This repo contains examples of using common code with both the Pytorch and Tensorflow ML/neural network frameworks.
Originally, this project was intended as a way to learn Tensorflow's eager execution mode by porting some typical Pytorch code. However, evidenced by the fact that it works with both frameworks, I have adopted this project structure as a good starting structure for new modelling projects.
Given additional time, I plan to implement more framework-agnostic models, optimizers, and losses.
The project contains the necessary pieces to spin up a docker container that doubles as a Jupyter server or an API, depending on the intended use. A number of docker shortcuts have been added to the Makefile in order to speed things up. However, if make is not installed, you can inspect the Makefile to get the relevant commands and run them manually.
Run make build to build the docker image. Then run make jupyter to start a docker container running jupyter lab. Navigate to localhost:38888 on the host machine to interact with the container.
main.py contains a example implementation of framework-agnostic code. The script generates some random data then creates a data loader, model, optimizer and loss function. The model is then fit to the generated data. The script accepts the following optional command-line arguments:
--framework: pytorch or tensorflow (default: pytorch)--samples: number of samples in generated random data (default: 10000)--features: number of input features (default: 128)--responses: number of output responses (default: 8)--batch_size: batch size (default: 128)--epochs: number of training epochs (default: 10)--classify: flag to switch between regression and classification--architecture: pg for policy gradient RL model, mlp for simple multi-layer perceptron (default: pg)
Examples:
To run RL classification model with pytorch:
python main.py --framework pytorch --architecture pg --classify.
To run MLP regression model with tensorflow:
python main.py --framework tensorflow --architecture mlp.
Any combination of the above parameters should work also, and all are optional.