{ "id": 487479, "title": "(1 stage only LB) VS (2 stage) here are my results", "url": "/competitions/hms-harmful-brain-activity-classification/discussion/487479", "author_name": "lhwcv", "post_date": "2024-03-29T07:10:56.089000", "votes": 22, "comment_count": 16, "views": 0, "content": "

same model, same data pipeline:

\n\n

what is your CV and LB if only use 1 stage (all eegs) ？
\nwill there be a shake up if only 2 stage?

", "messages": [ { "id": 2721717, "postDate": "2024-03-29T07:10:56.090Z", "content": "

same model, same data pipeline:

\n\n

what is your CV and LB if only use 1 stage (all eegs) ？
\nwill there be a shake up if only 2 stage?

", "rawMarkdown": "same model, same data pipeline:\n- one stage only: LB 0.33\n- 2 stage with votes>10: LB 0.25\n\nwhat is your CV and LB if only use 1 stage (all eegs) ？\nwill there be a shake up if only 2 stage?", "votes": 21 }, { "id": 2733931, "postDate": "2024-04-03T21:09:16.610Z", "content": "\n

Seems the 1st stage has a chance to be held in private so we betting on that it can improve a bit more. In stage 1 we observe the ensemble is not giving that big a boost only 0.01 so far ..

", "rawMarkdown": "- 1 stage LB : ~~0.36~~ 0.33 now\n- 2 stage LB : ~~0.3~~ 0.28 Now\n\nSeems the `1st stage` has a chance to be held in private so we betting on that it can improve a bit more. In stage 1 we observe the ensemble is not giving that big a boost only 0.01 so far .. ", "votes": 1 }, { "id": 2721768, "postDate": "2024-03-29T07:39:02.277Z", "content": "

What does the test data look like, and what are the possibilities?：

\n
    \n
  1. For each EEG, only 1 section annotated, they randomly selected 50 seconds, and the 10-minute Kaggle spec was annotated by experts
    2. \n
  2. Each EEG may have multiple sections, all annotated, and each section is used for testing
    3. \n
  3. Each EEG may have multiple sections, all annotated, but only the section with the most votes is used for testing
    4. \n
  4. Each EEG may have multiple sections, all annotated, but a section of the middle signal is cut out, the label is the average of all sections, and used for testing
    5. \n
  5. Each EEG may have multiple sections, all annotated, in the end, only those without differences of opinion are used for testing.
    6. \n
", "rawMarkdown": "What does the test data look like, and what are the possibilities?：\n\n1. For each EEG, only 1 section annotated, they randomly selected 50 seconds, and the 10-minute Kaggle spec was annotated by experts\n2. Each EEG may have multiple sections, all annotated, and each section is used for testing\n3. Each EEG may have multiple sections, all annotated, but only the section with the most votes is used for testing\n4. Each EEG may have multiple sections, all annotated, but a section of the middle signal is cut out, the label is the average of all sections, and used for testing\n5. Each EEG may have multiple sections, all annotated, in the end, only those without differences of opinion are used for testing.", "votes": 2 }, { "id": 2731972, "postDate": "2024-04-03T00:15:48.437Z", "content": "

(1 stage only LB) VS (2 stage) here are my results🫡

", "rawMarkdown": "(1 stage only LB) VS (2 stage) here are my results🫡" }, { "id": 2728585, "postDate": "2024-04-02T10:32:06.027Z", "content": "

@lihaoweicvch can I ask what you train on in the 1st stage? all votes? or votes < 10?

", "rawMarkdown": "@lihaoweicvch can I ask what you train on in the 1st stage? all votes? or votes < 10?" }, { "id": 2727833, "postDate": "2024-04-02T00:30:27.277Z", "content": "

Update
\nsingle model
\n1 stage only: LB 0.34
\n2 stage with votes>=10: LB 0.25

\n

will there be a shake if use 2 stage?

", "rawMarkdown": "Update\nsingle model\n1 stage only: LB 0.34\n2 stage with votes>=10: LB 0.25\n\nwill there be a shake if use 2 stage?", "replies": [ { "id": 2728508, "postDate": "2024-04-02T09:20:56.080Z", "content": "

You can consider this section of posts and how you are doing your CV. 2 stage or >=10 scores may be dependent on public LB content, it is hard to say. Am wary of private LB

", "rawMarkdown": "You can consider [this section of posts](https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/485697#2713543) and how you are doing your CV. 2 stage or >=10 scores may be dependent on public LB content, it is hard to say. Am wary of private LB ", "replies": [ { "id": 2728520, "postDate": "2024-04-02T09:27:26.897Z", "content": "

thank you for your sharing

", "rawMarkdown": "thank you for your sharing" }, { "id": 2732681, "postDate": "2024-04-03T09:16:09.377Z", "content": "

@lihaoweicvch - just a qualification: there would be differences in votes for eeg data and spectrogram data. e.g., if for eeg/spec data only using central 50s or 10s would want to consider relevant votes for that time section for eeg_id. And have noticed various ways of aggregating votes (first, sum, mean) while train provides votes at the label id level.
\nYou may already have taken this into consideration but just in case you want to review.

