pytorch-plug-n-play

Pytorch lightning code that allows you to just set up your dataset and model and you are good to go. I set this up so that new projects can get running and off the ground quickly and also because I hate re-writing boilerplate code.

Quick tour

The project is configured from .yaml files found in the config/ directory. I recommend you create a new config for each experiment that you run so that you can track past configurations (even though these are recorded on wandb anyways)

The code runs from run.py which can be ran as python -m plug.run "YOUR_CONFIG.yaml" after which the checkpoints will be created in the checkpoints/ directory as written in checkpoin_dir in the yaml config file.

sweep.py is a file to use for hyperperameter tuning.

Usage

1. Create conda environment using conda create -n PROJECT_NAME and run pip install -e . to install all dependencies in setup.py
2. Create Pytorch model that inherits from the BaseModel class in plug.models and import into run.py
3. Vice versa for datasets and dataloaders but there is no class to inheret from. (I recommend you use pytorch lightnings dataloader superclass)

Small notes

• The default dtype for tensors is set as float32 on medium accuracy in the run.py file which you will have to change manually within the file.