Tensor Decompositions in Autoencoders for Electrocardiography Classification

Implementation in python of my thesis work in unsupervised pre-training using tensor decomposed autoencoders for ECG classification.

Thesis can be found at: repository.tudelft.nl

Getting started

1. Clone the repository to your local machine.
2. Create a virtual environment (recommended).
3. Install the required packages with: pip3 install -r requirements.txt
4. Create directories for the raw ECG files. The data generation scripts look for the raw data in this structure:

   ../
   └───physionet
   │   └───cpsc/
   │   └───incart/
   │   └───ptbxl/
   │   └───mitbih/
   │   └───g12ec/
   

5. Download the raw ECG files as a zip from physionet.org (recommended) or download with wget:
   • MIT-BIH: wget -r -N -c -np https://physionet.org/files/mitdb/1.0.0/.
   • INCART: wget -r -N -c -np https://physionet.org/files/incartdb/1.0.0/.
   • PTB-XL: wget -r -N -c -np https://physionet.org/files/ptb-xl/1.0.3/.
   • CPSC: wget -r -N -c -np https://physionet.org/files/cpsc2021/1.0.0/.
   • G12EC has to be downloaded directly at: https://www.kaggle.com/datasets/bjoernjostein/georgia-12lead-ecg-challenge-database?resource=download.
6. Place the .dat, .mat or .atr files from each of these datasets in their corresponding folders from step 4.
7. Create a data/ directory at the root of the project.
8. Run each of the data generation files under generation/. Ensure the correct directories are set at the top of each of the files. Example command for MIT-BIH: python3 generation/mitbih.py

Models

This thesis contains five different models:

• ResNet autoencoder
• ConvNeXt autoencoder
• U-Net
• Basic autoencoder
• CPD Basic autoencoder

Running the experiments

To run the experiment execute the four corresponding files: exp1.py, exp2_1.py, exp2_2.py and exp2_3.py. Each of the experiments can be run with the following arguments:

Parameter	Description	Default	type
--model	Name of model to train	default	str
--num_runs	Number of repeat runs	3	int
--batch_size_mae	Per GPU batch size	256	int
--epochs_mae	Number of epochs to train the autoencoder	50	int
--warmup_epochs_mae	Number of epochs to warmup LR for autoencoder	0	int
--batch_size_class	Per GPU batch size for classifier	256	int
--epochs_class	Number of epochs to train the classifier	20	int
--warmup_epochs_class	Number of epochs to warmup LR for classifier	0	int
--weight_decay_mae	Weight decay for autoencoder	1e-4	int
--lr_mae	Learning rate for autoencoder	5e-4	int
--min_lr_mae	Lower LR bound for cyclic schedulers for autoencoder	1e-5	int
--weight_decay_class	Weight decay for classifier	1e-4	int
--lr_class	Learning rate for classifier	5e-4	int
--min_lr_class	Lower LR bound for cyclic schedulers for classifier	1e-5	int
--gpu	Specify to use single GPU or all GPUs	all	str

Flags

Parameter	Description
--contrun	Flag to continue from last run
--no_train_mae	Do not train autoencoder
--no_train_class	Do not train classifier
--no_save_mae	Do not save autoencoder model
--no_save_class	Do not save classifier model

To reconstruct images use the function reconstruct_img in nn_funcs.py.