Multi View Contrastive learning

[coderookie1994]

Authors: Heyan Gao (heyan3@illinois.edu), Sharthak Ghosh(sghos52@illinois.edu)
Paper: https://arxiv.org/pdf/2506.22393

Overview

This PR adds the following: -

1. SleepEEG multi view contrastive learning task
2. Multi view contrastive learning based on time series (paper implementation)
3. Generic n-view multi view contrastive learning (extension)

Task

1. The MVCLTrainingSleepEEG task returns sleep windows per patient based on the openly available SleepEDF dataset.
2. The key properties returned are xt, xd, xf. These are the Time, Derivative and Frequency tensors respectively.

Examples

The example is here examples\mvcl_training_sleepedf.ipynb.

1. A valid SleepEDF dataset path is necessary
2. It runs the task associated with it and runs the generic multi-view model

Models

There are two models, MultiViewContrastiveModel which is the generic model that can work with n number of views and MultiViewContrastiveTimeSeriesModel that specifically implments the paper as mentioned above.

Associated Unit-Tests

Task tests here tests\core\test_mvcl_training_sleepedf_task.py
Model tests here tests\core\test_mvcl.py

Dataset sources

https://physionet.org/content/sleep-edfx/1.0.0/

Complete List of files added/modified

1. docs/api/models/pyhealth.models.MVCL.rst
2. docs/api/tasks/pyhealth.tasks.MVCLTrainingSleepEEG.rst
3. docs/api/models.rst
4. docs/api/tasks.rst
5. examples/mvcl_training_sleepedf.ipynb
6. pyhealth/datasets/sleepedf.py
7. pyhealth/models/__init__.py
8. pyhealth/models/multi_view_contrastive_time_series_model.py
9. pyhealth/models/mvcl_model.py
10. pyhealth/tasks/__init__.py
11. pyhealth/tasks/mvcl_training_sleepedf_task.py
12. tests/core/test_mvcl.py
13. tests/core/test_mvcl_training_sleepedf_task.py

Bug Fix

In pyhealth/datasets/sleepedf.py the dev flag wasn't being propagated to the base class.