
)
if (epoch + 1) % 10 == 0:
with torch.no_grad():
z = torch.randn(16, latent_dim, device=device)
generated = generator(z).detach().cpu()
# Arrange generated images into a grid for display, normalizing pixel values.
grid = torchvision.utils.make_grid(generated,\
nrow=4, normalize=True)
plt.imshow(np.transpose(grid, (1, 2, 0)))
plt.axis("off")
plt.show()
plt.close()
Epoch [1/30] Batch 100/1563 Discriminator Loss: 0.6310 Generator Loss: 1.0856
Epoch [1/30] Batch 200/1563 Discriminator Loss: 0.7243 Generator Loss: 1.0722
Epoch [1/30] Batch 300/1563 Discriminator Loss: 0.7337 Generator Loss: 1.1005
Epoch [1/30] Batch 400/1563 Discriminator Loss: 0.6917 Generator Loss: 1.1717
Epoch [1/30] Batch 500/1563 Discriminator Loss: 0.5310 Generator Loss: 1.1474
Epoch [1/30] Batch 600/1563 Discriminator Loss: 0.7355 Generator Loss: 0.8708
Epoch [1/30] Batch 700/1563 Discriminator Loss: 0.6745 Generator Loss: 1.2315
Epoch [1/30] Batch 800/1563 Discriminator Loss: 0.7858 Generator Loss: 0.8245
Epoch [1/30] Batch 900/1563 Discriminator Loss: 0.6330 Generator Loss: 0.9612
Epoch [1/30] Batch 1000/1563 Discriminator Loss: 0.6853 Generator Loss: 1.1186
Epoch [1/30] Batch 1100/1563 Discriminator Loss: 0.5872 Generator Loss: 1.3537
Epoch [1/30] Batch 1200/1563 Discriminator Loss: 0.6135 Generator Loss: 1.0075
Epoch [1/30] Batch 1300/1563 Discriminator Loss: 0.5315 Generator Loss: 1.4053
Epoch [1/30] Batch 1400/1563 Discriminator Loss: 0.7756 Generator Loss: 0.8813
Epoch [1/30] Batch 1500/1563 Discriminator Loss: 0.6498 Generator Loss: 0.8439
Epoch [2/30] Batch 100/1563 Discriminator Loss: 0.6724 Generator Loss: 0.7996
Epoch [2/30] Batch 200/1563 Discriminator Loss: 0.8157 Generator Loss: 0.7308
Epoch [2/30] Batch 300/1563 Discriminator Loss: 0.5542 Generator Loss: 1.0256
Epoch [2/30] Batch 400/1563 Discriminator Loss: 0.5977 Generator Loss: 1.1029
Epoch [2/30] Batch 500/1563 Discriminator Loss: 0.6931 Generator Loss: 0.8717
Epoch [2/30] Batch 600/1563 Discriminator Loss: 0.5693 Generator Loss: 1.3365
Epoch [2/30] Batch 700/1563 Discriminator Loss: 0.6488 Generator Loss: 0.9503
Epoch [2/30] Batch 800/1563 Discriminator Loss: 0.6761 Generator Loss: 0.8909
Epoch [2/30] Batch 900/1563 Discriminator Loss: 0.5708 Generator Loss: 0.9972
Epoch [2/30] Batch 1000/1563 Discriminator Loss: 0.6001 Generator Loss: 1.1221
Epoch [2/30] Batch 1100/1563 Discriminator Loss: 0.6978 Generator Loss: 0.8663
Epoch [2/30] Batch 1200/1563 Discriminator Loss: 0.7402 Generator Loss: 1.1249
Epoch [2/30] Batch 1300/1563 Discriminator Loss: 0.6372 Generator Loss: 1.0888
Epoch [2/30] Batch 1400/1563 Discriminator Loss: 0.6317 Generator Loss: 0.9584
Epoch [2/30] Batch 1500/1563 Discriminator Loss: 0.5677 Generator Loss: 1.0163
Epoch [3/30] Batch 100/1563 Discriminator Loss: 0.6370 Generator Loss: 1.3195
Epoch [3/30] Batch 200/1563 Discriminator Loss: 0.5646 Generator Loss: 1.3755
Epoch [3/30] Batch 300/1563 Discriminator Loss: 0.7698 Generator Loss: 0.8023
Epoch [3/30] Batch 400/1563 Discriminator Loss: 0.5839 Generator Loss: 1.0319
Epoch [3/30] Batch 500/1563 Discriminator Loss: 0.5613 Generator Loss: 1.2391
Epoch [3/30] Batch 600/1563 Discriminator Loss: 0.5798 Generator Loss: 1.1867
Epoch [3/30] Batch 700/1563 Discriminator Loss: 0.6622 Generator Loss: 1.4665
Epoch [3/30] Batch 800/1563 Discriminator Loss: 0.7389 Generator Loss: 1.3259
Epoch [3/30] Batch 900/1563 Discriminator Loss: 0.6855 Generator Loss: 1.4243
Epoch [3/30] Batch 1000/1563 Discriminator Loss: 0.5348 Generator Loss: 1.5141
Epoch [3/30] Batch 1100/1563 Discriminator Loss: 0.4696 Generator Loss: 1.3948
Epoch [3/30] Batch 1200/1563 Discriminator Loss: 0.5467 Generator Loss: 0.6435
Epoch [3/30] Batch 1300/1563 Discriminator Loss: 0.6741 Generator Loss: 0.9211
Epoch [3/30] Batch 1400/1563 Discriminator Loss: 0.4664 Generator Loss: 1.0165
Epoch [3/30] Batch 1500/1563 Discriminator Loss: 0.4799 Generator Loss: 0.9388
Epoch [4/30] Batch 100/1563 Discriminator Loss: 0.6887 Generator Loss: 0.9108
Epoch [4/30] Batch 200/1563 Discriminator Loss: 0.7852 Generator Loss: 0.8538
Epoch [4/30] Batch 300/1563 Discriminator Loss: 0.5036 Generator Loss: 0.7702
Epoch [4/30] Batch 400/1563 Discriminator Loss: 0.4476 Generator Loss: 1.3409
Epoch [4/30] Batch 500/1563 Discriminator Loss: 0.6550 Generator Loss: 1.2383
Epoch [4/30] Batch 600/1563 Discriminator Loss: 0.7891 Generator Loss: 1.6518
Epoch [4/30] Batch 700/1563 Discriminator Loss: 0.5458 Generator Loss: 1.3395
Epoch [4/30] Batch 800/1563 Discriminator Loss: 0.6637 Generator Loss: 0.9561
Epoch [4/30] Batch 900/1563 Discriminator Loss: 0.6408 Generator Loss: 1.4673
Epoch [4/30] Batch 1000/1563 Discriminator Loss: 0.5879 Generator Loss: 0.8866
