misakidesu/kiduki-care-ai-gru-multitask

Kiduki Care AI GRU Multitask

This repository contains GRU-based multitask classifiers trained from first-person caregiving video clips.

Included checkpoints

• best_model_all_data.pt
  • Trained on all data (train_on_all_data=true)
  • No held-out validation/test in this setting
• best_model_baseline.pt
  • Train/val/test split baseline
  • Test metrics are available in metrics_baseline.json

Tasks

The model predicts four labels per clip:

1. Task label (task)
2. Stage label (stage)
3. Transfer type (transfer_type)
4. Transfer route (transfer_route)

Files

• best_model_all_data.pt: all-data trained checkpoint
• best_model_baseline.pt: baseline checkpoint
• label_maps_all_data.json, label_maps_baseline.json: label index maps
• metrics_all_data.json, metrics_baseline.json: training/evaluation metrics
• train_gru_baseline.py: training script
• predict_video.py: video inference script

Notes

• This is an application-oriented project for caregiving support workflows.
• The baseline checkpoint includes held-out test results.
• The all-data checkpoint may have higher in-sample performance but has no held-out test in current artifacts.