### 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

# 📚 PyTorch Tutorial for Beginners - Part 6/6: Advanced Architectures & Production Deployment
#PyTorch #DeepLearning #GraphNNs #NeuralODEs #ModelServing #ExplainableAI

Welcome to the final part of our PyTorch series! This comprehensive lesson covers cutting-edge architectures, model interpretation techniques, production deployment strategies, and the broader PyTorch ecosystem.

---

## 🔹 Graph Neural Networks (GNNs)
### 1. Core Concepts

Key Components:
- Node Features: Characteristics of each graph node
- Edge Features: Properties of connections between nodes
- Message Passing: Nodes aggregate information from neighbors
- Graph Pooling: Reduces graph to fixed-size representation

### 2. Implementing GNN with PyTorch Geometric

import torch_geometric as tg
from torch_geometric.nn import GCNConv, global_mean_pool

class GNN(torch.nn.Module):
    def __init__(self, node_features, hidden_dim, num_classes):
        super().__init__()
        self.conv1 = GCNConv(node_features, hidden_dim)
        self.conv2 = GCNConv(hidden_dim, hidden_dim)
        self.classifier = nn.Linear(hidden_dim, num_classes)
        
    def forward(self, data):
        x, edge_index, batch = data.x, data.edge_index, data.batch
        
        # Message passing
        x = self.conv1(x, edge_index).relu()
        x = self.conv2(x, edge_index)
        
        # Graph-level pooling
        x = global_mean_pool(x, batch)
        
        # Classification
        return self.classifier(x)

# Example usage
dataset = tg.datasets.Planetoid(root='/tmp/Cora', name='Cora')
model = GNN(node_features=dataset.num_node_features, 
           hidden_dim=64, 
           num_classes=dataset.num_classes).to(device)

# Specialized DataLoader
loader = tg.data.DataLoader(dataset, batch_size=32, shuffle=True)

### 3. Advanced GNN Architectures

# Graph Attention Network (GAT)
class GAT(torch.nn.Module):
    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.conv1 = tg.nn.GATConv(in_channels, 8, heads=8, dropout=0.6)
        self.conv2 = tg.nn.GATConv(8*8, out_channels, heads=1, concat=False, dropout=0.6)
        
    def forward(self, data):
        x, edge_index = data.x, data.edge_index
        x = F.dropout(x, p=0.6, training=self.training)
        x = F.elu(self.conv1(x, edge_index))
        x = F.dropout(x, p=0.6, training=self.training)
        x = self.conv2(x, edge_index)
        return F.log_softmax(x, dim=1)

# Graph Isomorphism Network (GIN)
class GIN(torch.nn.Module):
    def __init__(self, in_channels, hidden_channels, out_channels):
        super().__init__()
        self.conv1 = tg.nn.GINConv(
            nn.Sequential(
                nn.Linear(in_channels, hidden_channels),
                nn.ReLU(),
                nn.Linear(hidden_channels, hidden_channels)
            ), train_eps=True)
        self.conv2 = tg.nn.GINConv(
            nn.Sequential(
                nn.Linear(hidden_channels, hidden_channels),
                nn.ReLU(),
                nn.Linear(hidden_channels, out_channels)
            ), train_eps=True)
        
    def forward(self, data):
        x, edge_index = data.x, data.edge_index
        x = self.conv1(x, edge_index)
        x = F.relu(x)
        x = self.conv2(x, edge_index)
        return x

---

## 🔹 Neural Ordinary Differential Equations (Neural ODEs)
### 1. Core Concepts

- Continuous-depth networks: Replace discrete layers with ODE solver
- Memory efficiency: Constant memory cost regardless of "depth"
- Adaptive computation: ODE solver adjusts evaluation points

### 2. Implementation with TorchDiffEq

from torchdiffeq import odeint_adjoint as odeint

class ODEBlock(nn.Module):
    def __init__(self, odefunc):
        super().__init__()
        self.odefunc = odefunc
        self.integration_time = torch.tensor([0, 1]).float()
        
    def forward(self, x):
        self.integration_time = self.integration_time.to(x.device)
        out = odeint(self.odefunc, x, self.integration_time, 
                    rtol=1e-3, atol=1e-4)
        return out[1]

class ODEFunc(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(dim, dim),
            nn.Tanh(),
            nn.Linear(dim, dim)
        )
        
    def forward(self, t, x):
        return self.net(x)

class NeuralODE(nn.Module):
    def __init__(self, input_dim, hidden_dim, output_dim):
        super().__init__()
        self.downsampling = nn.Linear(input_dim, hidden_dim)
        self.odeblock = ODEBlock(ODEFunc(hidden_dim))
        self.upsampling = nn.Linear(hidden_dim, output_dim)
        
    def forward(self, x):
        x = self.downsampling(x)
        x = self.odeblock(x)
        return self.upsampling(x)

