Add support for mne.Epochs in _EEGNeuralNet #529
Conversation
|
Be aware we are getting rid of mne epochs for efficiency reasons in #515 |
|
Yes when datasets are too large to load in memory.
This PR is to simplify the case when all the data can be loaded in memory and you typically just pass a numpy array. Now you will be able to pass an mne.Epochs
|
|
Hm no idea was to move away from mne epochs entirely within the braindecode dataset classes, both for code simplicity as well as due to long times instantiating mne epochs in the create windows functions. But here it looks like anyways there are no braindecode dataset classes involved? might be good to have both some examples of this for docs as well as tests so it can be better understood :) |
|
Exactly, here there is no braindecode dataset involved. It would just allow to do this: epo = mne.Epochs(...)
y = ...
net = EEGClassifier(...)
net.fit(epochs, y)It's only to facilitate the usage for small datasets. |
|
I loved =) |
|
Elaborating more... I think it's an interesting counterproposal to what Robin is removing... We removed this internal support to make it more external, this way the neuro user can use it more easily. Do you agree @robintibor? |
Codecov Report
@@ Coverage Diff @@
## master #529 +/- ##
==========================================
- Coverage 84.51% 84.51% -0.01%
==========================================
Files 63 63
Lines 4664 4662 -2
==========================================
- Hits 3942 3940 -2
Misses 722 722 |
|
In general yes, I think right now for me hard to keep overview over everything, once we have things implemented with examples we should see from user perspective whether things are clear. Like also if it's clear when to use which option etc. Also from developer perspective how much overhead we have from multiple ways of specifying the data... Here this part seems minimal so that seems good :) |
* Set signal-related parameters in check_data * Merge tests for EEGClassifier and EEGRegressor * Fix coquille * Update test to use module mixin * Rename clf to eegneuralnet in test * Add preds fixture * Update eegneuralnet.py and subclasses * Update test_eegneuralnet.py * restored previous tests behavior where nn layer is ignored and always returns mocked values * Update whats_new.rst * Use two different mock modules for test * Rename mock modules * Use set_params instead of vars * Deprecate passing an initialized module and skip setting signal args in that case * Remove unnecessary f-strings * Try fix python 3.8 * Fix case with non torch dataset * Try fix python 3.8 * Add docstrings for fit and partial_fit (already including #529) * Test initialized module * Fix Flake8 * Add email * Remove deprecation of initialized module --------- Co-authored-by: Maciej Sliwowski <macieksliwowski@gmail.com> Co-authored-by: Bru <a.bruno@aluno.ufabc.edu.br>
|
Thank you @PierreGtch and @robintibor =) |
* Set signal-related parameters in check_data * Merge tests for EEGClassifier and EEGRegressor * Fix coquille * Update test to use module mixin * Rename clf to eegneuralnet in test * Add preds fixture * Update eegneuralnet.py and subclasses * Update test_eegneuralnet.py * restored previous tests behavior where nn layer is ignored and always returns mocked values * Update whats_new.rst * Use two different mock modules for test * Rename mock modules * Use set_params instead of vars * Deprecate passing an initialized module and skip setting signal args in that case * Remove unnecessary f-strings * Try fix python 3.8 * Fix case with non torch dataset * Try fix python 3.8 * Add docstrings for fit and partial_fit (already including braindecode#529) * Test initialized module * Fix Flake8 * Add email * Remove deprecation of initialized module --------- Co-authored-by: Maciej Sliwowski <macieksliwowski@gmail.com> Co-authored-by: Bru <a.bruno@aluno.ufabc.edu.br>
Passing mne.Epochs will allow to automatically populate more parameters of the model (cf #517 )