miniTorch

miniTorch is a minimal, educational deep learning framework inspired by PyTorch, designed for learning, research, and rapid prototyping. Built on top of NumPy, miniTorch provides a clear and concise implementation of core deep learning concepts, including tensors, autograd, neural network modules, optimizers, and data utilities.

👤 Maintainers

• Shardul Junagade
• Umang Shikarvar
• Soham Gaonkar

✨ Key Features

• NumPy-based Tensor Engine: Custom Tensor class with full support for broadcasting, dtype management, and automatic differentiation (autograd).
• Neural Network Building Blocks:
  • PyTorch-like Module base class for custom layers and models
  • Prebuilt layers: Linear, Sequential, and activation modules (ReLU, LeakyReLU, Sigmoid, Tanh)
  • Parameter management and easy extensibility
• Functional API: Core tensor operations (sum, exp, log, pow, transpose, relu, sigmoid, tanh, leaky_relu, softmax) for building custom computations
• Loss Functions: Ready-to-use MSELoss and BCELoss for regression and classification tasks
• Optimizers: Implementations of SGD and Adam for training neural networks
• Data Utilities: Simple Dataset and DataLoader classes for batching, shuffling, and iterating over data
• Computation Graph Visualization: Visualize your model's computation graph using Graphviz for better debugging and understanding.
• Educational Codebase: Clean, well-documented code ideal for students, educators, and researchers

🚀 Installation

Clone the repository:

git clone https://github.com/Umang-Shikarvar/miniTorch
cd miniTorch
pip install -r requirements.txt

To use minitorch as a library in your own projects, install it in editable mode:

pip install -e ./minitorch

Now you can import minitorch from anywhere in your environment and your changes will be picked up automatically.

Example:

import minitorch
from minitorch.nn.modules import Linear, Sequential
from minitorch.optim import SGD

# Define a simple model
model = Sequential(
    Linear(4, 8),
    minitorch.nn.modules.ReLU(),
    Linear(8, 1)
)

# Create optimizer
optimizer = SGD(model.parameters(), lr=0.01)

🛠️ Documentation & Tips

• For tips and useful commands, see docs/tips_and_commands.md.
• Explore the minitorch/ directory for source code and examples.

🤝 Contributing

We welcome contributions from everyone!

• Maintainers: see docs/CONTRIBUTING (maintainers).md
• Non-maintainers: see docs/CONTRIBUTING.md

For questions, suggestions, or discussions, open an issue or start a discussion on GitHub.

License

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

Happy learning and building with miniTorch! ⭐