A bitwise reproducible deep learning framework
4 months ago
- #deep-learning
- #PyTorch
- #reproducibility
- RepDL is a library ensuring bitwise identical outcomes in deep learning across different hardware platforms.
- Before setting up RepDL, PyTorch and the corresponding CUDA version must be installed.
- Installation involves cloning the GitHub repository and using pip to install the package.
- RepDL enables reproducible inference and training, with examples provided for both.
- Many PyTorch operations are non-reproducible by default, but RepDL provides reproducible alternatives.
- RepDL defines reproducible operations in `repdl.ops` and implements them in `repdl.backend`.
- Contributions to RepDL require adherence to guidelines for reproducible operations and a Contributor License Agreement (CLA).
- The project follows Microsoft's Open Source Code of Conduct and trademark guidelines.