Sign Language Fingerspelling Recognition (SLFR) Examples

Virtual environment

Miniconda was used to manage the virtual environment, and it is suggested to do so. To get the environment up and running, execute:

conda env create -f environment.yml

Dataset

A copy of the dataset from the American Sign Language Fingerspelling Recognition Kaggle competition is located in /home/temporal2/jsoutelo/datasets/GSLFR. It is suggested to create a symlink to it inside the ./data folder. Run:

ln -s  /home/temporal2/jsoutelo/datasets/GSLFR absolute/path/to/data

Data splits

The data splits are located inside the ./data_gen directory. Just as training (or inference), they have a configuration file associated to each of them, they can be found in ./data_gen/config. To generate a new split run:

python -m data_gen.generate_splits --config ./data_gen/config/config_file.yaml

As a result, a folder containing the splits will be created. It will have the same name as the configuration file used to generate them. I woudl recommend to start by generating the baseline.yaml split.

Raw data

In case that you want to use the raw data from the competition directly, it is provided two scripts inside ./raw_data_scripts folder. They handle the extraction logic of the raw .parquet files. This is not recommended.

Fine-tuning and Inference

To switch between the two modes it is just enough to edit the phase field in the config file to run either train or test. Once configured, simply execute the following:

python main.py --config ./path/to/config/file.yaml

Two examples are provided in the ./config folder for reference.

Notes

Problems with torch_edit_distance

If some kind of problem airses with the package torch_edit_distance please reinstall. A compiled version for the baiona server can be found in /home/temporal2/jsoutelo/builds/pytorch-edit-distance-bai. With a virtual environment activated, run:

cd /home/temporal2/jsoutelo/builds/pytorch-edit-distance-bai
python setup.py install

If the error persists feel free to reach out to me.