Epoch [4/30] Batch 1100/1563 Discriminator Loss: 0.5762 Generator Loss: 1.2028
Epoch [4/30] Batch 1200/1563 Discriminator Loss: 0.4307 Generator Loss: 1.8184
Epoch [4/30] Batch 1300/1563 Discriminator Loss: 0.8079 Generator Loss: 0.7371
Epoch [4/30] Batch 1400/1563 Discriminator Loss: 0.5739 Generator Loss: 0.9877
Epoch [4/30] Batch 1500/1563 Discriminator Loss: 0.5317 Generator Loss: 1.7946
Epoch [5/30] Batch 100/1563 Discriminator Loss: 0.4036 Generator Loss: 1.1946
Epoch [5/30] Batch 200/1563 Discriminator Loss: 0.4312 Generator Loss: 1.4645
Epoch [5/30] Batch 300/1563 Discriminator Loss: 0.4584 Generator Loss: 1.0263
Epoch [5/30] Batch 400/1563 Discriminator Loss: 0.5096 Generator Loss: 1.4634
Epoch [5/30] Batch 500/1563 Discriminator Loss: 0.6048 Generator Loss: 1.7682
Epoch [5/30] Batch 600/1563 Discriminator Loss: 0.7375 Generator Loss: 1.0852
Epoch [5/30] Batch 700/1563 Discriminator Loss: 0.6004 Generator Loss: 1.3448
Epoch [5/30] Batch 800/1563 Discriminator Loss: 0.5066 Generator Loss: 1.0084
Epoch [5/30] Batch 900/1563 Discriminator Loss: 0.6089 Generator Loss: 1.0463
Epoch [5/30] Batch 1000/1563 Discriminator Loss: 0.4824 Generator Loss: 1.2175
Epoch [5/30] Batch 1100/1563 Discriminator Loss: 0.7544 Generator Loss: 0.8785
Epoch [5/30] Batch 1200/1563 Discriminator Loss: 0.4876 Generator Loss: 1.2693
Epoch [5/30] Batch 1300/1563 Discriminator Loss: 0.4419 Generator Loss: 0.9601
Epoch [5/30] Batch 1400/1563 Discriminator Loss: 0.5805 Generator Loss: 1.0689
Epoch [5/30] Batch 1500/1563 Discriminator Loss: 0.5700 Generator Loss: 1.5043
Epoch [6/30] Batch 100/1563 Discriminator Loss: 0.3690 Generator Loss: 1.8759
Epoch [6/30] Batch 200/1563 Discriminator Loss: 0.4619 Generator Loss: 1.5484
Epoch [6/30] Batch 300/1563 Discriminator Loss: 0.5607 Generator Loss: 1.1555
Epoch [6/30] Batch 400/1563 Discriminator Loss: 0.3786 Generator Loss: 1.5910
Epoch [6/30] Batch 500/1563 Discriminator Loss: 0.6497 Generator Loss: 0.9485
Epoch [6/30] Batch 600/1563 Discriminator Loss: 0.4534 Generator Loss: 1.4235
Epoch [6/30] Batch 700/1563 Discriminator Loss: 0.5807 Generator Loss: 0.6237
Epoch [6/30] Batch 800/1563 Discriminator Loss: 0.5381 Generator Loss: 1.4090
Epoch [6/30] Batch 900/1563 Discriminator Loss: 0.4686 Generator Loss: 1.3971
Epoch [6/30] Batch 1000/1563 Discriminator Loss: 0.4787 Generator Loss: 1.4606
Epoch [6/30] Batch 1100/1563 Discriminator Loss: 0.6769 Generator Loss: 1.3275
Epoch [6/30] Batch 1200/1563 Discriminator Loss: 0.4402 Generator Loss: 1.3996
Epoch [6/30] Batch 1300/1563 Discriminator Loss: 0.6449 Generator Loss: 1.2970
Epoch [6/30] Batch 1400/1563 Discriminator Loss: 0.7675 Generator Loss: 1.1988
Epoch [6/30] Batch 1500/1563 Discriminator Loss: 0.6860 Generator Loss: 1.0752
Epoch [7/30] Batch 100/1563 Discriminator Loss: 0.5656 Generator Loss: 0.5764
Epoch [7/30] Batch 200/1563 Discriminator Loss: 0.4979 Generator Loss: 1.2684
Epoch [7/30] Batch 300/1563 Discriminator Loss: 0.6714 Generator Loss: 0.9369
Epoch [7/30] Batch 400/1563 Discriminator Loss: 0.6402 Generator Loss: 1.0803
Epoch [7/30] Batch 500/1563 Discriminator Loss: 0.5717 Generator Loss: 1.3460
Epoch [7/30] Batch 600/1563 Discriminator Loss: 0.4720 Generator Loss: 1.2296
Epoch [7/30] Batch 700/1563 Discriminator Loss: 0.5592 Generator Loss: 1.6116
Epoch [7/30] Batch 800/1563 Discriminator Loss: 0.7146 Generator Loss: 1.3487
Epoch [7/30] Batch 900/1563 Discriminator Loss: 0.4830 Generator Loss: 1.3931
Epoch [7/30] Batch 1000/1563 Discriminator Loss: 0.4894 Generator Loss: 1.7213
Epoch [7/30] Batch 1100/1563 Discriminator Loss: 0.8574 Generator Loss: 1.3492
Epoch [7/30] Batch 1200/1563 Discriminator Loss: 0.8195 Generator Loss: 0.9998
Epoch [7/30] Batch 1300/1563 Discriminator Loss: 0.7399 Generator Loss: 1.3046
Epoch [7/30] Batch 1400/1563 Discriminator Loss: 0.5688 Generator Loss: 0.9374
Epoch [7/30] Batch 1500/1563 Discriminator Loss: 0.4341 Generator Loss: 1.2548
Epoch [8/30] Batch 100/1563 Discriminator Loss: 0.6533 Generator Loss: 1.1707
Epoch [8/30] Batch 200/1563 Discriminator Loss: 0.4287 Generator Loss: 1.1481
Epoch [8/30] Batch 300/1563 Discriminator Loss: 0.5281 Generator Loss: 1.7373
Epoch [8/30] Batch 400/1563 Discriminator Loss: 0.5967 Generator Loss: 0.9080
Epoch [8/30] Batch 500/1563 Discriminator Loss: 0.6105 Generator Loss: 1.0256
Epoch [8/30] Batch 600/1563 Discriminator Loss: 0.5247 Generator Loss: 1.5951
Epoch [8/30] Batch 700/1563 Discriminator Loss: 0.7669 Generator Loss: 1.3100
Epoch [8/30] Batch 800/1563 Discriminator Loss: 0.5892 Generator Loss: 1.1584
Epoch [8/30] Batch 900/1563 Discriminator Loss: 0.8667 Generator Loss: 0.8409
Epoch [8/30] Batch 1000/1563 Discriminator Loss: 0.8007 Generator Loss: 0.6958
