[WIP] EEG-Inception #390

bruAristimunha · 2022-06-09T11:41:02Z

I am proposing to implement EEG-Inception. This model was used by the winners of last year BEETL Competition.

Is this suitable for library?

codecov · 2022-06-09T11:50:42Z

Codecov Report

Merging #390 (6b2720a) into master (884a78b) will increase coverage by 0.21%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #390      +/-   ##
==========================================
+ Coverage   82.81%   83.03%   +0.21%     
==========================================
  Files          54       55       +1     
  Lines        3789     3838      +49     
==========================================
+ Hits         3138     3187      +49     
  Misses        651      651

braindecode/models/eeginception.py

robintibor · 2022-06-09T12:10:03Z

In general more models are very suitable thank you very much!! Can you add some tests with different input sizes just checking forward works and output shape is as expected?

agramfort · 2022-06-09T12:16:25Z

great stuff as @robintibor would say :)

to get this merged we would need a test (see how done for others) and a what's new entry

btw I checked the license of https://github.com/Grifcc/EEG/tree/90e412a407c5242dfc953d5ffb490bdb32faf022 and it's Apache2 so it's ok for here (BSD).

bruAristimunha · 2022-06-09T12:42:18Z

Sure, I will do the tests. Glad I'm helping =)

robintibor · 2022-08-01T11:33:37Z

Do you need any support here? :) Is this still ongoing? :)

bruAristimunha · 2022-08-01T12:51:33Z

thanks for reminding me @robintibor, I will try to finish this week.

bruAristimunha · 2022-08-27T00:51:00Z

Hallo @robintibor!

I think I'm going to need some help X=
It is a little more difficult than expected to adjust the dimensions of the model.

bruAristimunha · 2022-08-30T21:58:31Z

Hello @sliwy, how are you?

I'm also in need of code review. I'm doing something wrong in the code and it's not working for dimensions other than the default value.

sliwy · 2022-08-31T08:23:49Z

Hi @bruAristimunha , I may try to help. I need more info from you when it fails, what is the error, maybe how to reproduce.

I think that we may be able to simplify a little bit by creating blocks and using nn.Sequential in the EEGInception implementation as well. I'm going to put some comments in the review.

sliwy

@bruAristimunha really great contribution with the Inception model, needed in the library for sure!

sliwy · 2022-08-31T08:28:58Z

braindecode/models/eeginception.py

+from .functions import squeeze_final_output
+
+
+class CustomPad(nn.Module):


I like that, we can use it as well in this PR #400

braindecode/models/eeginception.py

sliwy · 2022-08-31T09:03:27Z

braindecode/models/eeginception.py

+        self.sfreq = sfreq
+        self.n_filters = n_filters
+
+        scales_samples = [int(s * sfreq / input_window_samples) for s in scales_time]


I am not sure if this should depend on input_window_samples. I think that if we increase the size of the input window we may still want to keep the same scales, corresponding to length of selected miliseconds.

Maybe we can define kernel lengths in the inception in seconds. To me it is more natural.

…reases.

bruAristimunha · 2022-09-01T14:04:57Z

Hi @sliwy,

Strangely, the change to sequential increases the number of parameters. It still has a weird error when changing input values. I re-added the tests that were failing. I have no idea what's wrong or how to fix. I think I'll wait for the conclusion of the other PR and adapt based on the other code. The networks are similar.

sliwy · 2022-09-06T17:13:00Z

@bruAristimunha to debug easily the number of parameters change in the layers you can use https://github.com/TylerYep/torchinfo

fix conflict

bruAristimunha · 2022-09-06T18:48:48Z

I refactored it based on the other model, and here is the result of the torchinfo summary. I will go into more depth on this issue, but possibly it will take me a while.

===============================================================
Layer (type:depth-idx)                   Param #
=================================================================
EEGInception                             --
├─ELU: 1-1                               --
├─Ensure4d: 1-2                          --
├─Expression: 1-3                        --
├─_InceptionBlock: 1-4                   --
│    └─ModuleList: 2-1                   --
│    │    └─Sequential: 3-1              6,016
│    │    └─Sequential: 3-2              2,984
│    │    └─Sequential: 3-3              1,492
├─AvgPool2d: 1-5                         --
├─AvgPool2d: 1-6                         --
├─Sequential: 1-7                        --
│    └─CustomPad: 2-2                    --
│    └─Conv2d: 2-3                       2,304
│    └─BatchNorm2d: 2-4                  24
│    └─ELU: 2-5                          --
│    └─AvgPool2d: 2-6                    --
│    └─Dropout: 2-7                      --
│    └─CustomPad: 2-8                    --
│    └─Conv2d: 2-9                       288
│    └─BatchNorm2d: 2-10                 12
│    └─ELU: 2-11                         --
│    └─AvgPool2d: 2-12                   --
│    └─Dropout: 2-13                     --
├─Sequential: 1-8                        --
│    └─Flatten: 2-14                     --
│    └─Linear: 2-15                      3,474
│    └─Softmax: 2-16                     --
=================================================================
Total params: 16,594
Trainable params: 16,594
Non-trainable params: 0
=================================================================

