# 📚 PyTorch Tutorial for Beginners - Part 5/6: Generative Models & Advanced Topics
#PyTorch #DeepLearning #GANs #VAEs #ReinforcementLearning #Deployment

Welcome to Part 5 of our PyTorch series! This comprehensive lesson explores generative modeling, reinforcement learning, model optimization, and deployment strategies with practical implementations.

---

## 🔹 Generative Adversarial Networks (GANs)
### 1. GAN Core Concepts

Key Components:
- Generator: Creates fake samples from noise (typically a transposed CNN)
- Discriminator: Distinguishes real vs. fake samples (CNN classifier)
- Adversarial Training: The two networks compete in a minimax game

### 2. DCGAN Implementation

class Generator(nn.Module):
    def __init__(self, latent_dim, img_channels, features_g):
        super().__init__()
        self.net = nn.Sequential(
            # Input: N x latent_dim x 1 x 1
            nn.ConvTranspose2d(latent_dim, features_g*8, 4, 1, 0, bias=False),
            nn.BatchNorm2d(features_g*8),
            nn.ReLU(),
            # 4x4
            nn.ConvTranspose2d(features_g*8, features_g*4, 4, 2, 1, bias=False),
            nn.BatchNorm2d(features_g*4),
            nn.ReLU(),
            # 8x8
            nn.ConvTranspose2d(features_g*4, features_g*2, 4, 2, 1, bias=False),
            nn.BatchNorm2d(features_g*2),
            nn.ReLU(),
            # 16x16
            nn.ConvTranspose2d(features_g*2, img_channels, 4, 2, 1, bias=False),
            nn.Tanh()
            # 32x32
        )

    def forward(self, x):
        return self.net(x)

class Discriminator(nn.Module):
    def __init__(self, img_channels, features_d):
        super().__init__()
        self.net = nn.Sequential(
            # Input: N x img_channels x 32 x 32
            nn.Conv2d(img_channels, features_d, 4, 2, 1, bias=False),
            nn.LeakyReLU(0.2),
            # 16x16
            nn.Conv2d(features_d, features_d*2, 4, 2, 1, bias=False),
            nn.BatchNorm2d(features_d*2),
            nn.LeakyReLU(0.2),
            # 8x8
            nn.Conv2d(features_d*2, features_d*4, 4, 2, 1, bias=False),
            nn.BatchNorm2d(features_d*4),
            nn.LeakyReLU(0.2),
            # 4x4
            nn.Conv2d(features_d*4, 1, 4, 1, 0, bias=False),
            nn.Sigmoid()
        )

    def forward(self, x):
        return self.net(x)

# Initialize
gen = Generator(latent_dim=100, img_channels=3, features_g=64).to(device)
disc = Discriminator(img_channels=3, features_d=64).to(device)

# Loss and optimizers
criterion = nn.BCELoss()
opt_gen = optim.Adam(gen.parameters(), lr=0.0002, betas=(0.5, 0.999))
opt_disc = optim.Adam(disc.parameters(), lr=0.0002, betas=(0.5, 0.999))

### 3. GAN Training Loop

def train_gan(gen, disc, loader, num_epochs):
    fixed_noise = torch.randn(32, 100, 1, 1).to(device)
    
    for epoch in range(num_epochs):
        for batch_idx, (real, _) in enumerate(loader):
            real = real.to(device)
            noise = torch.randn(real.size(0), 100, 1, 1).to(device)
            fake = gen(noise)
            
            # Train Discriminator
            disc_real = disc(real).view(-1)
            loss_disc_real = criterion(disc_real, torch.ones_like(disc_real))
            disc_fake = disc(fake.detach()).view(-1)
            loss_disc_fake = criterion(disc_fake, torch.zeros_like(disc_fake))
            loss_disc = (loss_disc_real + loss_disc_fake) / 2
            disc.zero_grad()
            loss_disc.backward()
            opt_disc.step()
            
            # Train Generator
            output = disc(fake).view(-1)
            loss_gen = criterion(output, torch.ones_like(output))
            gen.zero_grad()
            loss_gen.backward()
            opt_gen.step()
            
        # Visualization
        with torch.no_grad():
            fake = gen(fixed_noise)
            save_image(fake, f"gan_samples/epoch_{epoch}.png", normalize=True)

### 2. Pruning

parameters_to_prune = (
    (model.conv1, 'weight'),
    (model.fc1, 'weight'),
)

prune.global_unstructured(
    parameters_to_prune,
    pruning_method=prune.L1Unstructured,
    amount=0.2
)

# Remove pruning reparameterization
for module, param in parameters_to_prune:
    prune.remove(module, param)

### 3. ONNX Export

dummy_input = torch.randn(1, 3, 224, 224)
torch.onnx.export(
    model,
    dummy_input,
    "model.onnx",
    input_names=["input"],
    output_names=["output"],
    dynamic_axes={
        "input": {0: "batch_size"},
        "output": {0: "batch_size"}
    }
)