Epoch [8/30] Batch 1100/1563 Discriminator Loss: 0.6248 Generator Loss: 1.0027
Epoch [8/30] Batch 1200/1563 Discriminator Loss: 0.9791 Generator Loss: 1.2779
Epoch [8/30] Batch 1300/1563 Discriminator Loss: 0.6273 Generator Loss: 1.3308
Epoch [8/30] Batch 1400/1563 Discriminator Loss: 0.4397 Generator Loss: 2.0570
Epoch [8/30] Batch 1500/1563 Discriminator Loss: 0.5385 Generator Loss: 1.1689
Epoch [9/30] Batch 100/1563 Discriminator Loss: 0.8302 Generator Loss: 1.3391
Epoch [9/30] Batch 200/1563 Discriminator Loss: 0.4323 Generator Loss: 1.5036
Epoch [9/30] Batch 300/1563 Discriminator Loss: 0.6001 Generator Loss: 2.3062
Epoch [9/30] Batch 400/1563 Discriminator Loss: 0.5612 Generator Loss: 1.0222
Epoch [9/30] Batch 500/1563 Discriminator Loss: 0.5333 Generator Loss: 1.5129
Epoch [9/30] Batch 600/1563 Discriminator Loss: 0.4739 Generator Loss: 0.9228
Epoch [9/30] Batch 700/1563 Discriminator Loss: 0.6237 Generator Loss: 1.3650
Epoch [9/30] Batch 800/1563 Discriminator Loss: 0.6755 Generator Loss: 0.7273
Epoch [9/30] Batch 900/1563 Discriminator Loss: 0.5411 Generator Loss: 1.7899
Epoch [9/30] Batch 1000/1563 Discriminator Loss: 0.5930 Generator Loss: 1.0269
Epoch [9/30] Batch 1100/1563 Discriminator Loss: 0.5476 Generator Loss: 1.8822
Epoch [9/30] Batch 1200/1563 Discriminator Loss: 0.7660 Generator Loss: 1.4145
Epoch [9/30] Batch 1300/1563 Discriminator Loss: 0.5074 Generator Loss: 1.5135
Epoch [9/30] Batch 1400/1563 Discriminator Loss: 0.5582 Generator Loss: 1.1927
Epoch [9/30] Batch 1500/1563 Discriminator Loss: 0.8168 Generator Loss: 1.2279
Epoch [10/30] Batch 100/1563 Discriminator Loss: 0.2423 Generator Loss: 1.7337
Epoch [10/30] Batch 200/1563 Discriminator Loss: 0.9663 Generator Loss: 1.0028
Epoch [10/30] Batch 300/1563 Discriminator Loss: 0.5886 Generator Loss: 0.9407
Epoch [10/30] Batch 400/1563 Discriminator Loss: 0.7274 Generator Loss: 1.0736
Epoch [10/30] Batch 500/1563 Discriminator Loss: 0.6266 Generator Loss: 0.8596
Epoch [10/30] Batch 600/1563 Discriminator Loss: 0.5225 Generator Loss: 2.0611
Epoch [10/30] Batch 700/1563 Discriminator Loss: 0.4990 Generator Loss: 1.6639
Epoch [10/30] Batch 800/1563 Discriminator Loss: 0.8305 Generator Loss: 0.8100
Epoch [10/30] Batch 900/1563 Discriminator Loss: 0.5522 Generator Loss: 1.2785
Epoch [10/30] Batch 1000/1563 Discriminator Loss: 0.8009 Generator Loss: 1.3512
Epoch [10/30] Batch 1100/1563 Discriminator Loss: 0.6749 Generator Loss: 1.1727
Epoch [10/30] Batch 1200/1563 Discriminator Loss: 0.8850 Generator Loss: 1.1001
Epoch [10/30] Batch 1300/1563 Discriminator Loss: 0.6387 Generator Loss: 1.4558
Epoch [10/30] Batch 1400/1563 Discriminator Loss: 0.6613 Generator Loss: 2.5028
Epoch [10/30] Batch 1500/1563 Discriminator Loss: 0.5191 Generator Loss: 0.9412
Epoch [11/30] Batch 100/1563 Discriminator Loss: 0.6146 Generator Loss: 1.1872
Epoch [11/30] Batch 200/1563 Discriminator Loss: 0.7205 Generator Loss: 1.0533
Epoch [11/30] Batch 300/1563 Discriminator Loss: 0.6223 Generator Loss: 1.1095
Epoch [11/30] Batch 400/1563 Discriminator Loss: 0.5828 Generator Loss: 0.7897
Epoch [11/30] Batch 500/1563 Discriminator Loss: 0.5107 Generator Loss: 1.1712
Epoch [11/30] Batch 600/1563 Discriminator Loss: 0.5452 Generator Loss: 1.1918
Epoch [11/30] Batch 700/1563 Discriminator Loss: 0.4733 Generator Loss: 1.6043
Epoch [11/30] Batch 800/1563 Discriminator Loss: 0.6569 Generator Loss: 1.1823
Epoch [11/30] Batch 900/1563 Discriminator Loss: 0.7731 Generator Loss: 0.9131
Epoch [11/30] Batch 1000/1563 Discriminator Loss: 0.7283 Generator Loss: 0.9966
Epoch [11/30] Batch 1100/1563 Discriminator Loss: 0.6360 Generator Loss: 1.4022
Epoch [11/30] Batch 1200/1563 Discriminator Loss: 0.6779 Generator Loss: 1.6380
Epoch [11/30] Batch 1300/1563 Discriminator Loss: 0.5721 Generator Loss: 1.0732
Epoch [11/30] Batch 1400/1563 Discriminator Loss: 0.5500 Generator Loss: 1.6293
Epoch [11/30] Batch 1500/1563 Discriminator Loss: 0.6260 Generator Loss: 1.6629
Epoch [12/30] Batch 100/1563 Discriminator Loss: 0.2396 Generator Loss: 1.8887
Epoch [12/30] Batch 200/1563 Discriminator Loss: 0.8208 Generator Loss: 0.7347
Epoch [12/30] Batch 300/1563 Discriminator Loss: 0.7865 Generator Loss: 1.2288
Epoch [12/30] Batch 400/1563 Discriminator Loss: 0.8839 Generator Loss: 1.5515
Epoch [12/30] Batch 500/1563 Discriminator Loss: 0.4100 Generator Loss: 1.6313
Epoch [12/30] Batch 600/1563 Discriminator Loss: 0.5070 Generator Loss: 0.7579
Epoch [12/30] Batch 700/1563 Discriminator Loss: 0.4470 Generator Loss: 1.4843
Epoch [12/30] Batch 800/1563 Discriminator Loss: 0.3804 Generator Loss: 1.9155
Epoch [12/30] Batch 900/1563 Discriminator Loss: 0.8740 Generator Loss: 0.9627
Epoch [12/30] Batch 1000/1563 Discriminator Loss: 0.6253 Generator Loss: 1.2338
