Add MaskEncoding augmentation

[gustavohenriquesr]

Hi, I've implemented another data augmentation class called MaskEncoding.

The idea is similar to SmoothTimeMask but there is no smoothing of the masking. In addition, the masking can be divided into parts and spread across the signal (these parts can overlap, and I chose to leave it this way). This implementation was initially based on this article, and I modified it as I saw appropriate.

Adresses #628

[codecov]

Codecov Report

Attention: Patch coverage is 96.42857% with 1 line in your changes missing coverage. Please review.

Project coverage is 87.34%. Comparing base (726ee1a) to head (52a1d66).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #631      +/-   ##
==========================================
+ Coverage   87.29%   87.34%   +0.04%     
==========================================
  Files          67       67              
  Lines        6055     6083      +28     
==========================================
+ Hits         5286     5313      +27     
- Misses        769      770       +1     

[bruAristimunha]

Hey @gustavohenriquesr,

The code is already good, so I will ask something else: can you measure the impact of augmentation using the ShallowNet and BCNI datasets? To integrate something, we need to have some interesting results!

The idea would be to do a cross-subject approach with 5-fold, very similar to Cedric's tutorial, but with a splitter for cross-validation with the augmentation. You can decide whether to validate the augmentation parameters or the probability. Use the same parameters from https://github.com/eeg-augmentation-benchmark/eeg-augmentation-benchmark-2022/blob/1df778f7359581b1297d18dbdabf2b2fd4fcfa2f/eeg_augmentation_benchmark/models.py#L64C1-L78C57

Small inspiration to build the cross-validation object:

• the BraindecodeKFold from @dengemann, @hubertjb and @gemeinl: https://github.com/meeg-ml-benchmarks/brain-age-benchmark-paper/blob/c57eba38c8b90dac9354f0f4f8148dbed7e56029/deep_learning_utils.py#L25

• the benchopt splitters from @tomMoral: https://github.com/benchopt/benchmark_bci/blob/main/benchmark_utils/splitter.py

If you need an extension for the deadline, ping me on Discord ;)

[gustavohenriquesr]

Hello, @bruAristimunha

Thank you for the patience.

I did the tests you requested using a simplified strategy of what @cedricrommel proposed earlier for some other augmentations. The results show that using more than one segment benefits the model, while very high masking ratios when n_segments = 1 worsen performance. When we increase the number of segments, as in this case I tested with 5 segments, the model responds well up to a certain masking ratio.

Leave it to you to decide whether it's worth adding this method to Braindecode.

Later, with more time, I can do more tests with a wider range of values if you like.

[bruAristimunha]

Amazing work @gustavohenriquesr ;)

@agramfort, the green button is your!