{ "id": 574257, "title": "What is your best single model without YOLO?", "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/574257", "author_name": "SUNFADED", "post_date": "2025-04-21T00:56:49.314000", "votes": 14, "comment_count": 29, "views": 0, "content": "

Mine is as below.
\nLB: 0.763(CV F2 score: 0.883)
\nencoder: efficientnet_v2_s

\n

I have no idea how to tackle the problem…

", "messages": [ { "id": 3183534, "postDate": "2025-04-21T00:56:49.313Z", "content": "

Mine is as below.
\nLB: 0.763(CV F2 score: 0.883)
\nencoder: efficientnet_v2_s

\n

I have no idea how to tackle the problem…

", "rawMarkdown": "Mine is as below.\nLB: 0.763(CV F2 score: 0.883)\nencoder: efficientnet_v2_s\n\nI have no idea how to tackle the problem...", "votes": 14 }, { "id": 3186615, "postDate": "2025-04-25T00:47:34.313Z", "content": "

I have a result with 0.968 fbeta score with some modifications. Now it is submitting.

", "rawMarkdown": "I have a result with 0.968 fbeta score with some modifications. Now it is submitting.", "votes": 3 }, { "id": 3212164, "postDate": "2025-05-29T12:03:56.273Z", "content": "

So here I realse all the improvements in this month after massive submissions and experiments:

\n

(7 hours submission)
\ncv: 0.968 lb: 0.845
\ncv: 0.973 lb: 0.856

\n

(11 hours submission)
\ncv: 0.983 lb: 0.865
\ncv: 0.9878 lb: ? (will submit tomorrow)

", "rawMarkdown": "So here I realse all the improvements in this month after massive submissions and experiments:\n\n(7 hours submission)\ncv: 0.968 lb: 0.845\ncv: 0.973 lb: 0.856\n\n(11 hours submission)\ncv: 0.983 lb: 0.865\ncv: 0.9878 lb: ? (will submit tomorrow)", "votes": 2, "replies": [ { "id": 3212236, "postDate": "2025-05-29T14:28:21.983Z", "content": "

That cv score is really impressive. Looking forward to read your final solution!

", "rawMarkdown": "That cv score is really impressive. Looking forward to read your final solution!" }, { "id": 3213437, "postDate": "2025-05-30T02:31:59.710Z", "content": "

Impressive scores! Did you calculate CV scores in all tomographs or that has 1 or 0 motors?

", "rawMarkdown": "Impressive scores! Did you calculate CV scores in all tomographs or that has 1 or 0 motors?" } ] }, { "id": 3185589, "postDate": "2025-04-23T14:35:27.773Z", "content": "

convnext_small.in12k_ft_in1k based on @junkoda LB：0.759

", "rawMarkdown": "convnext_small.in12k_ft_in1k based on @junkoda LB：0.759", "votes": 1, "replies": [ { "id": 3186155, "postDate": "2025-04-24T09:20:41.157Z", "content": "

Update: after they fixed the labels I got 0.73😭

", "rawMarkdown": "Update: after they fixed the labels I got 0.73😭" } ] }, { "id": 3183978, "postDate": "2025-04-21T13:48:44.580Z", "content": "

Same as your results, LB 0.763 with efficientnet backbones, and I just found that we are adjacent on LB haha.

", "rawMarkdown": "Same as your results, LB 0.763 with efficientnet backbones, and I just found that we are adjacent on LB haha.", "votes": 1 }, { "id": 3183954, "postDate": "2025-04-21T13:22:17.570Z", "content": "

Are you working with 2d or 3d? 35min is quite fast!

", "rawMarkdown": "Are you working with 2d or 3d? 35min is quite fast!", "votes": 1, "replies": [ { "id": 3183976, "postDate": "2025-04-21T13:46:00.170Z", "content": "

Thanks. I am using 2.5d models(using a small subset of z-axis slices).

", "rawMarkdown": "Thanks. I am using 2.5d models(using a small subset of z-axis slices).", "votes": 1, "replies": [ { "id": 3184022, "postDate": "2025-04-21T14:43:34.967Z", "content": "

Did the LB score differ significantly between 2.5D and 2D?