Epoch [12/30] Batch 1100/1563 Discriminator Loss: 0.5772 Generator Loss: 0.7276
Epoch [12/30] Batch 1200/1563 Discriminator Loss: 0.6984 Generator Loss: 1.2770
Epoch [12/30] Batch 1300/1563 Discriminator Loss: 0.5281 Generator Loss: 1.6734
Epoch [12/30] Batch 1400/1563 Discriminator Loss: 0.7011 Generator Loss: 1.5269
Epoch [12/30] Batch 1500/1563 Discriminator Loss: 0.6542 Generator Loss: 1.1141
Epoch [13/30] Batch 100/1563 Discriminator Loss: 0.3590 Generator Loss: 1.6407
Epoch [13/30] Batch 200/1563 Discriminator Loss: 0.6046 Generator Loss: 0.6858
Epoch [13/30] Batch 300/1563 Discriminator Loss: 0.5175 Generator Loss: 2.0719
Epoch [13/30] Batch 400/1563 Discriminator Loss: 0.6386 Generator Loss: 1.1742
Epoch [13/30] Batch 500/1563 Discriminator Loss: 0.5023 Generator Loss: 1.7407
Epoch [13/30] Batch 600/1563 Discriminator Loss: 0.4565 Generator Loss: 1.8108
Epoch [13/30] Batch 700/1563 Discriminator Loss: 0.3615 Generator Loss: 0.7747
Epoch [13/30] Batch 800/1563 Discriminator Loss: 0.5649 Generator Loss: 1.1583
Epoch [13/30] Batch 900/1563 Discriminator Loss: 0.5132 Generator Loss: 1.2286
Epoch [13/30] Batch 1000/1563 Discriminator Loss: 0.6191 Generator Loss: 1.4292
Epoch [13/30] Batch 1100/1563 Discriminator Loss: 0.4916 Generator Loss: 1.4518
Epoch [13/30] Batch 1200/1563 Discriminator Loss: 0.5353 Generator Loss: 1.0762
Epoch [13/30] Batch 1300/1563 Discriminator Loss: 0.6027 Generator Loss: 0.7657
Epoch [13/30] Batch 1400/1563 Discriminator Loss: 0.6896 Generator Loss: 1.5476
Epoch [13/30] Batch 1500/1563 Discriminator Loss: 0.5865 Generator Loss: 0.7674
Epoch [14/30] Batch 100/1563 Discriminator Loss: 0.3891 Generator Loss: 1.4870
Epoch [14/30] Batch 200/1563 Discriminator Loss: 0.4392 Generator Loss: 0.7682
Epoch [14/30] Batch 300/1563 Discriminator Loss: 0.9042 Generator Loss: 0.7021
Epoch [14/30] Batch 400/1563 Discriminator Loss: 0.5336 Generator Loss: 1.6118
Epoch [14/30] Batch 500/1563 Discriminator Loss: 0.6186 Generator Loss: 1.0359
Epoch [14/30] Batch 600/1563 Discriminator Loss: 0.7181 Generator Loss: 0.8545
Epoch [14/30] Batch 700/1563 Discriminator Loss: 0.8549 Generator Loss: 0.8246
Epoch [14/30] Batch 800/1563 Discriminator Loss: 0.8046 Generator Loss: 1.3709
Epoch [14/30] Batch 900/1563 Discriminator Loss: 0.4450 Generator Loss: 1.2583
Epoch [14/30] Batch 1000/1563 Discriminator Loss: 0.4912 Generator Loss: 1.8047
Epoch [14/30] Batch 1100/1563 Discriminator Loss: 0.4890 Generator Loss: 1.1271
Epoch [14/30] Batch 1200/1563 Discriminator Loss: 0.7053 Generator Loss: 1.3877
Epoch [14/30] Batch 1300/1563 Discriminator Loss: 0.8480 Generator Loss: 0.9890
Epoch [14/30] Batch 1400/1563 Discriminator Loss: 0.3800 Generator Loss: 1.6878
Epoch [14/30] Batch 1500/1563 Discriminator Loss: 0.4641 Generator Loss: 1.5752
Epoch [15/30] Batch 100/1563 Discriminator Loss: 0.6491 Generator Loss: 1.2644
Epoch [15/30] Batch 200/1563 Discriminator Loss: 0.5590 Generator Loss: 0.9048
Epoch [15/30] Batch 300/1563 Discriminator Loss: 0.7712 Generator Loss: 1.0408
Epoch [15/30] Batch 400/1563 Discriminator Loss: 0.6446 Generator Loss: 1.2220
Epoch [15/30] Batch 500/1563 Discriminator Loss: 0.5891 Generator Loss: 0.8177
Epoch [15/30] Batch 600/1563 Discriminator Loss: 0.4539 Generator Loss: 1.5968
Epoch [15/30] Batch 700/1563 Discriminator Loss: 0.9036 Generator Loss: 1.2203
Epoch [15/30] Batch 800/1563 Discriminator Loss: 0.6573 Generator Loss: 1.2048
Epoch [15/30] Batch 900/1563 Discriminator Loss: 0.5110 Generator Loss: 0.8048
Epoch [15/30] Batch 1000/1563 Discriminator Loss: 0.6662 Generator Loss: 1.6221
Epoch [15/30] Batch 1100/1563 Discriminator Loss: 0.4542 Generator Loss: 1.4416
Epoch [15/30] Batch 1200/1563 Discriminator Loss: 0.5644 Generator Loss: 1.5006
Epoch [15/30] Batch 1300/1563 Discriminator Loss: 0.5741 Generator Loss: 1.5613
Epoch [15/30] Batch 1400/1563 Discriminator Loss: 0.5394 Generator Loss: 1.6310
Epoch [15/30] Batch 1500/1563 Discriminator Loss: 0.2746 Generator Loss: 1.2002
Epoch [16/30] Batch 100/1563 Discriminator Loss: 0.9168 Generator Loss: 1.0404
Epoch [16/30] Batch 200/1563 Discriminator Loss: 0.3522 Generator Loss: 2.0820
Epoch [16/30] Batch 300/1563 Discriminator Loss: 1.0174 Generator Loss: 1.2016
Epoch [16/30] Batch 400/1563 Discriminator Loss: 0.9039 Generator Loss: 0.9897
Epoch [16/30] Batch 500/1563 Discriminator Loss: 0.7441 Generator Loss: 1.5617
Epoch [16/30] Batch 600/1563 Discriminator Loss: 0.6339 Generator Loss: 0.7029
Epoch [16/30] Batch 700/1563 Discriminator Loss: 0.6705 Generator Loss: 1.2205
Epoch [16/30] Batch 800/1563 Discriminator Loss: 0.4608 Generator Loss: 1.9291
Epoch [16/30] Batch 900/1563 Discriminator Loss: 0.3003 Generator Loss: 2.2707
