FeatureKD: "Feature Based Knowledge Distillation for Fall Detection"

Implementation of FeatureKD proposed in paper titled "SmartFallMM: A Multimodal Dataset For Enhanced Fall Detection"

Getting started

• Create an pip environment and use the requirements.txt to install all the neccasary files.

pip install -r requirements.txt

This requirements file doesn't have the instructions to install pytorch. Please install pytorch 1.13.0 for the experiments

Get the Dataset

• Download the SmartFallMM data from this link. Put the dataset under data folder.

Choose and configure models

• Model configuration for InertialTransformer model is kept under config/smartfallmm/student.yaml.

• Model Configuration or BiScaleFormer is kept under config/smartfallmm/teacher.yaml for SmartFallMM dataset.

• Configuration for Distillation is stored in config/smartfallmm/distill.yaml for SmartFallMM dataset.

Step for Knowledge Distillation

• Train a teacher model first using main.py and config/smartfallmm/teacher.yaml. You can do it by uncommenting line 57 in train.sh .
• Perform knowledge distillation using distill.py and config/smartfallmm/distill.yaml . You can perform the distillation by uncomenting line 61.

Train and test

Give execution access to train.sh with

chmod +x ./train.sh

Run the train.sh to train and test the multimodal and accelerometer models. Log and weights would be saved under working directory. Use the following command to run the train.sh script.

./train.sh