Repository for my DL4CV project at THM. Project only tested with Python 3.10.12
You can use this link to download my pretrained checkpoint.
Create new environment and install dependencies:
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
You must clone the Merkel Podcast Corpus repository and execute the "download_video.py" script. This may take a while.
After you're done you can execute the following to preprocess the data for training.
python ./make_dataset.py --workers N /path/to/corpus
where N is the amount of worker threads you wish to use. Preprocessed files will be saved by default into the data
directory.
Following example command trains with Wandb logging enabled and also preloads the entire dataset into memory.
python ./train.py --enable-logging --preload
This command runs an evaluation on a random subset of the dataset (controlled by --size
) and calculates mean STOI and ESTOI metrics.
python ./eval.py --checkpoint /path/to/checkpoint.pth --size 0.1
Starts the Gradio demo.
python ./demo.py --checkpoint /path/to/checkpoint.pth