Epoch [16/30] Batch 1000/1563 Discriminator Loss: 0.4935 Generator Loss: 0.9453
Epoch [16/30] Batch 1100/1563 Discriminator Loss: 0.6426 Generator Loss: 2.0306
Epoch [16/30] Batch 1200/1563 Discriminator Loss: 0.3256 Generator Loss: 1.9450
Epoch [16/30] Batch 1300/1563 Discriminator Loss: 0.4460 Generator Loss: 1.3895
Epoch [16/30] Batch 1400/1563 Discriminator Loss: 0.5697 Generator Loss: 1.8839
Epoch [16/30] Batch 1500/1563 Discriminator Loss: 0.3510 Generator Loss: 1.2868
Epoch [17/30] Batch 100/1563 Discriminator Loss: 0.4061 Generator Loss: 1.5765
Epoch [17/30] Batch 200/1563 Discriminator Loss: 0.6059 Generator Loss: 1.4607
Epoch [17/30] Batch 300/1563 Discriminator Loss: 0.3888 Generator Loss: 2.0352
Epoch [17/30] Batch 400/1563 Discriminator Loss: 0.9055 Generator Loss: 1.5344
Epoch [17/30] Batch 500/1563 Discriminator Loss: 0.4919 Generator Loss: 2.0243
Epoch [17/30] Batch 600/1563 Discriminator Loss: 0.5129 Generator Loss: 1.3443
Epoch [17/30] Batch 700/1563 Discriminator Loss: 0.5908 Generator Loss: 1.3944
Epoch [17/30] Batch 800/1563 Discriminator Loss: 0.6198 Generator Loss: 1.9401
Epoch [17/30] Batch 900/1563 Discriminator Loss: 0.4722 Generator Loss: 1.3389
Epoch [17/30] Batch 1000/1563 Discriminator Loss: 0.6337 Generator Loss: 1.2892
Epoch [17/30] Batch 1100/1563 Discriminator Loss: 0.4962 Generator Loss: 1.3870
Epoch [17/30] Batch 1200/1563 Discriminator Loss: 0.4610 Generator Loss: 0.8908
Epoch [17/30] Batch 1300/1563 Discriminator Loss: 0.7297 Generator Loss: 1.1273
Epoch [17/30] Batch 1400/1563 Discriminator Loss: 0.7066 Generator Loss: 0.8907
Epoch [17/30] Batch 1500/1563 Discriminator Loss: 0.4157 Generator Loss: 1.4886
Epoch [18/30] Batch 100/1563 Discriminator Loss: 0.4349 Generator Loss: 2.3884
Epoch [18/30] Batch 200/1563 Discriminator Loss: 0.6107 Generator Loss: 0.9189
Epoch [18/30] Batch 300/1563 Discriminator Loss: 0.6645 Generator Loss: 0.9639
Epoch [18/30] Batch 400/1563 Discriminator Loss: 0.6846 Generator Loss: 1.6454
Epoch [18/30] Batch 500/1563 Discriminator Loss: 0.6662 Generator Loss: 1.3799
Epoch [18/30] Batch 600/1563 Discriminator Loss: 0.3959 Generator Loss: 1.6886
Epoch [18/30] Batch 700/1563 Discriminator Loss: 0.7645 Generator Loss: 0.4036
Epoch [18/30] Batch 800/1563 Discriminator Loss: 0.3437 Generator Loss: 1.5305
Epoch [18/30] Batch 900/1563 Discriminator Loss: 0.7294 Generator Loss: 1.9114
Epoch [18/30] Batch 1000/1563 Discriminator Loss: 0.3701 Generator Loss: 2.1974
Epoch [18/30] Batch 1100/1563 Discriminator Loss: 0.3689 Generator Loss: 1.6446
Epoch [18/30] Batch 1200/1563 Discriminator Loss: 0.4035 Generator Loss: 1.3158
Epoch [18/30] Batch 1300/1563 Discriminator Loss: 0.6356 Generator Loss: 1.4208
Epoch [18/30] Batch 1400/1563 Discriminator Loss: 0.6068 Generator Loss: 1.0243
Epoch [18/30] Batch 1500/1563 Discriminator Loss: 0.7674 Generator Loss: 1.6801
Epoch [19/30] Batch 100/1563 Discriminator Loss: 0.4846 Generator Loss: 1.2064
Epoch [19/30] Batch 200/1563 Discriminator Loss: 0.6070 Generator Loss: 1.7031
Epoch [19/30] Batch 300/1563 Discriminator Loss: 0.7450 Generator Loss: 1.0996
Epoch [19/30] Batch 400/1563 Discriminator Loss: 0.6458 Generator Loss: 1.0814
Epoch [19/30] Batch 500/1563 Discriminator Loss: 0.4784 Generator Loss: 1.4959
Epoch [19/30] Batch 600/1563 Discriminator Loss: 0.3522 Generator Loss: 1.8114
Epoch [19/30] Batch 700/1563 Discriminator Loss: 0.6813 Generator Loss: 1.4926
Epoch [19/30] Batch 800/1563 Discriminator Loss: 0.3082 Generator Loss: 1.2779
Epoch [19/30] Batch 900/1563 Discriminator Loss: 0.4817 Generator Loss: 1.6282
Epoch [19/30] Batch 1000/1563 Discriminator Loss: 0.6549 Generator Loss: 1.8198
Epoch [19/30] Batch 1100/1563 Discriminator Loss: 0.5999 Generator Loss: 2.1225
Epoch [19/30] Batch 1200/1563 Discriminator Loss: 0.2709 Generator Loss: 1.7114
Epoch [19/30] Batch 1300/1563 Discriminator Loss: 0.5228 Generator Loss: 1.3422
Epoch [19/30] Batch 1400/1563 Discriminator Loss: 0.6171 Generator Loss: 1.0001
Epoch [19/30] Batch 1500/1563 Discriminator Loss: 0.5075 Generator Loss: 1.5396
Epoch [20/30] Batch 100/1563 Discriminator Loss: 0.7957 Generator Loss: 1.0529
Epoch [20/30] Batch 200/1563 Discriminator Loss: 0.4388 Generator Loss: 0.9466
Epoch [20/30] Batch 300/1563 Discriminator Loss: 0.3565 Generator Loss: 1.0153
Epoch [20/30] Batch 400/1563 Discriminator Loss: 0.5776 Generator Loss: 1.3406
Epoch [20/30] Batch 500/1563 Discriminator Loss: 0.5628 Generator Loss: 1.7318
Epoch [20/30] Batch 600/1563 Discriminator Loss: 0.6587 Generator Loss: 1.0502
Epoch [20/30] Batch 700/1563 Discriminator Loss: 0.3938 Generator Loss: 1.5252
Epoch [20/30] Batch 800/1563 Discriminator Loss: 0.5031 Generator Loss: 1.4253