### 3. Applications

# Time-series prediction
ode_ts = NeuralODE(input_dim=10, hidden_dim=64, output_dim=5)

# Continuous normalizing flows
class CNF(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.odefunc = ODEFunc(dim)
        self.odeblock = ODEBlock(self.odefunc)
        
    def forward(self, x):
        return self.odeblock(x)

---

## 🔹 Model Interpretation Techniques
### 1. SHAP Values

import shap

# Create explainer
background = train_data[:100].to(device)
explainer = shap.DeepExplainer(model, background)

# Calculate SHAP values
test_sample = test_data[0:1].to(device)
shap_values = explainer.shap_values(test_sample)

# Visualize
shap.image_plot(shap_values, -test_sample.cpu().numpy())

### 2. Integrated Gradients

from captum.attr import IntegratedGradients

ig = IntegratedGradients(model)
attributions = ig.attribute(input_tensor, 
                          target=pred_class_idx,
                          n_steps=50)

# Visualization
plt.imshow(attributions[0].cpu().detach().numpy().transpose(1,2,0))
plt.colorbar()
plt.show()

### 3. Attention Visualization

# For transformer models
def plot_attention(attention_weights, input_tokens):
    fig, ax = plt.subplots(figsize=(10, 10))
    im = ax.imshow(attention_weights.cpu().detach().numpy())
    
    ax.set_xticks(range(len(input_tokens)))
    ax.set_yticks(range(len(input_tokens))))
    ax.set_xticklabels(input_tokens, rotation=45)
    ax.set_yticklabels(input_tokens)
    
    plt.colorbar(im)
    plt.show()

---

## 🔹 Production Deployment
### 1. TorchServe

# Package model
torch-model-archiver --model-name mymodel --version 1.0 \
                    --serialized-file model.pth \
                    --export-path model_store \
                    --handler my_handler.py

# Start server
torchserve --start --model-store model_store --models mymodel=mymodel.mar

# Query model
curl https://localhost:8080/predictions/mymodel -T sample_input.json

### 2. Flask API

from flask import Flask, request, jsonify
import torch

app = Flask(__name__)
model = torch.load('model.pth', map_location='cpu')
model.eval()

@app.route('/predict', methods=['POST'])
def predict():
    data = request.get_json()
    tensor = torch.FloatTensor(data['input'])
    with torch.no_grad():
        output = model(tensor)
    return jsonify({'prediction': output.tolist()})

if __name__ == '__main__':
    app.run(host='0.0.0.0', port=5000)

### 3. ONNX Runtime

import onnxruntime as ort

# Create inference session
ort_session = ort.InferenceSession("model.onnx")

# Run inference
inputs = {"input": input_array}
outputs = ort_session.run(None, inputs)

### 4. TensorRT Optimization

# Convert ONNX to TensorRT
trt_logger = trt.Logger(trt.Logger.WARNING)
with trt.Builder(trt_logger) as builder:
    with builder.create_network(1) as network:
        with trt.OnnxParser(network, trt_logger) as parser:
            with open("model.onnx", "rb") as model:
                parser.parse(model.read())
            engine = builder.build_cuda_engine(network)

---

## 🔹 PyTorch Ecosystem
### 1. TorchVision

from torchvision.models import efficientnet_b0
from torchvision.ops import nms, roi_align

# Pretrained models
model = efficientnet_b0(pretrained=True)

# Computer vision ops
boxes = torch.tensor([[10, 20, 50, 60], [15, 25, 40, 70]])
scores = torch.tensor([0.9, 0.8])
keep = nms(boxes, scores, iou_threshold=0.5)

### 2. TorchText

from torchtext.data import Field, BucketIterator
from torchtext.datasets import IMDB

# Define fields
TEXT = Field(tokenize='spacy', lower=True, include_lengths=True)
LABEL = Field(sequential=False, dtype=torch.float)

# Load dataset
train_data, test_data = IMDB.splits(TEXT, LABEL)

# Build vocabulary
TEXT.build_vocab(train_data, max_size=25000)
LABEL.build_vocab(train_data)

### 3. TorchAudio

import torchaudio
import torchaudio.transforms as T

# Load audio
waveform, sample_rate = torchaudio.load('audio.wav')

# Spectrogram
spectrogram = T.Spectrogram()(waveform)

# MFCC
mfcc = T.MFCC(sample_rate=sample_rate)(waveform)

# Audio augmentation
augmented = T.TimeStretch()(waveform, n_freq=0.5)

---

## 🔹 Best Practices Summary
1. For GNNs: Normalize node features and use appropriate pooling
2. For Neural ODEs: Monitor ODE solver statistics during training
3. For Interpretability: Combine multiple explanation methods
4. For Deployment: Profile models before deployment (latency/throughput)
5. For Production: Implement monitoring for model drift

---

### 📌 Final Thoughts
Congratulations on completing this comprehensive PyTorch journey! You've learned:

✔️ Core PyTorch fundamentals
✔️ Deep neural networks & CNNs
✔️ Sequence modeling with RNNs/Transformers
✔️ Generative models & reinforcement learning
✔️ Advanced architectures & deployment

