{ "id": 473821, "title": "Raw EEG CV vs LB", "url": "/competitions/hms-harmful-brain-activity-classification/discussion/473821", "author_name": "", "post_date": "2024-02-06T07:17:23.245120800Z", "votes": 33, "comment_count": 30, "views": 0, "content": "

What is your best cv/lb using only raw egg?
\nOurs is CV: 0.6 ; LB: 0.41 (single 5-folds model)

", "messages": [ { "id": "2638267", "postDate": "02/06/2024 07:17:23", "content": "

What is your best cv/lb using only raw egg?
\nOurs is CV: 0.6 ; LB: 0.41 (single 5-folds model)

", "rawMarkdown": "What is your best cv/lb using only raw egg?\nOurs is **CV: 0.6 ; LB: 0.41** (single 5-folds model)", "votes": null }, { "id": "2638286", "postDate": "02/06/2024 07:50:29", "content": "

I couldn't go beyond CV 0.65/0.66 (LB?) using 1D on raw eeg. Your CV is for just eegs is quite nice! Are you using 2D models on raw eegs?

", "rawMarkdown": "I couldn't go beyond CV 0.65/0.66 (LB?) using 1D on raw eeg. Your CV is for just eegs is quite nice! Are you using 2D models on raw eegs?", "votes": null }, { "id": "2638335", "postDate": "02/06/2024 08:21:13", "content": "

So far I can't go below 0.49 on the LB but I have only used 8 Channels + ECG so far. Is yours on all channels?
\nBest,
\nJan

", "rawMarkdown": "So far I can't go below 0.49 on the LB but I have only used 8 Channels + ECG so far. Is yours on all channels?\nBest,\nJan", "votes": null }, { "id": "2638404", "postDate": "02/06/2024 08:55:53", "content": "

Only 8 channels: the double banana montage.

", "rawMarkdown": "Only 8 channels: the double banana montage.", "votes": null }, { "id": "2638566", "postDate": "02/06/2024 11:25:05", "content": "

Great results, thanks for sharing. So far I have CV: .74 and LB .45 although I think there's a lot of room to improve my results.

\n

For my models (and the public waveform models) the CV->LB gap seems to be much bigger for the raw waveform than the public spec models, with the exception of yours, which gets similarish scores to Chris Deotte's efficientnet starter CV 0.59 - LB 0.43.

\n

Have you any intuition about why this is?

", "rawMarkdown": "Great results, thanks for sharing. So far I have **CV: .74 and LB .45** although I think there's a lot of room to improve my results.\n\nFor my models (and the public waveform models) the CV->LB gap seems to be much bigger for the raw waveform than the public spec models, with the exception of yours, which gets similarish scores to Chris Deotte's efficientnet starter CV 0.59 - LB 0.43.\n\nHave you any intuition about why this is?", "votes": null }, { "id": "2638602", "postDate": "02/06/2024 11:54:22", "content": "

Thanks for the reply! Do also downsample or do you use the 10000 time steps?
\nBest,
\nJan

", "rawMarkdown": "Thanks for the reply! Do also downsample or do you use the 10000 time steps?\nBest,\nJan", "votes": null }, { "id": "2638689", "postDate": "02/06/2024 12:38:14", "content": "

My Cv was 0.61 and LB 0.42 using the pythorch model

", "rawMarkdown": "My Cv was 0.61 and LB 0.42 using the pythorch model", "votes": null }, { "id": "2638694", "postDate": "02/06/2024 12:41:37", "content": "

yes I do, same as Chris

", "rawMarkdown": "yes I do, same as Chris", "votes": null }, { "id": "2640354", "postDate": "02/06/2024 21:01:45", "content": "

Thanks for your feedback, I still have not found my secret key to getting below CV .70 somehow but you helped by making sure that these two conditions are not the reason.

\n

Best,
\nJan

", "rawMarkdown": "Thanks for your feedback, I still have not found my secret key to getting below CV .70 somehow but you helped by making sure that these two conditions are not the reason.\n\nBest,\nJan", "votes": null }, { "id": "2640873", "postDate": "02/07/2024 05:58:48", "content": "

In my experiment,
\nusing EfficientNet for 8spectrograms (Kaggle spectrograms and EEG spectrograms)

\n", "rawMarkdown": "In my experiment,\nusing EfficientNet for 8spectrograms (Kaggle spectrograms and EEG spectrograms) \n* CV:0.6272\n* LB:0.41", "votes": null }, { "id": "2640950", "postDate": "02/07/2024 07:16:57", "content": "

@ihebch are you using the CV scheme of @cdeotte or @nischaydnk to achieve the CV of .6 ?