\n

I got an LB 0.757 with 2D ConvNeXt backbones.

", "rawMarkdown": "Did the LB score differ significantly between 2.5D and 2D?\n\nI got an LB 0.757 with 2D ConvNeXt backbones.", "votes": 2 }, { "id": 3184313, "postDate": "2025-04-22T01:18:44.637Z", "content": "

I have not tried 2d models yet, but ConvNeXt did not converge in my environment.

", "rawMarkdown": "I have not tried 2d models yet, but ConvNeXt did not converge in my environment.", "votes": 1 }, { "id": 3188153, "postDate": "2025-04-27T05:46:47.990Z", "content": "

Hi there, What is your 2D ConvNeXt score after rescore?

", "rawMarkdown": "Hi there, What is your 2D ConvNeXt score after rescore?" }, { "id": 3188302, "postDate": "2025-04-27T11:03:48.783Z", "content": "

It dropped from 0.757 to 0.719 😅

", "rawMarkdown": "It dropped from 0.757 to 0.719 😅", "votes": 1 } ] }, { "id": 3189292, "postDate": "2025-04-29T04:17:32.927Z", "content": "

Hi, What's the approach with efficientnet & ConvNeXt, are you training it as a classification task.

", "rawMarkdown": "Hi, What's the approach with efficientnet & ConvNeXt, are you training it as a classification task.", "replies": [ { "id": 3189436, "postDate": "2025-04-29T07:51:41.810Z", "content": "

I guess they are used as backbone.

", "rawMarkdown": "I guess they are used as backbone." } ] } ] }, { "id": 3183993, "postDate": "2025-04-21T14:13:21.577Z", "content": "

Besides YOLO, my best single model is based on @junkoda great notebook. I used convnext_small.in12k_ft_in1k as the encoder. I've tried 2.5d but it didn't seem to improve so I've stick to 2d.
\nCross validation scores:
\n0.8289, 0.8099, 0.8200, 0.8475, 0.8462 → CV: 0.8305 ± 0.0164
\nPublic LB: 0.674

", "rawMarkdown": "Besides YOLO, my best single model is based on @junkoda [great notebook](https://www.](https://www.kaggle.com/code/junkoda/speed-up-inference). I used convnext_small.in12k_ft_in1k as the encoder. I've tried 2.5d but it didn't seem to improve so I've stick to 2d.\nCross validation scores:\n0.8289, 0.8099, 0.8200, 0.8475, 0.8462 → CV: 0.8305 ± 0.0164\nPublic LB: 0.674", "votes": 2 }, { "id": 3183546, "postDate": "2025-04-21T01:25:15.540Z", "content": "

Could I ask if your local cv is calculated for a single fold or the average of 5 folds?

", "rawMarkdown": "Could I ask if your local cv is calculated for a single fold or the average of 5 folds?", "votes": 2, "replies": [ { "id": 3183556, "postDate": "2025-04-21T01:49:07.313Z", "content": "

Single fold, the computational cost of 5 folds is too high. I'm worried that my private LB score will be shaken.

", "rawMarkdown": "Single fold, the computational cost of 5 folds is too high. I'm worried that my private LB score will be shaken.", "votes": 2 }, { "id": 3183567, "postDate": "2025-04-21T02:28:45.433Z", "content": "

Mine is 5 folds(by tomo_id). Inference time(5 folds average) is about 35 min.

", "rawMarkdown": "Mine is 5 folds(by tomo_id). Inference time(5 folds average) is about 35 min.", "votes": 3, "replies": [ { "id": 3183570, "postDate": "2025-04-21T02:38:04.780Z", "content": "

That's really impressive. Did you also use 5-fold cross validation during the training phase?

", "rawMarkdown": "That's really impressive. Did you also use 5-fold cross validation during the training phase?", "votes": 3 }, { "id": 3183571, "postDate": "2025-04-21T02:40:53.957Z", "content": "

Yes. Just using KFold and split tomo_ids into 5 folds.

