This is the source code of the solution of Team ⭐⭐⭐in the prize line⭐⭐⭐ for the competition Google - Isolated Sign Language Recognition (https://www.kaggle.com/competitions/asl-signs/overview).

We trained our models with 4x A100 GPUs with the batch size 128(32 * 4). The input files should be put under the directory of the source codes.

0. use pip install -r requirements.txt to install dependency.
1. Run makedataset.py to prepare the dataset for dataloader.
2. Run sh run.sh to train the model ddp.
3. Use the notebook merge.ipynb for making submissions for the competition.

Our final submission consists of 2 models. One with more parameters with mainly the following settings,

1. embedding size: 480
2. number of head: 16
3. 3 layers of transformer

and another with,

1. embedding size: 240
2. number of head: 16
3. 3 layers of transformer