#PyTorch #DeepLearning #MachineLearning 🎓🚀

Final Practice Exercises:
1. Implement a GNN for molecular property prediction
2. Train a Neural ODE on irregularly-sampled time series
3. Deploy a model with TorchServe and create a monitoring dashboard
4. Compare SHAP and Integrated Gradients for your CNN model
5. Optimize a transformer model with TensorRT

# Molecular GNN starter
class MolecularGNN(nn.Module):
    def __init__(self, node_features, edge_features, hidden_dim):
        super().__init__()
        self.node_encoder = nn.Linear(node_features, hidden_dim)
        self.edge_encoder = nn.Linear(edge_features, hidden_dim)
        self.conv = tg.nn.MessagePassing(aggr='mean')
        
    def forward(self, data):
        x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr
        x = self.node_encoder(x)
        edge_attr = self.edge_encoder(edge_attr)
        return self.conv(x, edge_index, edge_attr)

🌟 Vision Transformer (ViT) Tutorial – Part 1: From CNNs to Transformers – The Revolution in Computer Vision

Let's start: https://hackmd.io/@husseinsheikho/vit-1

#VisionTransformer #ViT #DeepLearning #ComputerVision #Transformers #AI #MachineLearning #NeuralNetworks #ImageClassification #AttentionIsAllYouNeed

✉️ Our Telegram channels: https://t.iss.one/addlist/0f6vfFbEMdAwODBk

📱 Our WhatsApp channel: https://whatsapp.com/channel/0029VaC7Weq29753hpcggW2A

🌟 Vision Transformer (ViT) Tutorial – Part 2: Implementing ViT from Scratch in PyTorch

Let's start: https://hackmd.io/@husseinsheikho/vit-2

#VisionTransformer #ViTFromScratch #PyTorch #DeepLearning #ComputerVision #Transformers #AI #MachineLearning #CodingTutorial #AttentionIsAllYouNeed

✉️ Our Telegram channels: https://t.iss.one/addlist/0f6vfFbEMdAwODBk

📱 Our WhatsApp channel: https://whatsapp.com/channel/0029VaC7Weq29753hpcggW2A

🌟 Vision Transformer (ViT) Tutorial – Part 3: Pretraining, Transfer Learning & Real-World Applications

Let's start: https://hackmd.io/@husseinsheikho/vit-3

#VisionTransformer #TransferLearning #HuggingFace #ImageNet #FineTuning #AI #DeepLearning #ComputerVision #Transformers #ModelZoo

✉️ Our Telegram channels: https://t.iss.one/addlist/0f6vfFbEMdAwODBk

🔥 Master Vision Transformers with 65+ MCQs! 🔥

Are you preparing for AI interviews or want to test your knowledge in Vision Transformers (ViT)?

🧠 Dive into 65+ curated Multiple Choice Questions covering the fundamentals, architecture, training, and applications of ViT — all with answers!

🌐 Explore Now: https://hackmd.io/@husseinsheikho/vit-mcq

🔹 Table of Contents
Basic Concepts (Q1–Q15)
Architecture & Components (Q16–Q30)
Attention & Transformers (Q31–Q45)
Training & Optimization (Q46–Q55)
Advanced & Real-World Applications (Q56–Q65)
Answer Key & Explanations

#VisionTransformer #ViT #DeepLearning #ComputerVision #Transformers #AI #MachineLearning #MCQ #InterviewPrep

✉️ Our Telegram channels: https://t.iss.one/addlist/0f6vfFbEMdAwODBk

📱 Our WhatsApp channel: https://whatsapp.com/channel/0029VaC7Weq29753hpcggW2A

5 minutes of work - 127,000$ profit!

Opened access to the Jay Welcome Club where the AI bot does all the work itself💻

Usually you pay crazy money to get into this club, but today access is free for everyone!

23,432% on deposit earned by club members in the last 6 months📈

Just follow Jay's trades and earn! 👇

https://t.iss.one/+mONXtEgVxtU5NmZl

Join our WhatsApp channel

There are dedicated resources only for WhatsApp users

https://whatsapp.com/channel/0029VaC7Weq29753hpcggW2A

🚀 Become an Agentic AI Builder — Free 12‑Week Certification by Ready Tensor

Ready Tensor’s Agentic AI Developer Certification is a free, project first 12‑week program designed to help you build and deploy real-world agentic AI systems. You'll complete three portfolio-ready projects using tools like LangChain, LangGraph, and vector databases, while deploying production-ready agents with FastAPI or Streamlit.

The course focuses on developing autonomous AI agents that can plan, reason, use memory, and act safely in complex environments. Certification is earned not by watching lectures, but by building — each project is reviewed against rigorous standards.

You can start anytime, and new cohorts begin monthly. Ideal for developers and engineers ready to go beyond chat prompts and start building true agentic systems.

👉 Apply now: https://www.readytensor.ai/agentic-ai-cert/