kdFCN-4-tsc

This is the companion repository for our paper entitled "A study of Knowledge Distillation in Fully Convolutional Network for Time Series Classification", submitted to IJCNN 2022.

Datasets used

All our experiments were performed on the widely used time series classification benchmark UCR/UEA archive. We used the latest version (2018) and considered 112 datasets by discarding datasets containing series of unequal length or missing values, as well as the Fungi dataset, which only provides a single train case for each class label. The UCR archive is necesary to run the code.

Code

The code is divided as follows:

• The main.py python file contains the necessary code to run the experiments.
• The teacher.py python file contains the teacher implementation.
• The student.py python file contains the student implementation.
• The studentAlone.py python file contains the studentAlone implementation.
• The utils.py python file contains the necessary functions to read the datasets.
• The constants.py python file contains the definition of various constants.
• The copy_best_teacher.py python file contains the code to copy the results of the best teacher (in training)

Run the experiments

To simply run the experiments, you can type:

python3 main.py

It will run the experiments according to the constants.py configuration

Configuration

You can setup the configuration in the constants.py python file. In particular, you may need to modify:

• PATH_DATA: path to the folder where the UCR Archive is
• PATH_OUT: path to the folder where results will be saved
• CLASSIFIERS: list of classifiers to consider among 'teacher', 'StudentAlone' and 'Student'
• BEST_TEACHER_ONLY: set to True if you want to consider only the best teacher (in training) in while distilling to the student
• ALPHALIST: list of different alpha hyper-parameter to consider
• TEMPERATURELIST: list of different temperature hyper-parameter to consider
• FILTERS: list of number of filters to consider in first and third layers of student architectures
• FILTERS2: list of number of filters to consider in second layer of student architectures
• LAYER: number of layers to consider in student architectures
• SEPARABLE_CONV: set to True if you want to use depthwise separable convolutions

Steps

If you want to reproduce the experiments of the paper for a particular configuration:

1. Run the main.py script for 'teacher' and 'studentAlone' only
2. Run the copy_best_teacher.py script to copy the best teacher models
3. Run again the main.py script for 'student' model