", "rawMarkdown": "ihebch are you using the CV scheme of @cdeotte or @nischaydnk to achieve the CV of .6 ?", "votes": null }, { "id": "2640959", "postDate": "02/07/2024 07:33:03", "content": "

5 GKF same as Chris

", "rawMarkdown": "5 GKF same as Chris", "votes": null }, { "id": "2642812", "postDate": "02/08/2024 12:41:16", "content": "

for me CV=0.613, LB =0.44 using raw eeg signals only.

", "rawMarkdown": "for me CV=0.613, LB =0.44 using raw eeg signals only.", "votes": null }, { "id": "2642926", "postDate": "02/08/2024 14:03:58", "content": "

Can you share your notebook? And what kind of model did you use? Thanks!

", "rawMarkdown": "Can you share your notebook? And what kind of model did you use? Thanks!", "votes": null }, { "id": "2642998", "postDate": "02/08/2024 14:55:20", "content": "

This is awesome @chaudharypriyanshu .. if you can share .Do you use augmentation ?

", "rawMarkdown": "This is awesome @chaudharypriyanshu .. if you can share .Do you use augmentation ?", "votes": null }, { "id": "2643043", "postDate": "02/08/2024 15:27:31", "content": "

yes i do. But all of them are custom augmentations.

", "rawMarkdown": "yes i do. But all of them are custom augmentations.", "votes": null }, { "id": "2650672", "postDate": "02/13/2024 16:09:22", "content": "

Different version of @nischaydnk 's 1D EEGNet LB 0.41

", "rawMarkdown": "Different version of @nischaydnk 's 1D EEGNet LB 0.41", "votes": null }, { "id": "2651047", "postDate": "02/13/2024 21:10:45", "content": "

What was the CV , if you dont mind telling ..

", "rawMarkdown": "What was the CV , if you dont mind telling ..", "votes": null }, { "id": "2659451", "postDate": "02/19/2024 20:55:59", "content": "

1D model (pytorch), raw EEG features

\n

OOF CV: 0.6765
\nLB: 0.42

\n

CV is calculated on 17k unique ids

", "rawMarkdown": "1D model (pytorch), raw EEG features \n\n`OOF CV: 0.6765 `\n`LB: 0.42 `\n\nCV is calculated on 17k unique ids", "votes": null }, { "id": "2659691", "postDate": "02/20/2024 05:00:13", "content": "

This is great .. I had a model with CV .667 and LB is only .44 another one with CV .656 and LB is .45 . What do you mean by 17K Unique IDs ? That means oof CV of 17k records right ?

", "rawMarkdown": "This is great .. I had a model with CV .667 and LB is only .44 another one with CV .656 and LB is .45 . What do you mean by 17K Unique IDs ? That means oof CV of 17k records right ?", "votes": null }, { "id": "2659803", "postDate": "02/20/2024 06:36:05", "content": "

Sorry I can't tell the CV score. Our raw EEG best model is LB 0.29 (close to 0.28) now.

", "rawMarkdown": "Sorry I can't tell the CV score. Our raw EEG best model is LB 0.29 (close to 0.28) now.", "votes": null }, { "id": "2659822", "postDate": "02/20/2024 06:59:11", "content": "

This result is too amazing！ If you don't mind, have you ever used augmentation on the original EEG?

", "rawMarkdown": "This result is too amazing！ If you don't mind, have you ever used augmentation on the original EEG?", "votes": null }, { "id": "2659977", "postDate": "02/20/2024 08:57:12", "content": "

right, my OOF is 17,089 x 6 (unique eeg_ids)

\n

regarding CV-LB correlation I'm on the same side

", "rawMarkdown": "right, my OOF is 17,089 x 6 (unique eeg_ids) \n\nregarding CV-LB correlation I'm on the same side", "votes": null }, { "id": "2659979", "postDate": "02/20/2024 09:01:16", "content": "

This is awesome .. good luck to you @gunesevitan

", "rawMarkdown": "This is awesome .. good luck to you @gunesevitan", "votes": null }, { "id": "2660012", "postDate": "02/20/2024 09:34:26", "content": "

Managed to get down to 0.42 on the LB through a lot of augmentations by now. I think there is still room for improvement as the examples of others seem to show. Appreciate every hint!

\n

Best,
\nJan

", "rawMarkdown": "Managed to get down to 0.42 on the LB through a lot of augmentations by now. I think there is still room for improvement as the examples of others seem to show. Appreciate every hint!\n\nBest,\nJan", "votes": null }, { "id": "2673003", "postDate": "02/28/2024 12:26:58", "content": "

Model 1, 2, 3 -> 0.27
\nModel 4, 5 -> 0.28
\nEqual Blend -> 0.26

\n

All of them are using raw EEGs.

