Minute Grad

A CPU implementation of an automatic differentiation framework for deep learning.

Run Locally

The automatic differentiation framework only depends on NumPy for computation. For experiments, there are optional dependencies matplotlib, tqdm, and wandb for visualisation, progress meter and tracking results respectively.

The project uses uv for managing dependencies, so you need to install uv first. Then, run uv sync --no-dev to update your local environment. To run the experiments/all_experiments.py script for example, simply run uv run python experiments/all_experiments.py.

Project Structure

.
├── data                             # Provided dataset in NumPy format
│   └── ...
├── experiments                      # Experimental scripts and utilities
│   ├── experiments                  # Directory containing experiment results (metrics & plots)
│   │   └── results
│   │       └──...
│   ├── architecture_experiments.py  # Runs architecture experiments testing different number of layers, neurons, etc.
│   ├── hyperparameter_experiments.py# Hyperparameter tuning experiments
│   ├── ablation_test.py             # Code used for the ablation study
│   └── train_util.py                # Helper functions for training routines
├── layer.py                         # Implementation of network layers
├── main.py                          # An example training script on the provided dataset
├── optimiser.py                     # Definitions of optimizer algorithms (e.g., Adam)
├── tensor.py                        # Definition of tensor operations and structures
├── util.py                          # Miscellaneous helper functions
├── README.md                        # You are here.
├── pyproject.toml                   # Project configuration and dependency management
└── uv.lock                          # Dependency version lock file