PyTorch, more like MyTorch Check it out

A PyTorch-like deep learning framework implementation for educational purposes. This project implements a complete neural network library with automatic differentiation, common layer types, optimizers, and example implementations.

• Tensor operations with automatic differentiation
• Neural network modules (Linear, ReLU, etc.)
• Optimizers (SGD, Adam)
• MNIST classification example with 97.16% accuracy
• Comprehensive evaluation metrics

1. Clone the repository:

2. Install uv (if not already installed):

3. Create and activate a virtual environment with dependencies:

4. Download MNIST Dataset: The MNIST dataset files are not included in the repository due to their size. You'll need to:

a. Download the CSV files from Kaggle: MNIST CSV Format b. Place the files in the archive directory:

• archive/mnist_train.csv (for training data)
• archive/mnist_test.csv (for test data)

1. Tensor Operations (mytorch/__init__.py)

   • Basic arithmetic operations (+, -, *, /)
   • Matrix operations (dot product, transpose)
   • Automatic differentiation with backward pass
   • Gradient computation and accumulation

2. Neural Network Modules (mytorch/nn.py)

   • Linear layer with weights and biases
   • Activation functions (ReLU)
   • Loss functions (Cross Entropy)
   • Forward and backward propagation

3. Optimizers

   • SGD (Stochastic Gradient Descent)
   • Adam optimizer implementation

The framework follows a modular design similar to PyTorch:

Run the training script:

The script will:

1. Load and preprocess MNIST data
2. Create a neural network model
3. Train for specified epochs
4. Save model weights

Run the evaluation script:

• Accuracy: 97.16%
• Macro Precision: 97.23%
• Macro Recall: 97.14%
• Macro F1-score: 97.15%

1. Strong Overall Performance

   • 97.16% accuracy on test set
   • Consistent performance across classes
   • High precision and recall for most digits

2. Per-class Analysis

   • Best performance on digit 1 (F1: 98.69%)
   • Most challenging: digit 3 (F1: 94.55%)
   • Very high precision across all classes (>90%)

3. Error Patterns

   • Most confusion between visually similar digits (2↔3, 4↔9)
   • Minimal confusion between dissimilar digits
   • Balanced error distribution

Feel free to open issues or submit pull requests for improvements or bug fixes. Areas for potential enhancement:

1. Additional layer types (Conv2D, MaxPool, etc.)
2. More optimization algorithms
3. Data augmentation techniques
4. Support for other datasets
5. Performance optimizations

MIT License - feel free to use this code for educational purposes.