", "rawMarkdown": "Model 1, 2, 3 -> 0.27\nModel 4, 5 -> 0.28\nEqual Blend -> 0.26\n\nAll of them are using raw EEGs.", "votes": null }, { "id": "2673009", "postDate": "02/28/2024 12:32:10", "content": "

Surprised how powerful Raw EEG is

", "rawMarkdown": "Surprised how powerful Raw EEG is", "votes": null }, { "id": "2673010", "postDate": "02/28/2024 12:32:16", "content": "

Good job!
\nis your CV/LB correlated so far?

", "rawMarkdown": "Good job!\nis your CV/LB correlated so far?", "votes": null }, { "id": "2673014", "postDate": "02/28/2024 12:36:11", "content": "

Yes highly correlated so far.

", "rawMarkdown": "Yes highly correlated so far.", "votes": null }, { "id": "2673120", "postDate": "02/28/2024 13:47:04", "content": "

Hey Gunes, very cool! I am currently stuck at .38 LB (CV:.56) with my raw EEG model but you gave me hope to keep on improving - thanks for sharing!

\n

Best,
\nJan

", "rawMarkdown": "Hey Gunes, very cool! I am currently stuck at .38 LB (CV:.56) with my raw EEG model but you gave me hope to keep on improving - thanks for sharing!\n\nBest,\nJan", "votes": null }, { "id": "2673208", "postDate": "02/28/2024 15:02:49", "content": "

Jan - interested in general idea of the augmentations you made on raw data models?

", "rawMarkdown": "[Jan](https://www.kaggle.com/janbrederecke) - interested in general idea of the augmentations you made on raw data models?", "votes": null } ], "comments": [ { "id": 2638286, "author_name": "pheadrus", "author_url": "", "post_date": "02/06/2024 07:50:29", "content": "

I couldn't go beyond CV 0.65/0.66 (LB?) using 1D on raw eeg. Your CV is for just eegs is quite nice! Are you using 2D models on raw eegs?

", "votes": null, "replies": [] }, { "id": 2638335, "author_name": "janbrederecke", "author_url": "", "post_date": "02/06/2024 08:21:13", "content": "

So far I can't go below 0.49 on the LB but I have only used 8 Channels + ECG so far. Is yours on all channels?
\nBest,
\nJan

", "votes": null, "replies": [ { "id": 2638404, "author_name": "ihebch", "author_url": "", "post_date": "02/06/2024 08:55:53", "content": "

Only 8 channels: the double banana montage.

", "votes": null, "replies": [ { "id": 2638602, "author_name": "janbrederecke", "author_url": "", "post_date": "02/06/2024 11:54:22", "content": "

Thanks for the reply! Do also downsample or do you use the 10000 time steps?
\nBest,
\nJan

", "votes": null, "replies": [ { "id": 2638694, "author_name": "ihebch", "author_url": "", "post_date": "02/06/2024 12:41:37", "content": "

yes I do, same as Chris

", "votes": null, "replies": [ { "id": 2640354, "author_name": "janbrederecke", "author_url": "", "post_date": "02/06/2024 21:01:45", "content": "

Thanks for your feedback, I still have not found my secret key to getting below CV .70 somehow but you helped by making sure that these two conditions are not the reason.

\n

Best,
\nJan

", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2660012, "author_name": "janbrederecke", "author_url": "", "post_date": "02/20/2024 09:34:26", "content": "

Managed to get down to 0.42 on the LB through a lot of augmentations by now. I think there is still room for improvement as the examples of others seem to show. Appreciate every hint!

\n

Best,
\nJan

", "votes": null, "replies": [ { "id": 2673208, "author_name": "pcjimmmy", "author_url": "", "post_date": "02/28/2024 15:02:49", "content": "

Jan - interested in general idea of the augmentations you made on raw data models?

", "votes": null, "replies": [] } ] } ] }, { "id": 2638566, "author_name": "feargalosullivan", "author_url": "", "post_date": "02/06/2024 11:25:05", "content": "

Great results, thanks for sharing. So far I have CV: .74 and LB .45 although I think there's a lot of room to improve my results.

\n

For my models (and the public waveform models) the CV->LB gap seems to be much bigger for the raw waveform than the public spec models, with the exception of yours, which gets similarish scores to Chris Deotte's efficientnet starter CV 0.59 - LB 0.43.

\n

Have you any intuition about why this is?

