!!! note For more information, visit the documentation.
Dependencies and virtual environments are managed using poetry.
- python (>=3.9 <3.12) and development headers
- libsndfile and development headers
- libsox and development headers
- a blas implementation and development headers
Install dependencies:
sudo apt update
sudo apt -y install python3-dev libblas-dev liblapack-dev libsndfile1-dev libsox-dev git-lfs
Install dependencies:
sudo yum install portaudio-devel libsndfile1-devel libsox-devel -y
Install a python 3.9 or python 3.10 version using pyenv.
curl -sSL https://install.python-poetry.org/ | python3
For alternative installation methods see: https://python-poetry.org/docs/#installation
Caution: this usually installs poetry to ~/.local/bin if this folder is not in your path you might need to run poetry as:
~/.local/bin/poetry
In the root directory of the project run:
git submodule update --init --recursive
poetry install -E vision
This creates a virtual environment under ~/.cache/pypoetry/virtualenvs.
The environment can be activated using:
poetry shell
We support installation of optional dependencies using poetry's -E commandline flag
We currently have the following optional dependencies:
Vision models require additional dependencies, these can be installed using:
poetry install -E vision
These dependencies include kornia and albumentations for image augmentations and timm (torch image models)
for baseline neural network models.
poetry install -E onnxrt-backend
Backend support for running models on onnx-runtime.
poetry install -E onnx-tf
Onnx based conversion of trained models to tensorflow/tensorflow-lite for external inference backends.
poetry install -E dvc
This installs dvc based model, data and experiment management. DVC support is highly experimental and subject to change.
1.) venv location
poetry installs the dependencies to a virtual environment in ~/.cache/pypoetry/virtualenvs
You can change the location of this directory using:
poetry config virtualenvs.path <desired virtual environment path>
Or move it to a subdirectory of the project directory using:
poetry config virtualenvs.in-project true
Datasets are downloaded automatically to the datasets data folder by default this is a subfolder of the dataset's data folder.
For the VAD Dataset the following Flag is needed to Download/Override the Dataset
dataset.override=True
Training is invoked by
hannah-train
If available the first GPU of the system will be used by default. Selecting another GPU is possible using the argument trainer.devices=[number]
e.g. for GPU 2 use:
hannah-train trainer.devices=[2]
Trained models are saved under trained_models/<experiment_id>/<model_name>.
Training of VAD is invoked by
hannah-train dataset=vad model.n_labels=2
Training of VAD_Extended is invoked by
hannah-train dataset=vad_extended model.n_labels=2
Selection of other Voice Dataset use dataset.variants="[UWNU, de, en, it, fr, es]"
hannah-train dataset=vad model.n_labels=2 dataset.variants="[UWNU, de, en, it, fr, es]"
Selection of other Noise Datasets use dataset.noise_dataset="[TUT]"
hannah-train dataset=vad model.n_labels=2 dataset.noise_dataset="[TUT]"
Selection of dataset Split use dataset.data_split="vad_balanced"
hannah-train dataset=vad model.n_labels=2 dataset.data_split="vad_balanced"
Create Vad_small Dataset
hannah-train dataset=vad model.n_labels=2 dataset.variants="[UWNU]" dataset.noise_dataset="[TUT]" dataset.data_split="vad_balanced"
Create VAD_big Dataset
hannah-train dataset=vad model.n_labels=2 dataset.variants="[UWNU, en, de, fr, es, it]" dataset.noise_dataset="[TUT, FSD50K]" dataset.data_split="vad_balanced"
Training of PAMAP2 human activity detection dataset is invoked by:
hannah-train -cn config_activity
Training of emergency siren detection dataset is invoked by:
hannah-train -cn config_siren_detection
To launch multiple optimizations in parallel you can use a hydra launcher.
Joblib launcher is installed by default:
hannah-train --multirun hydra/sweeper=optuna hydra/launcher=joblib optimizer.lr='interval(0.0001,0.1)' optimizer.weight_decay='interval(0, 0.1)' hydra.launcher.n_jobs=5
Launches optimizer hyperparameter optimization with 5 parallel jobs.
To stop training early when a validation metric does not improve, you can use lightning's early stopping callback:
hannah-train early_stopping=default
All experiments are logged to tensorboard: To visualize the results use:
tensorboard --logdir trained_models
or a subdirectory of trained models if only one experiment or model is of interest.
This project uses precommit hooks for auto formatting and static code analysis.
To enable precommit hooks run the following command in a poetry shell.
pre-commit install
Try to follow pep8 naming conventions and the rest of pep8 to the best of your abilities.
This project automatically mirrors its main branch and all branches prefixed with pub/ to its public github repository.
These branches are configured as protected branches by default.