MyTorch: A Custom Deep Learning Library

MyTorch is a lightweight neural‑network framework built from scratch in Python. It provides the core building blocks for constructing, training, and evaluating deep‑learning models without relying on PyTorch internals.

⚙️ Features

1. Tensor

• Tensor: n‑dimensional array with support for automatic gradient tracking, in-place operations, broadcasting, and backpropagation.

2. Activation Functions

• ReLU (relu.py)
• Leaky ReLU (leaky_relu.py)
• Sigmoid (sigmoid.py)
• Tanh (tanh.py)
• Softmax (softmax.py)
• Step Function (step.py)

3. Layers & Modules

• Linear (fully connected) (linear.py)
• Conv2D (conv2d.py)
• Pooling
  • MaxPool2D (max_pool2d.py)
  • AvgPool2D (avg_pool2d.py)
• Generic Layer Interface (layer.py)
• High‑level Model API (model.py): assemble and manage layers, forward pass orchestration.

4. Loss Functions

• Mean Squared Error (MSE) (mse.py)
• Binary Cross‑Entropy (BCE) (bce.py)
• Categorical Cross‑Entropy (CE) (ce.py)

5. Optimizers

• SGD (sgd.py)
• Momentum (momentum.py)
• RMSProp (rmsprop.py)
• Adam (adam.py)

All optimizers follow a uniform Optimizer interface defined in optimizer.py.

6. Data Handling

• Custom DataLoader (my_data_loader.py): mini‑batch iteration, shuffling, parallel loading support
• Utility Data Loader (util/data_loader.py)
• Flatten helper for reshaping tensors (util/flatten.py)
• Weight & Bias Initializers (util/initializer.py)

📓 Notebooks Overview

Five Jupyter notebooks demonstrate both native PyTorch models and their MyTorch counterparts across two datasets:

Phase 1: Breast Cancer Wisconsin (Diagnostic)

• Task 1 (PyTorch MLP)
  Build and train a multilayer perceptron in PyTorch.

• Task 2 (MyTorch MLP)
  Re‑implement the same MLP using MyTorch’s Tensor, Module and SGD classes.
  Demonstrates end‑to‑end custom backpropagation.

Phase 2: FER‑2013 Facial Expression Recognition

• Task 3 (PyTorch MLP)
  Benchmark MLP on FER‑2013 images (flattened input).

• Task 4 (PyTorch CNN)
  Build and train a convolutional network in PyTorch for emotion classification.

• Task 5 (MyTorch CNN)
  Re‑implement the CNN architecture in MyTorch, covering Conv2D, pooling and softmax layers.
  Illustrates how MyTorch handles multi‑dimensional gradient flow.