", "votes": null, "replies": [] }, { "id": 2638689, "author_name": "rafaelzimmermann1", "author_url": "", "post_date": "02/06/2024 12:38:14", "content": "

My Cv was 0.61 and LB 0.42 using the pythorch model

", "votes": null, "replies": [] }, { "id": 2640873, "author_name": "hideyukizushi", "author_url": "", "post_date": "02/07/2024 05:58:48", "content": "

In my experiment,
\nusing EfficientNet for 8spectrograms (Kaggle spectrograms and EEG spectrograms)

\n", "votes": null, "replies": [] }, { "id": 2640950, "author_name": "phoenix9032", "author_url": "", "post_date": "02/07/2024 07:16:57", "content": "

@ihebch are you using the CV scheme of @cdeotte or @nischaydnk to achieve the CV of .6 ?

", "votes": null, "replies": [ { "id": 2640959, "author_name": "ihebch", "author_url": "", "post_date": "02/07/2024 07:33:03", "content": "

5 GKF same as Chris

", "votes": null, "replies": [] } ] }, { "id": 2642812, "author_name": "chaudharypriyanshu", "author_url": "", "post_date": "02/08/2024 12:41:16", "content": "

for me CV=0.613, LB =0.44 using raw eeg signals only.

", "votes": null, "replies": [ { "id": 2642926, "author_name": "fraolgelana", "author_url": "", "post_date": "02/08/2024 14:03:58", "content": "

Can you share your notebook? And what kind of model did you use? Thanks!

", "votes": null, "replies": [] }, { "id": 2642998, "author_name": "phoenix9032", "author_url": "", "post_date": "02/08/2024 14:55:20", "content": "

This is awesome @chaudharypriyanshu .. if you can share .Do you use augmentation ?

", "votes": null, "replies": [ { "id": 2643043, "author_name": "chaudharypriyanshu", "author_url": "", "post_date": "02/08/2024 15:27:31", "content": "

yes i do. But all of them are custom augmentations.

", "votes": null, "replies": [] } ] } ] }, { "id": 2650672, "author_name": "gunesevitan", "author_url": "", "post_date": "02/13/2024 16:09:22", "content": "

Different version of @nischaydnk 's 1D EEGNet LB 0.41

", "votes": null, "replies": [ { "id": 2651047, "author_name": "phoenix9032", "author_url": "", "post_date": "02/13/2024 21:10:45", "content": "

What was the CV , if you dont mind telling ..

", "votes": null, "replies": [ { "id": 2659803, "author_name": "gunesevitan", "author_url": "", "post_date": "02/20/2024 06:36:05", "content": "

Sorry I can't tell the CV score. Our raw EEG best model is LB 0.29 (close to 0.28) now.

", "votes": null, "replies": [ { "id": 2659979, "author_name": "phoenix9032", "author_url": "", "post_date": "02/20/2024 09:01:16", "content": "

This is awesome .. good luck to you @gunesevitan

", "votes": null, "replies": [] } ] } ] }, { "id": 2659822, "author_name": "gentlezdh", "author_url": "", "post_date": "02/20/2024 06:59:11", "content": "

This result is too amazing！ If you don't mind, have you ever used augmentation on the original EEG?

", "votes": null, "replies": [] }, { "id": 2673003, "author_name": "gunesevitan", "author_url": "", "post_date": "02/28/2024 12:26:58", "content": "

Model 1, 2, 3 -> 0.27
\nModel 4, 5 -> 0.28
\nEqual Blend -> 0.26

\n

All of them are using raw EEGs.

", "votes": null, "replies": [ { "id": 2673009, "author_name": "gentlezdh", "author_url": "", "post_date": "02/28/2024 12:32:10", "content": "

Surprised how powerful Raw EEG is

", "votes": null, "replies": [] }, { "id": 2673010, "author_name": "ihebch", "author_url": "", "post_date": "02/28/2024 12:32:16", "content": "

Good job!
\nis your CV/LB correlated so far?

", "votes": null, "replies": [ { "id": 2673014, "author_name": "gunesevitan", "author_url": "", "post_date": "02/28/2024 12:36:11", "content": "

Yes highly correlated so far.

", "votes": null, "replies": [ { "id": 2673120, "author_name": "janbrederecke", "author_url": "", "post_date": "02/28/2024 13:47:04", "content": "

Hey Gunes, very cool! I am currently stuck at .38 LB (CV:.56) with my raw EEG model but you gave me hope to keep on improving - thanks for sharing!