", "rawMarkdown": "Yes. Just using KFold and split tomo_ids into 5 folds.", "votes": 2 } ] } ] }, { "id": 3204870, "postDate": "2025-05-19T02:16:12.737Z", "content": "

Update
\nLB: 0.809
\n2stage 3d models were used.

", "rawMarkdown": "Update\nLB: 0.809\n2stage 3d models were used.", "replies": [ { "id": 3209925, "postDate": "2025-05-26T14:01:48.417Z", "content": "

Is there any update? I currently have a LB:808 single YOLO 1stage, but I can hardly get a LB score like you. I'm very curious about how you do model ensemble because I can't achieve LB scores like you at all!

", "rawMarkdown": "Is there any update? I currently have a LB:808 single YOLO 1stage, but I can hardly get a LB score like you. I'm very curious about how you do model ensemble because I can't achieve LB scores like you at all!", "replies": [ { "id": 3210377, "postDate": "2025-05-27T05:22:56.090Z", "content": "

Using 2stage 3d models, we could get LB 0.811 without blending. Ensembling with YOLO will boost a lot in LB, so if you manage to get a decent score in 3d models, it will help you.

", "rawMarkdown": "Using 2stage 3d models, we could get LB 0.811 without blending. Ensembling with YOLO will boost a lot in LB, so if you manage to get a decent score in 3d models, it will help you.", "votes": 2 } ] } ] }, { "id": 3185967, "postDate": "2025-04-24T04:02:34.453Z", "content": "

update after rescore
\nLB: 0.798(CV F2 score: 0.923)

", "rawMarkdown": "update after rescore\nLB: 0.798(CV F2 score: 0.923)" }, { "id": 3185172, "postDate": "2025-04-23T00:52:18.173Z", "content": "

Did you only use the competition data or also the additional data in discussions? Can you provide some loss curves?

", "rawMarkdown": "Did you only use the competition data or also the additional data in discussions? Can you provide some loss curves? ", "replies": [ { "id": 3185177, "postDate": "2025-04-23T01:04:29.833Z", "content": "

I used only the competition data. The loss curves are secret(because I tried some kinds of custom losses).

", "rawMarkdown": "I used only the competition data. The loss curves are secret(because I tried some kinds of custom losses).", "votes": 1, "replies": [ { "id": 3185259, "postDate": "2025-04-23T04:08:12.830Z", "content": "

what backbone model did you use, how long did you train?

", "rawMarkdown": "what backbone model did you use, how long did you train?" } ] } ] }, { "id": 3183545, "postDate": "2025-04-21T01:22:39.313Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3186615, "author_name": "Tom", "author_url": "", "post_date": "2025-04-25T00:47:34.313000", "content": "

I have a result with 0.968 fbeta score with some modifications. Now it is submitting.

", "votes": 3, "replies": [] }, { "id": 3212164, "author_name": "Tom", "author_url": "", "post_date": "2025-05-29T12:03:56.273000", "content": "

So here I realse all the improvements in this month after massive submissions and experiments:

\n

(7 hours submission)
\ncv: 0.968 lb: 0.845
\ncv: 0.973 lb: 0.856

\n

(11 hours submission)
\ncv: 0.983 lb: 0.865
\ncv: 0.9878 lb: ? (will submit tomorrow)

", "votes": 2, "replies": [ { "id": 3212236, "author_name": "Yksin Young", "author_url": "", "post_date": "2025-05-29T14:28:21.983000", "content": "

That cv score is really impressive. Looking forward to read your final solution!

", "votes": 0, "replies": [] }, { "id": 3213437, "author_name": "SUNFADED", "author_url": "", "post_date": "2025-05-30T02:31:59.710000", "content": "

Impressive scores! Did you calculate CV scores in all tomographs or that has 1 or 0 motors?

