Development of a Generalizable Exercise Monitoring System using Kinect Data for Binary Classification of Performance

Instructions for running:

• Note: If you don't want to generate the dataset files again, proceed to step 4

1. Clone the repo
2. Run all cells in the UI-PRMD20pre-processing.ipynb notebook to generate the preprocessed .npy files of 3D skeletal data
3. Navigate to the datasets folder and run 2class-all/generate_splits.ipynb and 80-20-split/generate_splits/ipynb. This will generate the necessary PKL files for hyperparameter tuning and LOSO validation.
4. Open Anaconda Prompt and run the following commands to set up an environment for this project:

    conda create --name mmactionenv python=3.8 -y
    conda activate openmmlab
    conda install pytorch torchvision -c pytorch
    pip install -U openmim
    mim install mmengine
    mim install mmcv==2.1.0
    mim install mmdet==3.2.0
    mim install mmpose
    pip install mmaction2
   

5. Navigate to the mmaction2 directory and replace the tools/train.py file with the train.py file provided in the Testing_Code directory. This file has modified the code to run hyperparameter tuning and LOSO validation.
6. If you are doing hyperparameter tuning please uncomment the call for hyperparam_grid_search(cfg) and comment out the code below it. If you are doing LOSO validation please leave the file as is.
7. Whether you are doing hyperparameter tuning or LOSO validation, the command to train and validate the model is the same. Navigate to the Testing_Code directory and run:

   python ../mmaction2/tools/train.py stgcn_custom_exercise.py --seed 0 --deterministic
   

8. You will find the results from the process under its respective folders in the Testing_Code directory. To examine the results per split for LOSO, please navigate to Testing_Code\loso_2class_all\loso_split_s10.pkl_lr_0.001_bs_16_repeat_3\loso_results.txt where the metrics will be listed for each split.