lnl-examples

This repository contains a sequence of examples for using PyTorch Lightning and Lhotse, two Python libraries that help with a lot of the necessary plumbing that can otherwise be tricky/time-consuming to take care of yourself (e.g. automatic mixed precision, dynamic batching, etc.).

The examples are split between Google Colab notebooks in notebooks (e.g. notebooks/01_mwe.ipynb) and Python script in the root directory (e.g. train_mwe.py). The notebooks contain a lot of prose aimed at newcomers and builds up a component-by-component introduction towards a given Python script (e.g. everything you need to understand train_mwe.py is found in 01_mwe.ipynb).

Usage

Environment

We'll be using the latest stable versions of various packages as of early June 2023:

# For minimal working example in train_mwe.py
torch==2.0.1
torchaudio==2.02
lightning==2.0.2
lhotse==1.14.0

# For extras in train_mwe-extras.py and trainlnl.py
jiwer==3.0.2
pandas==2.0.2
hydra-core==1.3.2
wandb==0.15.4
torchinfo==1.8.0

Conda

You can create and activate a conda environment at in the env directory from the env.yaml file using:

conda env create --prefix ./env --file env.yaml
conda activate ./env

Docker

If you have docker and docker-compose you can also launch a container using:

docker compose run lnl-examples

Run examples

Minimal working examples

Once you've set up this environment (and optionally read the corresponding notebook), you can run a given example as:

python train_mwe.py

# or extras

python train_mwe-extras.py

LnL example

# To run without Weights & Biases configured
wandb disabled
python train_lnl.py \
    lnl.wnb_project=null \
    lnl.wnb_run=null

# To run with Weights & Biases configured
wandb enabled
python train_lnl.py \
    lnl.wnb_project=my_project \
    lnl.wnb_run='test run train_lnl.py'

# Use Hydra CLI overrides to change/add configs (see defaults in configs folder):
python train_lnl.py \
    lnl.wnb_project=my_project \
    lnl.wnb_run='test run train_lnl.py' \
    trainer.precision='16-mixed' \
    +trainer.gradient_clip_val=0.5

fauxneticien / lnl-examples