Epoch [20/30] Batch 900/1563 Discriminator Loss: 0.8166 Generator Loss: 0.8745
Epoch [20/30] Batch 1000/1563 Discriminator Loss: 0.8261 Generator Loss: 0.7913
Epoch [20/30] Batch 1100/1563 Discriminator Loss: 0.3707 Generator Loss: 1.8016
Epoch [20/30] Batch 1200/1563 Discriminator Loss: 0.5905 Generator Loss: 1.0425
Epoch [20/30] Batch 1300/1563 Discriminator Loss: 0.4316 Generator Loss: 1.2454
Epoch [20/30] Batch 1400/1563 Discriminator Loss: 0.6296 Generator Loss: 1.1763
Epoch [20/30] Batch 1500/1563 Discriminator Loss: 0.3132 Generator Loss: 1.5287
Epoch [21/30] Batch 100/1563 Discriminator Loss: 0.6704 Generator Loss: 0.9994
Epoch [21/30] Batch 200/1563 Discriminator Loss: 0.4294 Generator Loss: 1.6680
Epoch [21/30] Batch 300/1563 Discriminator Loss: 0.3051 Generator Loss: 1.7558
Epoch [21/30] Batch 400/1563 Discriminator Loss: 0.5725 Generator Loss: 1.0703
Epoch [21/30] Batch 500/1563 Discriminator Loss: 0.5754 Generator Loss: 0.8874
Epoch [21/30] Batch 600/1563 Discriminator Loss: 0.5921 Generator Loss: 0.9687
Epoch [21/30] Batch 700/1563 Discriminator Loss: 0.3578 Generator Loss: 2.0567
Epoch [21/30] Batch 800/1563 Discriminator Loss: 0.3976 Generator Loss: 1.6124
Epoch [21/30] Batch 900/1563 Discriminator Loss: 0.5874 Generator Loss: 1.4437
Epoch [21/30] Batch 1000/1563 Discriminator Loss: 0.3392 Generator Loss: 1.8707
Epoch [21/30] Batch 1100/1563 Discriminator Loss: 0.6126 Generator Loss: 0.8583
Epoch [21/30] Batch 1200/1563 Discriminator Loss: 0.6787 Generator Loss: 1.2624
Epoch [21/30] Batch 1300/1563 Discriminator Loss: 0.4979 Generator Loss: 1.1952
Epoch [21/30] Batch 1400/1563 Discriminator Loss: 0.4609 Generator Loss: 1.2393
Epoch [21/30] Batch 1500/1563 Discriminator Loss: 0.4389 Generator Loss: 1.0301
Epoch [22/30] Batch 100/1563 Discriminator Loss: 0.5087 Generator Loss: 0.8303
Epoch [22/30] Batch 200/1563 Discriminator Loss: 0.5395 Generator Loss: 1.7216
Epoch [22/30] Batch 300/1563 Discriminator Loss: 0.7014 Generator Loss: 1.3354
Epoch [22/30] Batch 400/1563 Discriminator Loss: 0.4920 Generator Loss: 1.2890
Epoch [22/30] Batch 500/1563 Discriminator Loss: 0.6030 Generator Loss: 0.8941
Epoch [22/30] Batch 600/1563 Discriminator Loss: 0.5582 Generator Loss: 0.8958
Epoch [22/30] Batch 700/1563 Discriminator Loss: 0.5013 Generator Loss: 2.0577
Epoch [22/30] Batch 800/1563 Discriminator Loss: 0.6041 Generator Loss: 1.3337
Epoch [22/30] Batch 900/1563 Discriminator Loss: 0.5399 Generator Loss: 0.9913
Epoch [22/30] Batch 1000/1563 Discriminator Loss: 0.6801 Generator Loss: 0.9158
Epoch [22/30] Batch 1100/1563 Discriminator Loss: 0.4355 Generator Loss: 1.8976
Epoch [22/30] Batch 1200/1563 Discriminator Loss: 0.6911 Generator Loss: 1.3129
Epoch [22/30] Batch 1300/1563 Discriminator Loss: 0.5185 Generator Loss: 1.0487
Epoch [22/30] Batch 1400/1563 Discriminator Loss: 0.7541 Generator Loss: 1.2654
Epoch [22/30] Batch 1500/1563 Discriminator Loss: 0.4369 Generator Loss: 1.2044
Epoch [23/30] Batch 100/1563 Discriminator Loss: 0.5012 Generator Loss: 0.8639
Epoch [23/30] Batch 200/1563 Discriminator Loss: 0.7417 Generator Loss: 1.0816
Epoch [23/30] Batch 300/1563 Discriminator Loss: 0.5031 Generator Loss: 1.6705
Epoch [23/30] Batch 400/1563 Discriminator Loss: 0.7800 Generator Loss: 1.1085
Epoch [23/30] Batch 500/1563 Discriminator Loss: 0.6581 Generator Loss: 0.7194
Epoch [23/30] Batch 600/1563 Discriminator Loss: 0.6880 Generator Loss: 1.2136
Epoch [23/30] Batch 700/1563 Discriminator Loss: 0.8604 Generator Loss: 1.2978
Epoch [23/30] Batch 800/1563 Discriminator Loss: 0.4069 Generator Loss: 1.5116
Epoch [23/30] Batch 900/1563 Discriminator Loss: 0.6084 Generator Loss: 1.3218
Epoch [23/30] Batch 1000/1563 Discriminator Loss: 0.3370 Generator Loss: 2.3237
Epoch [23/30] Batch 1100/1563 Discriminator Loss: 0.5806 Generator Loss: 1.6243
Epoch [23/30] Batch 1200/1563 Discriminator Loss: 0.5700 Generator Loss: 1.3030
Epoch [23/30] Batch 1300/1563 Discriminator Loss: 0.4817 Generator Loss: 1.0187
Epoch [23/30] Batch 1400/1563 Discriminator Loss: 0.5350 Generator Loss: 0.9343
Epoch [23/30] Batch 1500/1563 Discriminator Loss: 0.3988 Generator Loss: 1.9423
Epoch [24/30] Batch 100/1563 Discriminator Loss: 0.5867 Generator Loss: 1.8228
Epoch [24/30] Batch 200/1563 Discriminator Loss: 0.6380 Generator Loss: 0.7886
Epoch [24/30] Batch 300/1563 Discriminator Loss: 1.0712 Generator Loss: 1.7067
Epoch [24/30] Batch 400/1563 Discriminator Loss: 0.3691 Generator Loss: 1.6998
Epoch [24/30] Batch 500/1563 Discriminator Loss: 0.7596 Generator Loss: 0.6284
Epoch [24/30] Batch 600/1563 Discriminator Loss: 0.8459 Generator Loss: 1.5625
Epoch [24/30] Batch 700/1563 Discriminator Loss: 0.4499 Generator Loss: 1.0025