=================================================================
Layer (type:depth-idx)                   Param #
=================================================================
EEGInception                             --
├─ELU: 1-1                               --
├─Ensure4d: 1-2                          --
├─Expression: 1-3                        --
├─_InceptionBlock: 1-4                   --
│    └─ModuleList: 2-1                   --
│    │    └─Sequential: 3-1              6,016
│    │    └─Sequential: 3-2              2,984
│    │    └─Sequential: 3-3              1,492
├─AvgPool2d: 1-5                         --
├─AvgPool2d: 1-6                         --
├─Sequential: 1-7                        --
│    └─CustomPad: 2-2                    --
│    └─Conv2d: 2-3                       2,304
│    └─BatchNorm2d: 2-4                  24
│    └─ELU: 2-5                          --
│    └─AvgPool2d: 2-6                    --
│    └─Dropout: 2-7                      --
│    └─CustomPad: 2-8                    --
│    └─Conv2d: 2-9                       288
│    └─BatchNorm2d: 2-10                 12
│    └─ELU: 2-11                         --
│    └─AvgPool2d: 2-12                   --
│    └─Dropout: 2-13                     --
├─Sequential: 1-8                        --
│    └─Flatten: 2-14                     --
│    └─Linear: 2-15                      3,474
│    └─Softmax: 2-16                     --
=================================================================
Total params: 16,594
Trainable params: 16,594
Non-trainable params: 0
=================================================================

bruAristimunha · 2022-09-06T19:29:34Z

thank you so much! I hadn't noticed the wrong formatting, I think I'll stop for today ;p

in fact, I think I sent a duplicate. But good idea, I'll do it (module vs sequential)

test/unit_tests/models/test_models.py

bruAristimunha · 2022-09-08T14:11:30Z

Hello guys, @robintibor and @sliwy,

I've talked to @cedricrommel, and he's going to take a closer look at this code for the next week.

cedricrommel · 2022-09-15T15:32:14Z

Hello guys, @robintibor and @sliwy,

I've talked to @cedricrommel, and he's going to take a closer look at this code for the next week.

I found a few problems with the layer dimensions and fix them. Both tests are now passing. For the second test I found in the authors official code provided by @bruAristimunha that they set all biases to False except in the first inception block. After fixing that and the window size in the test, we now get the number of params announced in the paper.

cedricrommel · 2022-09-15T15:36:47Z

In case it can help:

agramfort

@bruAristimunha can you add this model to the doc before merging?

🙏

cedricrommel · 2022-09-15T16:01:45Z

Apparently math.prod was introduced after python 3.7 😓 Can you fix that @bruAristimunha?

bruAristimunha · 2022-09-15T17:29:31Z

Done @agramfort and @cedricrommel,

Many thanks, @sliwy and @robintibor, for the review, and thank you, @cedricrommel, for the hard work on this model! I spent a lot of time trying to find these little details, and I am pleased that you, as an expert professional, managed to find them.

Happy to contribute to the library with this model that won last year's competition. I am open for further revisions if you feel it necessary, @agramfort, @robintibor and @sliwy.

agramfort · 2022-09-15T18:16:33Z

looks good https://output.circle-artifacts.com/output/job/a9f2fd77-ee59-4cc3-977e-63b895e24f25/artifacts/0/dev/generated/braindecode.models.EEGInception.html !

🙏 @bruAristimunha @cedricrommel

bruAristimunha added 3 commits June 9, 2022 03:46

Adding pre-draft of EEGinception.

6c3949e

First draft, using example from https://github.com/Grifcc/EEG.

586e4df

flake8

d9fd0f6

robintibor reviewed Jun 9, 2022

View reviewed changes

braindecode/models/eeginception.py Outdated Show resolved Hide resolved

bruAristimunha added 2 commits June 16, 2022 02:44

Adding draft of the test. Fix later.

fcf4950

fixing variable. finish later also.

a0ddafc

bruAristimunha added 2 commits August 27, 2022 01:45

Updating test, and the model to include Ensure4d.

f6e00dc

flake8

703dab2

bruAristimunha requested a review from sliwy August 30, 2022 21:58

sliwy reviewed Aug 31, 2022

View reviewed changes

bruAristimunha added 3 commits August 31, 2022 12:59

Updating model, transforming to Sequential.