", "votes": 0, "replies": [] } ] }, { "id": 3185589, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-04-23T14:35:27.773000", "content": "

convnext_small.in12k_ft_in1k based on @junkoda LB：0.759

", "votes": 1, "replies": [ { "id": 3186155, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-04-24T09:20:41.157000", "content": "

Update: after they fixed the labels I got 0.73😭

", "votes": 0, "replies": [] } ] }, { "id": 3183978, "author_name": "Yksin Young", "author_url": "", "post_date": "2025-04-21T13:48:44.580000", "content": "

Same as your results, LB 0.763 with efficientnet backbones, and I just found that we are adjacent on LB haha.

", "votes": 1, "replies": [] }, { "id": 3183954, "author_name": "Carlos Pérez Ricardo", "author_url": "", "post_date": "2025-04-21T13:22:17.570000", "content": "

Are you working with 2d or 3d? 35min is quite fast!

", "votes": 1, "replies": [ { "id": 3183976, "author_name": "SUNFADED", "author_url": "", "post_date": "2025-04-21T13:46:00.170000", "content": "

Thanks. I am using 2.5d models(using a small subset of z-axis slices).

", "votes": 1, "replies": [ { "id": 3184022, "author_name": "Yohei Tomio", "author_url": "", "post_date": "2025-04-21T14:43:34.967000", "content": "

Did the LB score differ significantly between 2.5D and 2D?

\n

I got an LB 0.757 with 2D ConvNeXt backbones.

", "votes": 2, "replies": [] }, { "id": 3184313, "author_name": "SUNFADED", "author_url": "", "post_date": "2025-04-22T01:18:44.637000", "content": "

I have not tried 2d models yet, but ConvNeXt did not converge in my environment.

", "votes": 1, "replies": [] }, { "id": 3188153, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-04-27T05:46:47.990000", "content": "

Hi there, What is your 2D ConvNeXt score after rescore?

", "votes": 0, "replies": [] }, { "id": 3188302, "author_name": "Yohei Tomio", "author_url": "", "post_date": "2025-04-27T11:03:48.783000", "content": "

It dropped from 0.757 to 0.719 😅

", "votes": 1, "replies": [] } ] }, { "id": 3189292, "author_name": "Gowri Shankar Penugonda", "author_url": "", "post_date": "2025-04-29T04:17:32.927000", "content": "

Hi, What's the approach with efficientnet & ConvNeXt, are you training it as a classification task.

", "votes": 0, "replies": [ { "id": 3189436, "author_name": "ynhuhu", "author_url": "", "post_date": "2025-04-29T07:51:41.810000", "content": "

I guess they are used as backbone.

", "votes": 0, "replies": [] } ] } ] }, { "id": 3183993, "author_name": "Sergio Alvarez", "author_url": "", "post_date": "2025-04-21T14:13:21.577000", "content": "

Besides YOLO, my best single model is based on @junkoda great notebook. I used convnext_small.in12k_ft_in1k as the encoder. I've tried 2.5d but it didn't seem to improve so I've stick to 2d.
\nCross validation scores:
\n0.8289, 0.8099, 0.8200, 0.8475, 0.8462 → CV: 0.8305 ± 0.0164
\nPublic LB: 0.674

", "votes": 2, "replies": [] }, { "id": 3183546, "author_name": "AnnieGo", "author_url": "", "post_date": "2025-04-21T01:25:15.540000", "content": "

Could I ask if your local cv is calculated for a single fold or the average of 5 folds?

", "votes": 2, "replies": [ { "id": 3183556, "author_name": "Less", "author_url": "", "post_date": "2025-04-21T01:49:07.313000", "content": "

Single fold, the computational cost of 5 folds is too high. I'm worried that my private LB score will be shaken.

", "votes": 2, "replies": [] }, { "id": 3183567, "author_name": "SUNFADED", "author_url": "", "post_date": "2025-04-21T02:28:45.433000", "content": "

Mine is 5 folds(by tomo_id). Inference time(5 folds average) is about 35 min.

", "votes": 3, "replies": [ { "id": 3183570, "author_name": "Less", "author_url": "", "post_date": "2025-04-21T02:38:04.780000", "content": "

That's really impressive. Did you also use 5-fold cross validation during the training phase?