### 4. TorchScript

# Tracing
example_input = torch.rand(1, 3, 224, 224)
traced_script = torch.jit.trace(model, example_input)
traced_script.save("traced_model.pt")

# Scripting
scripted_model = torch.jit.script(model)
scripted_model.save("scripted_model.pt")

---

## 🔹 PyTorch Lightning Best Practices
### 1. LightningModule Structure

import pytorch_lightning as pl

class LitModel(pl.LightningModule):
    def __init__(self, learning_rate=1e-3):
        super().__init__()
        self.save_hyperparameters()
        self.model = nn.Sequential(
            nn.Linear(28*28, 128),
            nn.ReLU(),
            nn.Linear(128, 10)
        )
    
    def forward(self, x):
        return self.model(x)
    
    def training_step(self, batch, batch_idx):
        x, y = batch
        y_hat = self(x)
        loss = nn.functional.cross_entropy(y_hat, y)
        self.log('train_loss', loss)
        return loss
    
    def validation_step(self, batch, batch_idx):
        x, y = batch
        y_hat = self(x)
        loss = nn.functional.cross_entropy(y_hat, y)
        self.log('val_loss', loss)
    
    def configure_optimizers(self):
        return optim.Adam(self.parameters(), lr=self.hparams.learning_rate)

# Training
trainer = pl.Trainer(gpus=1, max_epochs=10)
model = LitModel()
trainer.fit(model, train_loader, val_loader)

### 2. Advanced Lightning Features

# Mixed Precision
trainer = pl.Trainer(precision=16)

# Distributed Training
trainer = pl.Trainer(gpus=2, accelerator='ddp')

# Callbacks
early_stop = pl.callbacks.EarlyStopping(monitor='val_loss')
checkpoint = pl.callbacks.ModelCheckpoint(monitor='val_loss')
trainer = pl.Trainer(callbacks=[early_stop, checkpoint])

# Logging
trainer = pl.Trainer(logger=pl.loggers.TensorBoardLogger('logs/'))

---

## 🔹 Best Practices Summary
1. For GANs: Use spectral norm, progressive growing, and TTUR
2. For VAEs: Monitor both reconstruction and KL divergence terms
3. For RL: Properly normalize rewards and use experience replay
4. For Deployment: Quantize, prune, and export to optimized formats
5. For Maintenance: Use PyTorch Lightning for reproducible experiments

---

### 📌 What's Next?
In Part 6 (Final), we'll cover:
➡️ Advanced Architectures (Graph NNs, Neural ODEs)
➡️ Model Interpretation Techniques
➡️ Production Deployment (TorchServe, Flask API)
➡️ PyTorch Ecosystem (TorchVision, TorchText, TorchAudio)

#PyTorch #DeepLearning #GANs #ReinforcementLearning 🚀

Practice Exercises:
1. Implement WGAN-GP with gradient penalty
2. Train a VAE on MNIST and visualize latent space
3. Build a DQN agent for CartPole environment
4. Quantize a pretrained ResNet and compare accuracy/speed
5. Convert a model to TorchScript and serve with Flask

# WGAN-GP Gradient Penalty
def compute_gradient_penalty(D, real_samples, fake_samples):
    alpha = torch.rand(real_samples.size(0), 1, 1, 1).to(device)
    interpolates = (alpha * real_samples + (1 - alpha) * fake_samples).requires_grad_(True)
    d_interpolates = D(interpolates)
    gradients = torch.autograd.grad(
        outputs=d_interpolates,
        inputs=interpolates,
        grad_outputs=torch.ones_like(d_interpolates),
        create_graph=True,
        retain_graph=True,
        only_inputs=True
    )[0]
    gradients = gradients.view(gradients.size(0), -1)
    gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean()
    return gradient_penalty

🤖🧠 NVIDIA, MIT, HKU and Tsinghua University Introduce QeRL: A Powerful Quantum Leap in Reinforcement Learning for LLMs

🗓️ 17 Oct 2025
📚 AI News & Trends

The rise of large language models (LLMs) has redefined artificial intelligence powering everything from conversational AI to autonomous reasoning systems. However, training these models especially through reinforcement learning (RL) is computationally expensive requiring massive GPU resources and long training cycles. To address this, a team of researchers from NVIDIA, Massachusetts Institute of Technology (MIT), The ...

#QuantumLearning #ReinforcementLearning #LLMs #NVIDIA #MIT #TsinghuaUniversity

🤖🧠 Agentic Entropy-Balanced Policy Optimization (AEPO): Balancing Exploration and Stability in Reinforcement Learning for Web Agents

🗓️ 17 Oct 2025
📚 AI News & Trends

AEPO (Agentic Entropy-Balanced Policy Optimization) represents a major advancement in the evolution of Agentic Reinforcement Learning (RL). As large language models (LLMs) increasingly act as autonomous web agents – searching, reasoning and interacting with tools – the need for balanced exploration and stability has become crucial. Traditional RL methods often rely heavily on entropy to ...

#AgenticRL #ReinforcementLearning #LLMs #WebAgents #EntropyBalanced #PolicyOptimization