EegModuleMixin extension

[sliwy]

Adding input_shape, get_output_shape, to_dense_prediction_model to EEGModuleMixin, cf #457

Deprecating get_output_shape and to_dense_prediction_model, will be removed in version 1.0.

When we have a proper way of removing logsoftmax we may replace functions with methods everywhere.

Solve #432 with raising error suggesting that input size may be too small.

[codecov]

Codecov Report

Merging #514 (c58853b) into master (a29166e) will increase coverage by 0.06%.
The diff coverage is 94.73%.

❗ Current head c58853b differs from pull request most recent head 94f2959. Consider uploading reports for the commit 94f2959 to get more accurate results

@@            Coverage Diff             @@
##           master     #514      +/-   ##
==========================================
+ Coverage   84.33%   84.39%   +0.06%     
==========================================
  Files          63       63              
  Lines        4514     4525      +11     
==========================================
+ Hits         3807     3819      +12     
+ Misses        707      706       -1     

📢 Have feedback on the report? Share it here.

[sliwy]

Responding to @tgnassou comment:

function get_output_shape only works for class which heritates from nn.Sequential. I would like to use it to replace the function _len_last_layer of SleepStagerChambon (

braindecode/braindecode/models/sleep_stager_chambon_2018.py

Line 120 in 57280a6

def _len_last_layer(self, n_channels, input_size):

).

Actually the problem is in trying to use get_output_shape in nn.Module in a moment when part of a model that is used in forward does not exist. I don't see an easy fix. Maybe creating a method that performs forward until last fc layer can be possible and have a kind of convention how this method should be named.

@tgnassou what kind of error do you receive? maybe it would be easier to think of solution

[sliwy]

and yes, for partially initialized models it will work only for nn.Sequential because nn.Sequential forward can be called even when all the modules are not yet added to the model.

In case of nn.Module it will be problematic to do that because parts of model do not exist and are used directly in the forward. We may think of try-exception clause in case some modules are not defined yet? but I don't think it;s worth implementing

[PierreGtch]

Responding to @tgnassou comment:

function get_output_shape only works for class which heritates from nn.Sequential. I would like to use it to replace the function _len_last_layer of SleepStagerChambon (

braindecode/braindecode/models/sleep_stager_chambon_2018.py

Line 120 in 57280a6

def _len_last_layer(self, n_channels, input_size):

).

Actually the problem is in trying to use get_output_shape in nn.Module in a moment when part of a model that is used in forward does not exist. I don't see an easy fix. Maybe creating a method that performs forward until last fc layer can be possible and have a kind of convention how this method should be named.

@tgnassou what kind of error do you receive? maybe it would be easier to think of solution

@sliwy, @tgnassou, once @brunaafl will be finished with #520, the next step will be to make methods swap_final_layer and drop_final_layer that we allow to easily replace the final layer with a placeholder identity function. This way we will be able to also compute the output shape for the nn.Module!

[sliwy]

@PierreGtch It's the convention I thought of, if we have a specific name of the last layer then it's possible