\n

Best,
\nJan

", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 2659451, "author_name": "imeintanis", "author_url": "", "post_date": "02/19/2024 20:55:59", "content": "

1D model (pytorch), raw EEG features

\n

OOF CV: 0.6765
\nLB: 0.42

\n

CV is calculated on 17k unique ids

", "votes": null, "replies": [ { "id": 2659691, "author_name": "phoenix9032", "author_url": "", "post_date": "02/20/2024 05:00:13", "content": "

This is great .. I had a model with CV .667 and LB is only .44 another one with CV .656 and LB is .45 . What do you mean by 17K Unique IDs ? That means oof CV of 17k records right ?

", "votes": null, "replies": [ { "id": 2659977, "author_name": "imeintanis", "author_url": "", "post_date": "02/20/2024 08:57:12", "content": "

right, my OOF is 17,089 x 6 (unique eeg_ids)

\n

regarding CV-LB correlation I'm on the same side

", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2638267": "What is your best cv/lb using only raw egg?\nOurs is **CV: 0.6 ; LB: 0.41** (single 5-folds model)", "2638286": "I couldn't go beyond CV 0.65/0.66 (LB?) using 1D on raw eeg. Your CV is for just eegs is quite nice! Are you using 2D models on raw eegs?", "2638335": "So far I can't go below 0.49 on the LB but I have only used 8 Channels + ECG so far. Is yours on all channels?\nBest,\nJan", "2638404": "Only 8 channels: the double banana montage.", "2638566": "Great results, thanks for sharing. So far I have **CV: .74 and LB .45** although I think there's a lot of room to improve my results.\n\nFor my models (and the public waveform models) the CV->LB gap seems to be much bigger for the raw waveform than the public spec models, with the exception of yours, which gets similarish scores to Chris Deotte's efficientnet starter CV 0.59 - LB 0.43.\n\nHave you any intuition about why this is?", "2638602": "Thanks for the reply! Do also downsample or do you use the 10000 time steps?\nBest,\nJan", "2638689": "My Cv was 0.61 and LB 0.42 using the pythorch model", "2638694": "yes I do, same as Chris", "2640354": "Thanks for your feedback, I still have not found my secret key to getting below CV .70 somehow but you helped by making sure that these two conditions are not the reason.\n\nBest,\nJan", "2640873": "In my experiment,\nusing EfficientNet for 8spectrograms (Kaggle spectrograms and EEG spectrograms) \n* CV:0.6272\n* LB:0.41", "2640950": "ihebch are you using the CV scheme of @cdeotte or @nischaydnk to achieve the CV of .6 ?", "2640959": "5 GKF same as Chris", "2642812": "for me CV=0.613, LB =0.44 using raw eeg signals only.", "2642926": "Can you share your notebook? And what kind of model did you use? Thanks!", "2642998": "This is awesome @chaudharypriyanshu .. if you can share .Do you use augmentation ?", "2643043": "yes i do. But all of them are custom augmentations.", "2650672": "Different version of @nischaydnk 's 1D EEGNet LB 0.41", "2651047": "What was the CV , if you dont mind telling ..", "2659451": "1D model (pytorch), raw EEG features \n\n`OOF CV: 0.6765 `\n`LB: 0.42 `\n\nCV is calculated on 17k unique ids", "2659691": "This is great .. I had a model with CV .667 and LB is only .44 another one with CV .656 and LB is .45 . What do you mean by 17K Unique IDs ? That means oof CV of 17k records right ?", "2659803": "Sorry I can't tell the CV score. Our raw EEG best model is LB 0.29 (close to 0.28) now.", "2659822": "This result is too amazing！ If you don't mind, have you ever used augmentation on the original EEG?", "2659977": "right, my OOF is 17,089 x 6 (unique eeg_ids) \n\nregarding CV-LB correlation I'm on the same side", "2659979": "This is awesome .. good luck to you @gunesevitan", "2660012": "Managed to get down to 0.42 on the LB through a lot of augmentations by now. I think there is still room for improvement as the examples of others seem to show. Appreciate every hint!\n\nBest,\nJan", "2673003": "Model 1, 2, 3 -> 0.27\nModel 4, 5 -> 0.28\nEqual Blend -> 0.26\n\nAll of them are using raw EEGs.", "2673009": "Surprised how powerful Raw EEG is", "2673010": "Good job!\nis your CV/LB correlated so far?", "2673014": "Yes highly correlated so far.", "2673120": "Hey Gunes, very cool! I am currently stuck at .38 LB (CV:.56) with my raw EEG model but you gave me hope to keep on improving - thanks for sharing!\n\nBest,\nJan", "2673208": "[Jan](https://www.kaggle.com/janbrederecke) - interested in general idea of the augmentations you made on raw data models?" }, "source": "meta" }