fsl-app-2526/fsl-express

FSL Interpreter

Implementing a transformer-based word-level FSL interpreter for emotion-aware sentence translation.

Code Structure

• generate sentence/pose landmarks = each uses MediaPipe holistic to generate landmark data and stores them into two HDF5 files.
• generate json data = splits the dataset into the 80-20 w/ 5-folds for training
• app = contains gradio front-end
• annotater = technically not needed but added for posterity, was used to split

Data Files

• results = container including results from study that was run from Jan 2026 -> May 2026

• Extract sentence_data and sliced_dataset tar into "fsl-data" folder in root dir. Move both sentence_landmarks.h5 and multitask_mediapipe.h5 files into "fsl-data" as well.