PyTorch Deep Learning & Computer Vision Foundations

A sophisticated deep learning and computer vision repository covering fundamental tensor mathematics, custom neural architectures, non-linear classification, convolutional vision pipelines (TinyVGG), dynamic data augmentation, device-agnostic GPU acceleration and modular CLI-driven training engines.

---

📌 Repository Structure

pytorch-deep-learning-foundations/
├── 00_pytorch_fundamentals/
│   ├── 00_pytorch_fundamentals.ipynb
│   └── 00_pytorch_fundamentals_solutions.ipynb
├── 01_pytorch_workflow/
│   ├── 01_pytorch_workflow.ipynb
│   └── 01_pytorch_workflow_solutions.ipynb
├── 02_pytorch_neural_network_classification/
│   ├── 02_pytorch_classification.ipynb
│   └── 02_pytorch_classification_solutions.ipynb
├── 03_pytorch_computer_vision/
│   ├── 03_pytorch_computer_vision.ipynb
│   └── 03_pytorch_computer_vision_solutions.ipynb
├── 04_pytorch_custom_datasets/
│   ├── 04_pytorch_custom_datasets.ipynb
│   └── 04_pytorch_custom_datasets_solutions.ipynb
├── 05_pytorch_going_modular/
│   ├── 05_pytorch_going_modular.ipynb
│   ├── 05_pytorch_going_modular_solutions.ipynb
│   └── going_modular/
│       ├── data_setup.py
│       ├── engine.py
│       ├── model_builder.py
│       ├── train.py
│       └── utils.py
├── .gitignore
├── LICENSE
├── requirements.txt
└── README.md

---

🔬 Module Breakdown

Module	Core Concepts & Systems Implemented	Key APIs / Mathematical Techniques
00_pytorch_fundamentals	Tensor initialization, memory layouts, dimensional manipulation, device-agnostic execution (CPU/CUDA), batch matrix multiplication.	torch.matmul, torch.squeeze, torch.permute, torch.manual_seed
01_pytorch_workflow	End-to-end regression pipeline, nn.Module subclassing, forward computational graphs, parameter updates, serialization.	nn.Linear, nn.L1Loss, torch.optim.SGD, state_dict
02_pytorch_neural_network_classification	Non-linear geometric separation (Circles, Blobs), logit-to-probability mapping, numerical stability under extreme loss bounds.	BCEWithLogitsLoss, CrossEntropyLoss, nn.ReLU, torchmetrics
03_pytorch_computer_vision	Multi-class image classification (FashionMNIST), NCHW tensor conventions, 2D convolutional filter design, max pooling downsampling.	nn.Conv2d, nn.MaxPool2d, torchvision.transforms, TinyVGG
04_pytorch_custom_datasets	Custom disk ingestion pipelines (Dataset/DataLoader), stochastic data augmentations (TrivialAugmentWide), transfer learning adaptation.	ImageFolder, __getitem__ override, torchvision.models
05_pytorch_going_modular	Refactoring notebook code into production-ready, reusable Python engines with parameterizable CLI entry points.	argparse, modular engine abstractions, automated asset caching

---

⚙️ Modular Engine Architecture (going_modular/)

The repository transitions research prototypes into a clean, decoupled production package:

going_modular/
├── data_setup.py      # Ingestion logic, directory parsing, transformations, and DataLoader pipeline generation
├── engine.py          # Device-agnostic train_step, test_step, and multi-epoch loops
├── model_builder.py   # Parameterized CNN architecture (TinyVGG) definition
├── train.py           # Making executable CLI orchestration script tying data loading, model instantiation, training loops with dynamic hyperparameters
└── utils.py           # Reusable utilities for saving model checkpoints and directory setup.

Modular CLI Usage

1. Dataset Ingestion

python going_modular/data_setup.py

2. Model Training with Hyperparameter Overrides

python going_modular/train.py \
    --model_name tinyvgg_food101.pth \
    --num_epochs 20 \
    --batch_size 32 \
    --hidden_units 64 \
    --learning_rate 0.001

3. Single-Sample Inference

python going_modular/predict.py \
    --model_path models/tinyvgg_food101.pth \
    --image data/pizza_steak_sushi/test/pizza/1152100.jpg

---

🛠️ Key Architectural Decisions

• Numerical Stability via Logit-Level Loss Formulation:
  Replaced raw BCELoss(Sigmoid(x)) with BCEWithLogitsLoss to combine the sigmoid layer and cross-entropy step into a single mathematically fused operation, preventing vanishing gradients and NaN instability during backpropagation.

• Device-Agnostic Acceleration:
  Standardized tensor and model transitions to automatically detect and leverage CUDA environments while maintaining seamless fallback to CPU execution:

  device = "cuda" if torch.cuda.is_available() else "cpu"

• Decoupled CLI & Modular Script Architecture: Transitioned exploratory Jupyter prototypes into an isolated, reusable production package (going_modular/). Separated dataset ingestion, model definition, training/evaluation loops, checkpoint persistence, and inference into independent, parameterizable CLI modules driven by argparse.

• Zero-Artifact Commit Policy: All evaluation metrics, decision boundary surfaces, confusion matrices, and loss curves are executed and embedded directly within the .ipynb cells to keep the repository lightweight and eliminate out-of-sync visual file dependencies.

---

🚀 Quickstart & Environment Setup

Prerequisites

• Python 3.10 or higher
• NVIDIA GPU with CUDA support (optional, but recommended)

Installation

# Clone the repository
git clone https://github.com/Subhrajyoti8520/pytorch-deep-learning-foundations.git
cd pytorch-deep-learning-foundations

# Create and activate a virtual environment
python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Launch Jupyter:

jupyter lab
# or
jupyter notebook

---

📜 License

This project is licensed under the MIT License — see the LICENSE file for details.