7de6a44

flake8

990d7c5

Updating the model and investigating why the number of parameters inc…

e01bce4

…reases.

bruAristimunha requested a review from sliwy September 1, 2022 14:01

bruAristimunha added 4 commits September 6, 2022 18:55

Merge branch 'eeginception' into master

0b33502

Merge pull request #2 from bruAristimunha/master

188a018

fix conflict

refactoring the eeginception based on eegitnet.

5fd9016

Updating testing and model.

b2b35e1

robintibor reviewed Sep 7, 2022

View reviewed changes

test/unit_tests/models/test_models.py Show resolved Hide resolved

removing an unnecessary test.

ff30799

cedricrommel added 12 commits September 15, 2022 16:23

Replace CustomPad by padding="same"

5d58dee

Set filter sizes in seconds

a6edbbe

Fix nfilter and BN dim after depthwise convolution

75ce820

Fix average pooling along time dim

f228677

Fix name of first inception block

392e497

Fix input dim of 2nd inception block

d4a4822

Fix final block

2ac00a9

Linting

155e116

Fix order of dropout and pooling

73e23cd

Fix window size and scales in n_params test

5e37496

Add name to script

f1a6068

Fix bias=True in first inception block

c83226c

cedricrommel marked this pull request as ready for review September 15, 2022 15:33

Updating whats_new.rst

9eb760f

bruAristimunha requested review from robintibor and agramfort September 15, 2022 15:42

agramfort reviewed Sep 15, 2022

View reviewed changes

bruAristimunha added 3 commits September 15, 2022 17:18

changing math to numpy.

9287bc9

Updating the model to the doc.

b4f1041

Updating the doc.

6b2720a

agramfort merged commit 9cc8c68 into braindecode:master Sep 15, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP] EEG-Inception #390

[WIP] EEG-Inception #390

bruAristimunha commented Jun 9, 2022

codecov bot commented Jun 9, 2022 •

edited

Loading

robintibor commented Jun 9, 2022

agramfort commented Jun 9, 2022

bruAristimunha commented Jun 9, 2022

robintibor commented Aug 1, 2022

bruAristimunha commented Aug 1, 2022

bruAristimunha commented Aug 27, 2022

bruAristimunha commented Aug 30, 2022

sliwy commented Aug 31, 2022

sliwy left a comment

sliwy Aug 31, 2022

sliwy Aug 31, 2022

bruAristimunha commented Sep 1, 2022

sliwy commented Sep 6, 2022

bruAristimunha commented Sep 6, 2022 •

edited by sliwy

Loading

bruAristimunha commented Sep 6, 2022

bruAristimunha commented Sep 8, 2022

cedricrommel commented Sep 15, 2022 •

edited

Loading

cedricrommel commented Sep 15, 2022

agramfort left a comment

cedricrommel commented Sep 15, 2022

bruAristimunha commented Sep 15, 2022

agramfort commented Sep 15, 2022

		from .functions import squeeze_final_output


		class CustomPad(nn.Module):

[WIP] EEG-Inception #390

[WIP] EEG-Inception #390

Conversation

bruAristimunha commented Jun 9, 2022

codecov bot commented Jun 9, 2022 • edited Loading

Codecov Report

robintibor commented Jun 9, 2022

agramfort commented Jun 9, 2022

bruAristimunha commented Jun 9, 2022

robintibor commented Aug 1, 2022

bruAristimunha commented Aug 1, 2022

bruAristimunha commented Aug 27, 2022

bruAristimunha commented Aug 30, 2022

sliwy commented Aug 31, 2022

sliwy left a comment

Choose a reason for hiding this comment

sliwy Aug 31, 2022

Choose a reason for hiding this comment

sliwy Aug 31, 2022

Choose a reason for hiding this comment

bruAristimunha commented Sep 1, 2022

sliwy commented Sep 6, 2022

bruAristimunha commented Sep 6, 2022 • edited by sliwy Loading

bruAristimunha commented Sep 6, 2022

bruAristimunha commented Sep 8, 2022

cedricrommel commented Sep 15, 2022 • edited Loading

cedricrommel commented Sep 15, 2022

agramfort left a comment

Choose a reason for hiding this comment

cedricrommel commented Sep 15, 2022

bruAristimunha commented Sep 15, 2022

agramfort commented Sep 15, 2022

codecov bot commented Jun 9, 2022 •

edited

Loading

bruAristimunha commented Sep 6, 2022 •

edited by sliwy

Loading

cedricrommel commented Sep 15, 2022 •

edited

Loading