AdeptML

Auto-differentiable embedding of Physics and Torch Machine Learning (AdePT-ML):

This is a convienience library built on top of PyTorch to enable easy integration and training of hybrid models involving physics and deep learning modules.

Features

1. Allows integration of torch.nn.module with numpy functions and enable training with torch optimizers.
2. Pre-defined Modules and configs for physics and MLP architectures.
3. Integrated training function with tensorboard support.

Installation

Installing with pip

pip install adeptml 

Requirements (Automatically installed with pip):

1. PyTorch (https://pytorch.org/)
2. Joblib (https://joblib.readthedocs.io/en/latest/) (For loading and saving model parameters)
3. Tensorboard

Usage:

The primary building block of this package is the Hybrid Model class. It neatly packages all the member models into one main Torch model and enables running forward inference as well as backpropagation. The class accepts as input an instance of the Hybrid Config class. This config is useful in defining all the constituent modules and their inputs.

As component modules, the Models module provides a straight forward MLP implementation as well as a Physics Module. This module is a torch Autograd wrapper which enables the integration of non-Torch numpy functions into a fully torch model and allows for training with torch optimizers.

API Documentation:

Visit our Read the docs page

Examples:

Refer to the tests file. Additional examples will be added soon.