Epoch [24/30] Batch 800/1563 Discriminator Loss: 0.3667 Generator Loss: 1.3655
Epoch [24/30] Batch 900/1563 Discriminator Loss: 0.8259 Generator Loss: 1.6259
Epoch [24/30] Batch 1000/1563 Discriminator Loss: 0.8847 Generator Loss: 1.0071
Epoch [24/30] Batch 1100/1563 Discriminator Loss: 0.2908 Generator Loss: 2.2473
Epoch [24/30] Batch 1200/1563 Discriminator Loss: 0.8809 Generator Loss: 0.8121
Epoch [24/30] Batch 1300/1563 Discriminator Loss: 0.5599 Generator Loss: 0.8316
Epoch [24/30] Batch 1400/1563 Discriminator Loss: 0.5931 Generator Loss: 1.6818
Epoch [24/30] Batch 1500/1563 Discriminator Loss: 0.4780 Generator Loss: 1.4393
Epoch [25/30] Batch 100/1563 Discriminator Loss: 0.4862 Generator Loss: 1.0551
Epoch [25/30] Batch 200/1563 Discriminator Loss: 0.7419 Generator Loss: 0.8418
Epoch [25/30] Batch 300/1563 Discriminator Loss: 0.3005 Generator Loss: 1.4868
Epoch [25/30] Batch 400/1563 Discriminator Loss: 0.4458 Generator Loss: 0.8691
Epoch [25/30] Batch 500/1563 Discriminator Loss: 0.4591 Generator Loss: 1.3558
Epoch [25/30] Batch 600/1563 Discriminator Loss: 0.2917 Generator Loss: 1.1712
Epoch [25/30] Batch 700/1563 Discriminator Loss: 0.6088 Generator Loss: 1.1128
Epoch [25/30] Batch 800/1563 Discriminator Loss: 0.8086 Generator Loss: 1.5543
Epoch [25/30] Batch 900/1563 Discriminator Loss: 0.5244 Generator Loss: 1.4409
Epoch [25/30] Batch 1000/1563 Discriminator Loss: 0.6720 Generator Loss: 1.2794
Epoch [25/30] Batch 1100/1563 Discriminator Loss: 0.5955 Generator Loss: 1.0705
Epoch [25/30] Batch 1200/1563 Discriminator Loss: 0.6215 Generator Loss: 1.2729
Epoch [25/30] Batch 1300/1563 Discriminator Loss: 1.2661 Generator Loss: 1.4346
Epoch [25/30] Batch 1400/1563 Discriminator Loss: 0.6111 Generator Loss: 1.4205
Epoch [25/30] Batch 1500/1563 Discriminator Loss: 0.5564 Generator Loss: 1.5329
Epoch [26/30] Batch 100/1563 Discriminator Loss: 0.7150 Generator Loss: 1.1040
Epoch [26/30] Batch 200/1563 Discriminator Loss: 0.5486 Generator Loss: 0.8081
Epoch [26/30] Batch 300/1563 Discriminator Loss: 0.4655 Generator Loss: 1.0543
Epoch [26/30] Batch 400/1563 Discriminator Loss: 0.6280 Generator Loss: 1.3089
Epoch [26/30] Batch 500/1563 Discriminator Loss: 0.6435 Generator Loss: 1.2938
Epoch [26/30] Batch 600/1563 Discriminator Loss: 0.5681 Generator Loss: 0.6704
Epoch [26/30] Batch 700/1563 Discriminator Loss: 0.5975 Generator Loss: 1.6594
Epoch [26/30] Batch 800/1563 Discriminator Loss: 0.5214 Generator Loss: 1.5375
Epoch [26/30] Batch 900/1563 Discriminator Loss: 0.7233 Generator Loss: 2.1745
Epoch [26/30] Batch 1000/1563 Discriminator Loss: 0.7600 Generator Loss: 1.9514
Epoch [26/30] Batch 1100/1563 Discriminator Loss: 0.4599 Generator Loss: 1.6351
Epoch [26/30] Batch 1200/1563 Discriminator Loss: 0.4130 Generator Loss: 0.6645
Epoch [26/30] Batch 1300/1563 Discriminator Loss: 0.7173 Generator Loss: 1.5409
Epoch [26/30] Batch 1400/1563 Discriminator Loss: 0.6444 Generator Loss: 0.4363
Epoch [26/30] Batch 1500/1563 Discriminator Loss: 0.8130 Generator Loss: 0.5472
Epoch [27/30] Batch 100/1563 Discriminator Loss: 0.3806 Generator Loss: 2.5330
Epoch [27/30] Batch 200/1563 Discriminator Loss: 0.6471 Generator Loss: 1.4548
Epoch [27/30] Batch 300/1563 Discriminator Loss: 0.4945 Generator Loss: 2.0417
Epoch [27/30] Batch 400/1563 Discriminator Loss: 0.8287 Generator Loss: 0.8313
Epoch [27/30] Batch 500/1563 Discriminator Loss: 0.2590 Generator Loss: 1.3648
Epoch [27/30] Batch 600/1563 Discriminator Loss: 0.7229 Generator Loss: 1.0506
Epoch [27/30] Batch 700/1563 Discriminator Loss: 0.5257 Generator Loss: 0.6630
Epoch [27/30] Batch 800/1563 Discriminator Loss: 0.7077 Generator Loss: 1.5787
Epoch [27/30] Batch 900/1563 Discriminator Loss: 0.2241 Generator Loss: 2.4144
Epoch [27/30] Batch 1000/1563 Discriminator Loss: 0.9249 Generator Loss: 1.0945
Epoch [27/30] Batch 1100/1563 Discriminator Loss: 0.5512 Generator Loss: 1.3305
Epoch [27/30] Batch 1200/1563 Discriminator Loss: 0.6963 Generator Loss: 0.6857
Epoch [27/30] Batch 1300/1563 Discriminator Loss: 0.6866 Generator Loss: 1.0320
Epoch [27/30] Batch 1400/1563 Discriminator Loss: 0.8760 Generator Loss: 0.5938
Epoch [27/30] Batch 1500/1563 Discriminator Loss: 0.6564 Generator Loss: 1.1705
Epoch [28/30] Batch 100/1563 Discriminator Loss: 0.6794 Generator Loss: 0.6070
Epoch [28/30] Batch 200/1563 Discriminator Loss: 0.5660 Generator Loss: 1.3422
Epoch [28/30] Batch 300/1563 Discriminator Loss: 0.4716 Generator Loss: 1.3143
Epoch [28/30] Batch 400/1563 Discriminator Loss: 0.5311 Generator Loss: 1.8503
Epoch [28/30] Batch 500/1563 Discriminator Loss: 0.5435 Generator Loss: 1.4696
Epoch [28/30] Batch 600/1563 Discriminator Loss: 0.4778 Generator Loss: 1.1126
Epoch [28/30] Batch 700/1563 Discriminator Loss: 0.4951 Generator Loss: 0.9386
