🧠 PyTorch Deep Learning Journey

My personal journey learning PyTorch — from basic tensor operations to designing, training, and optimizing neural networks. Every notebook here is a step in my hands-on deep learning education.

---

🎯 About This Repository

This is a structured, practical archive of my PyTorch learning path. It includes conceptual breakdowns, working notebooks, mini-projects, and optimization experiments — all built to reinforce deep learning foundations and explore advanced applications.

Why I built this: To reinforce my learning with real code, and to share tested, beginner-friendly examples that actually work.

---

📚 Learning Roadmap

🔰 Foundation Phase

Notebook	Key Learnings
00_Introduction to PyTorch.pdf	Overview of PyTorch and its ecosystem.
01_Tensors_in_pytorch.ipynb	Tensor creation, indexing, broadcasting, and operations.
02_pytorch_autograd.ipynb	Gradient computation and backpropagation via autograd.
03_pytorch_training_pipeline.ipynb	Manual training loops, loss calculation, and optimization basics.

🏗️ Building Better Models

Notebook	Focus Area
04_pytorch_nn_module.ipynb	Writing modular, reusable neural networks using nn.Module.
05_dataset_and_dataloader.ipynb	Clean data handling with Dataset and DataLoader.
06_ANN_using_pytorch.ipynb	My first complete multi-layer perceptron (MLP) network.

⚡ Performance & Optimization

Notebook	What I Explored
07_neural_network_training_on_GPU.ipynb	CUDA-powered training on GPU — major performance gains.
08_Optimize_neural_network.ipynb	Manual optimization techniques — adjusting learning rates, etc.
09_Optimize_neural_network_using_optuna.ipynb	Automated tuning with Optuna — model selection made efficient.

🧠 Advanced Architectures

Notebook	Project
10_a_CNN_on_fashion_mnist.ipynb	Basic convolutional neural network for image classification.
11_CNN_on_fashion_mnist.ipynb	Improved CNN with batch norm, dropout, and deeper layers.
12_RNN_using_pytorch.ipynb	Intro to RNNs for handling sequential data.
13_next_word_predictor.ipynb	RNN-based language model for next-word prediction.

---

📊 Datasets

File	Usage	Notes
06_fmnist_small.csv	MLP training	Lightweight dataset for quick tests
07_fashion-mnist_test.csv	CNN testing	Test set for benchmarking
07_fashion-mnist_train.zip	CNN training	Full training data
12_100_Unique_QA_Dataset.csv	NLP with RNNs	Custom Q&A dataset

---

⚙️ Getting Started

🧪 Setup Instructions

# 1. Clone the repo
git clone https://github.com/Sourabh-Kumar04/Pytorch.git
cd Pytorch

# 2. Create environment
conda create -n pytorch-learning python=3.9
conda activate pytorch-learning

# 3. Install dependencies
pip install torch torchvision matplotlib numpy optuna pandas jupyter

# 4. Run the notebooks
jupyter notebook

---

💡 Key Learnings & Insights

💥 What Stood Out

• Simple models first: My early ANN helped me understand more than any advanced CNN tutorial.
• GPU = Game changer: 10× training speed-up with CUDA.
• Pipeline precision: Clean data handling saved me hours of pain.
• Optuna: Discovered hyperparameters I wouldn't have considered.

✅ What Helped Me Most

• Following a week-by-week structure
• Running experiments on the same dataset to compare architectures
• Keeping visual checkpoints to understand overfitting/underfitting

🚧 Pitfalls I Hit

• Forgetting .to(device) calls → silent bugs on GPU
• Skipping normalization → exploding losses
• Wrong loss functions → poor convergence
• Overfitting small datasets → false sense of success

---

📈 Results Snapshot

Task	Initial Accuracy	Final Accuracy	Improvements
CNN (Fashion MNIST)	78%	91%	Model tuning & augmentation
Advanced CNN	91%	94%	Dropout, LR scheduler
Word Prediction (RNN)	65%	82%	Larger hidden layer, better preprocessing

---

🧰 Tools & Libraries

torch>=2.0.0          # Core deep learning framework
torchvision>=0.15.0   # Datasets, transforms, pretrained models
numpy>=1.21.0         # Numerical operations
pandas>=1.3.0         # Data wrangling
matplotlib>=3.5.0     # Visualizations
seaborn>=0.11.0       # Advanced plots
optuna>=3.0.0         # Hyperparameter optimization
jupyter>=1.0.0        # Notebook execution
tqdm>=4.62.0          # Loop progress bars

---

📅 Learning Schedule (What Worked for Me)

Week	Topics
Week 1	Tensors + Autograd (01, 02)
Week 2	Training loops & modules (03, 04)
Week 3	Data & MLP models (05, 06)
Week 4	GPU & Optimization (07, 08, 09)
Week 5	CNNs for CV (10, 11)
Week 6	RNNs for NLP (12, 13)

✅ Tip: Don’t rush the early modules — they’re the real foundation!

---

❓ Common Questions

Do I need a GPU?

Not at the start. You can run all notebooks up to `06_ANN` comfortably on CPU. For CNN and RNN, a GPU is highly recommended for speed.

Where should I start?

Begin with `00_Introduction to PyTorch.pdf`. Even if you know NumPy, PyTorch has its own quirks.

Can I use this for my own learning/projects?

Absolutely! It's MIT licensed — fork it, adapt it, build on it.

---

📚 Resources That Helped Me

Official Docs & Courses

• 📘 PyTorch Tutorials
• 📗 Deep Learning with PyTorch Book
• 🎓 Fast.ai Course

When Stuck

• 🧠 PyTorch Forums
• 🧩 Stack Overflow
• 📊 Papers with Code

---

👨‍💻 About Me

Sourabh Kumar 🎓 AI Programming with Python - Udacity Graduate

🎖️ AWS AI ML Scholar'24 | AI | ML | GenAI Explorer*

🌐 GitHub | LinkedIn

"Always learning, always building — one notebook at a time."

---

📄 License

This project is licensed under the MIT License.

---

⭐️ Support the Project

If you found this useful:

• Give a ⭐️ to the repository
• Fork it and make it your own
• Share feedback or suggestions
• Connect with me for collabs or questions

---

Happy Learning! 🚀

"The best way to learn deep learning is by building with it."