Deep Learning (20/21) WMAI017-05.2020-2021.2A

Uses data describing the ASL alphabet and tries to classify the images correctly using an adapted custom Neural Network, built with TensorFlow/Keras. Runs the model fitting multiple times for various optimizers, such that we can compare various optimizers against each other.

Demonstrates the sign language detection model using TensorFlow.js:

1. Follow the setup from the official Kaggle API repo. Make sure ~/.kaggle/kaggle.json exists.

2. Create a virtual environment and install the packages:

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Will install all required packages.

3. Download the data from Kaggle

sh util/data_download.sh <directory_to_store_data>

4. (optional) Combine datasets

python src/combine_datasets.py <directory_to_store_data>

5. Run training Test using just 32 samples:

python src/train_model_test.py

Do a full training cycle:

python src/train_model.py -d <directory_to_store_data>/<dataset_to_use>

Directory <dataset_to_use> must have subdirectories containing the names of the designated classes.

Follow the instructions above, will work for Peregrine just as well. Submit a job using:

sbatch util/peregrine.sh

Download the results using:

rsync -aP $PEREGRINE_USERNAME@peregrine.hpc.rug.nl:~/deep-learning/logs ./
rsync -aP $PEREGRINE_USERNAME@peregrine.hpc.rug.nl:~/deep-learning/models ./

Run the demo locally by running:

cd demo
yarn
yarn start

A browser tab should automatically be opened with the website ✨

By Jeroen Overschie and Loran Knol.