", "votes": 3, "replies": [] }, { "id": 3183571, "author_name": "SUNFADED", "author_url": "", "post_date": "2025-04-21T02:40:53.957000", "content": "

Yes. Just using KFold and split tomo_ids into 5 folds.

", "votes": 2, "replies": [] } ] } ] }, { "id": 3204870, "author_name": "SUNFADED", "author_url": "", "post_date": "2025-05-19T02:16:12.737000", "content": "

Update
\nLB: 0.809
\n2stage 3d models were used.

", "votes": 0, "replies": [ { "id": 3209925, "author_name": "shanzhong8", "author_url": "", "post_date": "2025-05-26T14:01:48.417000", "content": "

Is there any update? I currently have a LB:808 single YOLO 1stage, but I can hardly get a LB score like you. I'm very curious about how you do model ensemble because I can't achieve LB scores like you at all!

", "votes": 0, "replies": [ { "id": 3210377, "author_name": "SUNFADED", "author_url": "", "post_date": "2025-05-27T05:22:56.090000", "content": "

Using 2stage 3d models, we could get LB 0.811 without blending. Ensembling with YOLO will boost a lot in LB, so if you manage to get a decent score in 3d models, it will help you.

", "votes": 2, "replies": [] } ] } ] }, { "id": 3185967, "author_name": "SUNFADED", "author_url": "", "post_date": "2025-04-24T04:02:34.453000", "content": "

update after rescore
\nLB: 0.798(CV F2 score: 0.923)

", "votes": 0, "replies": [] }, { "id": 3185172, "author_name": "thisArmin", "author_url": "", "post_date": "2025-04-23T00:52:18.173000", "content": "

Did you only use the competition data or also the additional data in discussions? Can you provide some loss curves?

", "votes": 0, "replies": [ { "id": 3185177, "author_name": "SUNFADED", "author_url": "", "post_date": "2025-04-23T01:04:29.833000", "content": "

I used only the competition data. The loss curves are secret(because I tried some kinds of custom losses).

", "votes": 1, "replies": [ { "id": 3185259, "author_name": "thisArmin", "author_url": "", "post_date": "2025-04-23T04:08:12.830000", "content": "

what backbone model did you use, how long did you train?

", "votes": 0, "replies": [] } ] } ] }, { "id": 3183545, "author_name": "", "author_url": "", "post_date": "2025-04-21T01:22:39.313000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3183534": "Mine is as below.\nLB: 0.763(CV F2 score: 0.883)\nencoder: efficientnet_v2_s\n\nI have no idea how to tackle the problem...", "3186615": "I have a result with 0.968 fbeta score with some modifications. Now it is submitting.", "3212164": "So here I realse all the improvements in this month after massive submissions and experiments:\n\n(7 hours submission)\ncv: 0.968 lb: 0.845\ncv: 0.973 lb: 0.856\n\n(11 hours submission)\ncv: 0.983 lb: 0.865\ncv: 0.9878 lb: ? (will submit tomorrow)", "3185589": "convnext_small.in12k_ft_in1k based on @junkoda LB：0.759", "3183978": "Same as your results, LB 0.763 with efficientnet backbones, and I just found that we are adjacent on LB haha.", "3183954": "Are you working with 2d or 3d? 35min is quite fast!", "3183993": "Besides YOLO, my best single model is based on @junkoda [great notebook](https://www.](https://www.kaggle.com/code/junkoda/speed-up-inference). I used convnext_small.in12k_ft_in1k as the encoder. I've tried 2.5d but it didn't seem to improve so I've stick to 2d.\nCross validation scores:\n0.8289, 0.8099, 0.8200, 0.8475, 0.8462 → CV: 0.8305 ± 0.0164\nPublic LB: 0.674", "3183546": "Could I ask if your local cv is calculated for a single fold or the average of 5 folds?", "3204870": "Update\nLB: 0.809\n2stage 3d models were used.", "3185967": "update after rescore\nLB: 0.798(CV F2 score: 0.923)", "3185172": "Did you only use the competition data or also the additional data in discussions? Can you provide some loss curves? ", "3183545": "" } }