A Machine Learning framework from scratch in Mojo 🔥
| Task | Dataset | Forward | Backward | Training |
|---|---|---|---|---|
| REGRESSION | ✅ | ✅ | ✅ | ✅ |
| MNIST | ✅ | ✅ | ✅ | ✅ |
Regression example:
Running a regression example
mojo run -I . examples/housing.mojo
Compare to a alternative PyTorch implementation:
Install the requirements in python-requirements.txt
python examples/housing.py
❌: Not implemented
✅: Working (but might require changes because of not implemented dependencies)
WIP: Work in progress
| Task | Status |
|---|---|
| NODE | ✅ |
| GRAPH | ✅ |
| Task | Status |
|---|---|
| ADD | ✅ |
| SUB | ✅ |
| MUL | ✅ |
| DIV | ✅ |
| DOT | ✅ |
| EXP | ✅ |
| LOG | ✅ |
| POW | ✅ |
| SUM | ✅ |
| MAX | ✅ |
| TRANSPOSE | ✅ |
| FLATTEN | ✅ |
| RESHAPE | ✅ |
| CONV2D | ✅ |
| CONV3D | ❌ |
| MAXPOOL2D | ✅ |
| MAXPOOL3D | ❌ |
| Task | Status |
|---|---|
| MSE | ✅ |
| CE | ✅ |
| Task | Status |
|---|---|
| RELU | ✅ |
| SIGMOID | ✅ |
| TANH | ✅ |
| SOFTMAX | ✅ |
| LOGSOFTMAX | ✅ |
| LEAKYRELU | ❌ |
| Task | Status |
|---|---|
| ADAM | ✅ |
| Task | Status |
|---|---|
| SEQUENTIAL | ❌ |
| LINEAR | ✅ |
| DROPOUT | ❌ |
| CONV2D | ✅ |
| CONV3D | ❌ |
| MAXPOOL2D | ✅ |
| MAXPOOL3D | ❌ |
| Task | Status |
|---|---|
| Model abstractions (eval/save/load/...) | ❌ |
| Datasets (MNIST/Boston Housing) | ✅ |
| Dataloader | ✅ |
| Tensorutils | ✅ |
| Checkpoints | ❌ |
| Task | Original | Included Kaggle CSV |
|---|---|---|
| Bosten Housing Dataset | names, data | kaggle |
| MNIST DataSet | training images, training labels, test images, test labels |
(subset of) kaggle |
Policy WIP.
This project is in active development, and we welcome contributions:
- Fork the repository on GitHub.
- Create a new branch for your feature or bug fix.
- Commit your changes with clear, descriptive commit messages.
- Push your branch to GitHub and submit a pull request.
We appreciate any and all contributions, whether they're for bug fixes, new features, or documentation improvements.
This project is licensed under the Apache License 2.0 - see the LICENSE file for details.
A huge thanks to pranftw for the inspiration and support! This project is inspired by neograd.