Epoch [28/30] Batch 800/1563 Discriminator Loss: 0.5161 Generator Loss: 1.7677
Epoch [28/30] Batch 900/1563 Discriminator Loss: 0.5352 Generator Loss: 1.3093
Epoch [28/30] Batch 1000/1563 Discriminator Loss: 0.4307 Generator Loss: 1.0498
Epoch [28/30] Batch 1100/1563 Discriminator Loss: 0.8755 Generator Loss: 1.3385
Epoch [28/30] Batch 1200/1563 Discriminator Loss: 0.5321 Generator Loss: 1.0425
Epoch [28/30] Batch 1300/1563 Discriminator Loss: 0.6281 Generator Loss: 0.8529
Epoch [28/30] Batch 1400/1563 Discriminator Loss: 0.5690 Generator Loss: 1.3257
Epoch [28/30] Batch 1500/1563 Discriminator Loss: 1.0743 Generator Loss: 1.1915
Epoch [29/30] Batch 100/1563 Discriminator Loss: 0.6799 Generator Loss: 1.0841
Epoch [29/30] Batch 200/1563 Discriminator Loss: 0.6932 Generator Loss: 1.3634
Epoch [29/30] Batch 300/1563 Discriminator Loss: 0.4663 Generator Loss: 1.1661
Epoch [29/30] Batch 400/1563 Discriminator Loss: 0.7206 Generator Loss: 0.9654
Epoch [29/30] Batch 500/1563 Discriminator Loss: 0.7813 Generator Loss: 1.4285
Epoch [29/30] Batch 600/1563 Discriminator Loss: 0.8215 Generator Loss: 1.1129
Epoch [29/30] Batch 700/1563 Discriminator Loss: 0.5939 Generator Loss: 1.5154
Epoch [29/30] Batch 800/1563 Discriminator Loss: 0.6840 Generator Loss: 1.2734
Epoch [29/30] Batch 900/1563 Discriminator Loss: 0.4068 Generator Loss: 1.2772
Epoch [29/30] Batch 1000/1563 Discriminator Loss: 0.6416 Generator Loss: 0.4559
Epoch [29/30] Batch 1100/1563 Discriminator Loss: 0.4972 Generator Loss: 1.6791
Epoch [29/30] Batch 1200/1563 Discriminator Loss: 0.8609 Generator Loss: 1.1342
Epoch [29/30] Batch 1300/1563 Discriminator Loss: 0.6266 Generator Loss: 0.8422
Epoch [29/30] Batch 1400/1563 Discriminator Loss: 0.5542 Generator Loss: 1.1366
Epoch [29/30] Batch 1500/1563 Discriminator Loss: 0.5414 Generator Loss: 1.0412
Epoch [30/30] Batch 100/1563 Discriminator Loss: 1.0110 Generator Loss: 0.9997
Epoch [30/30] Batch 200/1563 Discriminator Loss: 0.4734 Generator Loss: 1.2250
Epoch [30/30] Batch 300/1563 Discriminator Loss: 0.3835 Generator Loss: 1.3170
Epoch [30/30] Batch 400/1563 Discriminator Loss: 0.3179 Generator Loss: 2.2588
Epoch [30/30] Batch 500/1563 Discriminator Loss: 0.3329 Generator Loss: 1.5280
Epoch [30/30] Batch 600/1563 Discriminator Loss: 0.3452 Generator Loss: 0.8436
Epoch [30/30] Batch 700/1563 Discriminator Loss: 0.8666 Generator Loss: 0.6735
Epoch [30/30] Batch 800/1563 Discriminator Loss: 0.6920 Generator Loss: 1.8045
Epoch [30/30] Batch 900/1563 Discriminator Loss: 0.6625 Generator Loss: 0.8388
Epoch [30/30] Batch 1000/1563 Discriminator Loss: 0.7458 Generator Loss: 2.1390
Epoch [30/30] Batch 1100/1563 Discriminator Loss: 1.4766 Generator Loss: 1.7256
Epoch [30/30] Batch 1200/1563 Discriminator Loss: 0.4399 Generator Loss: 1.4570
Epoch [30/30] Batch 1300/1563 Discriminator Loss: 0.4310 Generator Loss: 1.7740
Epoch [30/30] Batch 1400/1563 Discriminator Loss: 0.4642 Generator Loss: 1.1918
Epoch [30/30] Batch 1500/1563 Discriminator Loss: 0.5108 Generator Loss: 1.7462
Visualização da Arquitetura
-
Goodfellow, I., et al. (2014). Generative Adversarial Networks — NeurIPS. O problema da saturação e a correção não saturante estão os dois na seção 3. ↩
-
Radford, A., Metz, L., & Chintala, S. (2016). Unsupervised Representation Learning with Deep Convolutional GANs — ICLR. DCGAN. ↩
-
Silva, T. — A Short Introduction to Generative Adversarial Networks. ↩
-
Karras, T., Laine, S., & Aila, T. (2019). A Style-Based Generator Architecture for Generative Adversarial Networks — CVPR. StyleGAN e o espaço latente \(\mathcal{W}\). ↩
-
Kang, M., et al. (2023). Scaling up GANs for Text-to-Image Synthesis — CVPR. GigaGAN: qualidade competitiva, ordens de grandeza mais rápido e largamente ignorado porque a área já tinha se mudado. ↩
-
Gulrajani, I., et al. (2017). Improved Training of Wasserstein GANs — NeurIPS. ↩
-
Miyato, T., et al. (2018). Spectral Normalization for Generative Adversarial Networks — ICLR. Uma linha, efeito grande. ↩
-
Mescheder, L., Geiger, A., & Nowozin, S. (2018). Which Training Methods for GANs do actually Converge? — ICML. De onde vem o R1, com a análise de convergência. ↩
-
Karras, T., et al. (2020). Training Generative Adversarial Networks with Limited Data — NeurIPS. Augmentation adaptativa do discriminador. ↩
-
Esser, P., Rombach, R., & Ommer, B. (2021). Taming Transformers for High-Resolution Image Synthesis — CVPR. VQGAN: o decodificador adversarial + perceptual que todo modelo de difusão latente herdou. ↩
-
Sauer, A., Lorenz, D., Blattmann, A., & Rombach, R. (2023). Adversarial Diffusion Distillation. SDXL-Turbo; um discriminador é o que faz a amostragem de um passo parecer real. ↩