Vanilla Pytorch + Hydra boilerplate

Yet another Pytorch boilerplate ...... but targeting at research users who would love to

• 1. write customized steps
• 2. conduct extensive ablation studies

Key features

• Vanilla Pytorch with a trainer.

  • The coding style follows HuggingFace-Transformers-trainer.
  • The trainer consists of main functions like train, training_step, evaluation_step, etc.

• Hydra supports multi-run, a command line configuration, and verbose logging of hyper-parameters for ablation studies.

• MISC

  • Custom dataset.
  • Currently, it is a single-GPU implementation.
  • Coding style: one lesson I learned from my mentors is to avoid global variables; instead, pass the hyper-parameters down function by function, which is less prone to unnoticeable bugs.

Sample usage

• python main.py
• python main.py --multirun model.arch=RNN,CNN (reference: conf/config.yaml)

This demo is built for an audio classification task:

• The data/ folder comes with 6 short audio clips of piano or podcast. The exemplary datasets classes load the raw WAV files and segment them into short training and testing samples.
• It runs a vanilla RNN or CNN model that binary-classifies the two audio types.

Future work

• Support multi-GPU training using DDP.

License

Code released under WTFPL