Nimrod

Description

This is a repo with minimal tooling, modules, models and recipes to get easily get started with deep learning training and experimentation with an emphasis on speech, audio and language modeling.

Install

Requires Python 3.10 to 3.12 (PyTorch 2.9).

Install using Pip

pip install slg-nimrod

The core install covers the models, image and text datasets and training. Optional stacks are extras: audio, diffusion, image, tracking. nimrod.models.diffusion needs the diffusion extra, and nimrod.models.ngram needs kenlm, installed separately (https://github.com/kpu/kenlm). Install all extras with:

pip install "slg-nimrod[all]"

Usage

Check recipes in recipes/ folder. E.g. for a simple digit recognizer on MNIST:

git clone https://github.com/slegroux/nimrod.git
python train.py experiment=mnist_mlp data.num_workers=8 trainer.max_epochs=20 
head config/train.yaml

All the parameters of the experiment are editable and read from a .yaml file which details:

• data and logging directory paths
• data module with data source path and batching parameters
• model architecture
• trainer with hardware acceleration and number of epochs
• callbacks for early stopping and automatic logging to Wandb

Docker

You might want to use docker containers for reproductible development environment or run your project in the cloud

make docker-build
docker pull slegroux/nimrod
docker run -it --rm -p 8888:8888 slegroux/nimrod /bin/bash

You can also use docker-compose to define services and volumes

cd .devcontainer
docker-compose up
docker-compose down

Develop

pip install -e .

Hyperparameter tuning

to compare training results on different model parameters:

python train.py --multirun experiment=mnist_mlp model.nnet.n_h=16,64,256 logger=tensorboard trainer.max_epochs=5

Server

st webapp

Run a simple digit recognizer webapp with GUI

cd server
./run_st_app.sh

Authors

2023 Sylvain Le Groux [email protected]