Ready for the grAIt descent?
GrAIdient is a framework that allows deep learning models to be developed using the internal GPU of a Mac, unlocking researchers to more easily train and run AI models on their own computers.
GrAIdient exposes the graph of layers, providing a unique way to design deep learning models for greater understanding, control and reproducibility.
Though deeply grounded to the data driven pipeline, the goal is to challenge the very understanding of deep learning models and inject human intelligence where relevant; to transition from black box models to white box models, and all the gradients in between.
Check out this toy VGG example and its documentation to get started with GrAIdient today!
- direct access to the graph of layers and to the backward pass
- run natively on Mac GPU (Intel GPU, eGPU, Apple Silicon)
- written in Swift: a compiled language with strong typing
- gradient checking to validate backward operations
- PyTorch interoperability
- gradients per batch & per sample (needed for differential privacy)
- debug at the neuron level
Add the following dependency to your Package.swift manifest:
.package(url: "https://github.com/owkin/GrAIdient.git", .branch("main")),The documentation is divided into several sections:
Read below to learn how to take part in improving GrAIdient.
All notable changes to GrAIdient will be documented in the changelog.
Read our contributing guide to learn about our development process and how to build and test your changes to GrAIdient.
GrAIdient has adopted a Code of Conduct that we expect project participants to adhere to. Please read the full text so that you can understand what actions will and will not be tolerated.
GrAIdient, GrAITestsUtils and GrAITests are MIT licenced.
GrAIExamples and GrAITorchTests both depend on PythonKit and are Apache 2.0 licensed.