", "rawMarkdown": "@lihaoweicvch - just a qualification: there would be differences in votes for eeg data and spectrogram data. e.g., if for eeg/spec data only using central 50s or 10s would want to consider relevant votes for that time section for eeg_id. And have noticed various ways of aggregating votes (first, sum, mean) while train provides votes at the label id level. \nYou may already have taken this into consideration but just in case you want to review. ", "votes": 2 } ] } ] }, { "id": 2726433, "postDate": "2024-04-01T07:40:29.813Z", "content": "

one stage only: LB 0.32
\n2 stage with votes>10: LB 0.26

", "rawMarkdown": "one stage only: LB 0.32\n2 stage with votes>10: LB 0.26" }, { "id": 2726108, "postDate": "2024-04-01T01:13:01.483Z", "content": "

May i ask \"one stage only: LB 0.33\" is single model or ensemble ?

", "rawMarkdown": "May i ask \"one stage only: LB 0.33\" is single model or ensemble ?", "replies": [ { "id": 2726129, "postDate": "2024-04-01T02:04:08.783Z", "content": "

ensemble, single model 0.34 , 0.35

", "rawMarkdown": "ensemble, single model 0.34 , 0.35 ", "replies": [ { "id": 2726398, "postDate": "2024-04-01T06:58:02.827Z", "content": "

For me CV 0.55 LB 0.42 for 1 stage single model. CV 0.32 LB 0.32 for 2 stages

", "rawMarkdown": "For me CV 0.55 LB 0.42 for 1 stage single model. CV 0.32 LB 0.32 for 2 stages" }, { "id": 2726562, "postDate": "2024-04-01T09:03:17.457Z", "content": "

Hello @lihaoweicvch , Thanks for sharing this amazing insight!
\nSo, for the second stage of training, I just want to make sure that you have trained both of these two models. Am I right?

", "rawMarkdown": "Hello @lihaoweicvch , Thanks for sharing this amazing insight!\nSo, for the second stage of training, I just want to make sure that you have trained both of these two models. Am I right?" }, { "id": 2727831, "postDate": "2024-04-02T00:28:53.243Z", "content": "

@lmhongkhnh yes, these two models is trained

", "rawMarkdown": "@lmhongkhnh yes, these two models is trained" } ] } ] }, { "id": 2728584, "postDate": "2024-04-02T10:31:29.480Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2722053, "postDate": "2024-03-29T11:30:36.250Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2733931, "author_name": "Gaurav Rawat", "author_url": "", "post_date": "2024-04-03T21:09:16.610000", "content": "\n

Seems the 1st stage has a chance to be held in private so we betting on that it can improve a bit more. In stage 1 we observe the ensemble is not giving that big a boost only 0.01 so far ..

", "votes": 1, "replies": [] }, { "id": 2721768, "author_name": "lhwcv", "author_url": "", "post_date": "2024-03-29T07:39:02.277000", "content": "

What does the test data look like, and what are the possibilities?：

\n
    \n
  1. For each EEG, only 1 section annotated, they randomly selected 50 seconds, and the 10-minute Kaggle spec was annotated by experts
    2. \n
  2. Each EEG may have multiple sections, all annotated, and each section is used for testing
    3. \n
  3. Each EEG may have multiple sections, all annotated, but only the section with the most votes is used for testing
    4. \n
  4. Each EEG may have multiple sections, all annotated, but a section of the middle signal is cut out, the label is the average of all sections, and used for testing
    5. \n
  5. Each EEG may have multiple sections, all annotated, in the end, only those without differences of opinion are used for testing.
    6. \n
", "votes": 2, "replies": [] }, { "id": 2731972, "author_name": "oliviacsy", "author_url": "", "post_date": "2024-04-03T00:15:48.437000", "content": "

(1 stage only LB) VS (2 stage) here are my results🫡

", "votes": 0, "replies": [] }, { "id": 2728585, "author_name": "GabeTheHuman", "author_url": "", "post_date": "2024-04-02T10:32:06.027000", "content": "

@lihaoweicvch can I ask what you train on in the 1st stage? all votes? or votes < 10?

", "votes": 0, "replies": [] }, { "id": 2727833, "author_name": "lhwcv", "author_url": "", "post_date": "2024-04-02T00:30:27.277000", "content": "

Update
\nsingle model
\n1 stage only: LB 0.34
\n2 stage with votes>=10: LB 0.25

\n

will there be a shake if use 2 stage?

