This repository contains the code accompanying this video. The video also refers to a ROS package, which can be found here.
See the original source repository here
The original source repository contains models trained on the Charades dataset (see original README). This repository contains the code to fine-tune the RGB Charades model on a new dataset, namely, the training datasets from the HEART-MET Activity Recognition Challenge. A separate model will have to be trained for dataset A and dataset B.
To train this model, update the path to the dataset, labels, and the logs directory in run.sh and run the script:
./run.sh
To get the results on the validation/test set, update the path to the dataset and trained checkpoint in test.sh and run the script. This will generate a file submission.json. The submission.json file for each dataset will have to be merged based on the format specified on Codalab.
./test.sh
The code was tested with the following versions:
torch==1.10.2
pytorch-lightning==1.6.1
torchvision==0.11.3
av=10.0.0