Skip to content

Latest commit

 

History

History
75 lines (59 loc) · 3.59 KB

README.md

File metadata and controls

75 lines (59 loc) · 3.59 KB
Raw

"Self-supervised ECG Representation Learning for Emotion Recognition" (TF v1.14.0)

Folllowing are the papers associated with this project:

Journal version: Self-supervised ECG Representation Learning for Emotion Recognition Authors: Sarkar and Etemad

Conference version: Self-Supervised Learning for ECG-Based Emotion Recognition Authors: Sarkar and Etemad

Proposed architecture

our proposed architecture

Requirements

  • Python >=3.6
  • TensorFlow = 1.14.0
  • TensorBoard = 1.14.0
  • Scikit-Learn = 0.22.2
  • NumPy = 1.18.4
  • Tqdm = 4.36.1
  • Pandas = 0.25.1
  • Mlxtend = 0.17.0

Usage

  • implementation: this directory contains all of our source codes.
    • Please create similar directory structure in your working directory:
      • data_folder: Keep your data in numpy format here.
      • implementation: Keep the codes here.
      • summaries: Tensorboard summaries will be saved here.
      • output: Loss and Results will be stored here.
      • models: Self-supervised models will be stored here.

  • load_model: this directory contains the pretrained self-supervised model and sample codes to use it.

    • The saved pretrained model can be used in order to extract features from raw ECG signals, which can be further used to perform downstream tasks.
    • We provide sample code for the above: extract_features.py.
    • In order to extract features, the input arrays must be in format of batch_size x window_size. We selected window_size of 10 seconds X 256 Hz = 2560 samples, where 256 Hz refers to the sampling rate. A sample ECG signal is given here.
    • We also provide sample code in order to save the weights of our pretrained network: save_weights.py

  • tips:

    • Try using larger batch size in the downstream task, that would boost performance.
    • Try full fine-tuning rather than fc-tuning (which I did) to boost up performance.
    • Try using larger batch for pre-training as well, this may help!

  • note:

    • I have received few emails and messages regarding missing processed data. As per the EULA of the original dataset, I am not allowed to share the processed data, so I could not upload them in this repo. Originally, I processed the datasets in Matlab separately, I added separately the preprocessing codes in written in Python, you may use this as reference: #1 (comment).

Citation

Please cite our papers for any purpose of usage.

@misc{sarkar2020selfsupervised,
    title={Self-supervised ECG Representation Learning for Emotion Recognition},
    author={Pritam Sarkar and Ali Etemad},
    year={2020},
    eprint={2002.03898},
    archivePrefix={arXiv},
    primaryClass={eess.SP}
}

@INPROCEEDINGS{sarkar2019selfsupervised,
  author={P. {Sarkar} and A. {Etemad}},
  booktitle={ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Self-Supervised Learning for ECG-Based Emotion Recognition}, 
  year={2020},
  volume={},
  number={},
  pages={3217-3221},}

Question

If you have any query or want to chat with me regarding our work please reach me at pritam.sarkar@queensu.ca or connect me in LinkedIN.