toobrainless / as

ASR

This is a repository for ASR homework with implementation of Conformer model and some experiments with it. Also that repository can be used as a template for other projects.

Getting Started

These instructions will help you to run a project on your machine.

Prerequisites

Install poetry following the instructions in the official repository.

Installation

Clone the repository into your local folder:

cd <path/to/local/folder>
git clone https://github.com/toobrainless/dla_template -b asr

Setup requirements with poetry

cd <path/to/cloned/ASR/project>
poetry install

Now you can use poetry run to use environment. For example, poetry run python3 train.py, for more usability read poetry documentation.

Setup heavy stuff

poetry run python3 setup_lm_model.py
poetry run python3 setup_conformer.py

Training

To train model run with default config:

poetry run python3 train.py

To resume training from checkpoint

poetry run python3 train.py resume=path\to\saved\checkpoint.pth

To change config parameters use the following approach:

poetry run python3 train.py param1=value1 param2=value2

For example, to change batch size and number of epochs:

poetry run python3 train.py data.train.batch_size=32 trainer.epochs=10

The train config is stored in configs/train.yaml, it contains path to the datasets, model, optimizer, scheduler, trainer and other parameters. Also you can change it manually or create your own config file.

Testing

To evaluate model run the following script:

poetry run python3 test.py

The test config is stored in configs/test.yaml, it contains path to the checkpoint, datasets and metrics for evaluation. To change config parameters use the same approach as for training.