", "votes": 0, "replies": [ { "id": 2728508, "author_name": "something4kag", "author_url": "", "post_date": "2024-04-02T09:20:56.080000", "content": "

You can consider this section of posts and how you are doing your CV. 2 stage or >=10 scores may be dependent on public LB content, it is hard to say. Am wary of private LB

", "votes": 0, "replies": [ { "id": 2728520, "author_name": "lhwcv", "author_url": "", "post_date": "2024-04-02T09:27:26.897000", "content": "

thank you for your sharing

", "votes": 0, "replies": [] }, { "id": 2732681, "author_name": "something4kag", "author_url": "", "post_date": "2024-04-03T09:16:09.377000", "content": "

@lihaoweicvch - just a qualification: there would be differences in votes for eeg data and spectrogram data. e.g., if for eeg/spec data only using central 50s or 10s would want to consider relevant votes for that time section for eeg_id. And have noticed various ways of aggregating votes (first, sum, mean) while train provides votes at the label id level.
\nYou may already have taken this into consideration but just in case you want to review.

", "votes": 2, "replies": [] } ] } ] }, { "id": 2726433, "author_name": "Wisp Vale", "author_url": "", "post_date": "2024-04-01T07:40:29.813000", "content": "

one stage only: LB 0.32
\n2 stage with votes>10: LB 0.26

", "votes": 0, "replies": [] }, { "id": 2726108, "author_name": "Bùi Nhật Trường", "author_url": "", "post_date": "2024-04-01T01:13:01.483000", "content": "

May i ask \"one stage only: LB 0.33\" is single model or ensemble ?

", "votes": 0, "replies": [ { "id": 2726129, "author_name": "lhwcv", "author_url": "", "post_date": "2024-04-01T02:04:08.783000", "content": "

ensemble, single model 0.34 , 0.35

", "votes": 0, "replies": [ { "id": 2726398, "author_name": "Bùi Nhật Trường", "author_url": "", "post_date": "2024-04-01T06:58:02.827000", "content": "

For me CV 0.55 LB 0.42 for 1 stage single model. CV 0.32 LB 0.32 for 2 stages

", "votes": 0, "replies": [] }, { "id": 2726562, "author_name": "LHK", "author_url": "", "post_date": "2024-04-01T09:03:17.457000", "content": "

Hello @lihaoweicvch , Thanks for sharing this amazing insight!
\nSo, for the second stage of training, I just want to make sure that you have trained both of these two models. Am I right?

", "votes": 0, "replies": [] }, { "id": 2727831, "author_name": "lhwcv", "author_url": "", "post_date": "2024-04-02T00:28:53.243000", "content": "

@lmhongkhnh yes, these two models is trained

", "votes": 0, "replies": [] } ] } ] }, { "id": 2728584, "author_name": "", "author_url": "", "post_date": "2024-04-02T10:31:29.480000", "content": "", "votes": 0, "replies": [] }, { "id": 2722053, "author_name": "", "author_url": "", "post_date": "2024-03-29T11:30:36.250000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2721717": "same model, same data pipeline:\n- one stage only: LB 0.33\n- 2 stage with votes>10: LB 0.25\n\nwhat is your CV and LB if only use 1 stage (all eegs) ？\nwill there be a shake up if only 2 stage?", "2733931": "- 1 stage LB : ~~0.36~~ 0.33 now\n- 2 stage LB : ~~0.3~~ 0.28 Now\n\nSeems the `1st stage` has a chance to be held in private so we betting on that it can improve a bit more. In stage 1 we observe the ensemble is not giving that big a boost only 0.01 so far .. ", "2721768": "What does the test data look like, and what are the possibilities?：\n\n1. For each EEG, only 1 section annotated, they randomly selected 50 seconds, and the 10-minute Kaggle spec was annotated by experts\n2. Each EEG may have multiple sections, all annotated, and each section is used for testing\n3. Each EEG may have multiple sections, all annotated, but only the section with the most votes is used for testing\n4. Each EEG may have multiple sections, all annotated, but a section of the middle signal is cut out, the label is the average of all sections, and used for testing\n5. Each EEG may have multiple sections, all annotated, in the end, only those without differences of opinion are used for testing.", "2731972": "(1 stage only LB) VS (2 stage) here are my results🫡", "2728585": "@lihaoweicvch can I ask what you train on in the 1st stage? all votes? or votes < 10?", "2727833": "Update\nsingle model\n1 stage only: LB 0.34\n2 stage with votes>=10: LB 0.25\n\nwill there be a shake if use 2 stage?", "2726433": "one stage only: LB 0.32\n2 stage with votes>10: LB 0.26", "2726108": "May i ask \"one stage only: LB 0.33\" is single model or ensemble ?", "2728584": "", "2722053": "" } }