{ "id": 575108, "title": "Updated Best Solo Models?", "url": "/competitions/byu-locating-bacterial-flagellar-motors-2025/discussion/575108", "author_name": "Cody_Null", "post_date": "2025-04-26T03:18:44.715000", "votes": 12, "comment_count": 45, "views": 0, "content": "

Curious how people’s solo models are doing. Currently most our best solo models are at around .8 but we expect them to be much higher in about a week.

", "messages": [ { "id": 3188253, "postDate": "2025-04-27T09:45:53.073Z", "content": "

GNN solution 0.83+ without setting any confidence threshold, can be better with ensemble. I expect to achieve 0.87+.

\n

Basically it is a 3d approach.

\n

\"\"

", "rawMarkdown": "GNN solution 0.83+ without setting any confidence threshold, can be better with ensemble. I expect to achieve 0.87+. \n\nBasically it is a 3d approach.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4310004%2Fd9df241f2a0e54edf238323a962033cd%2Fyolo_pred_structure.png?generation=1745747696576552&alt=media)", "votes": 20, "replies": [ { "id": 3188314, "postDate": "2025-04-27T11:49:05.317Z", "content": "

This is one of the more creative solutions of this competition, I really applaud your out of the box approach to this problem. I think I am not the only one who can't wait to read the solutions once this finishes!

", "rawMarkdown": "This is one of the more creative solutions of this competition, I really applaud your out of the box approach to this problem. I think I am not the only one who can't wait to read the solutions once this finishes!", "votes": 2, "replies": [ { "id": 3188584, "postDate": "2025-04-27T21:56:37.620Z", "content": "

This is super interesting, yes I cant wait to hear more about it too! I tried a 3D approach at the beginning but wasn't working out the best. I think your solution is easily the most interesting and I also have 0 faith in myself to be able to replicate it with GNN so this will be a super cool learning opportunity!

", "rawMarkdown": "This is super interesting, yes I cant wait to hear more about it too! I tried a 3D approach at the beginning but wasn't working out the best. I think your solution is easily the most interesting and I also have 0 faith in myself to be able to replicate it with GNN so this will be a super cool learning opportunity!", "votes": 2 }, { "id": 3191999, "postDate": "2025-05-02T10:44:54.867Z", "content": "

@cody11null 3d approach can work (I use the codebase from @christofhenkel ctyo solution). The biggest problem is 3d unet or ResSegNet have hard time to model a single positive point. I saw people struggling on positive weight or focal loss hyperparamter turning. So we have to put more effort on label development and prediction design. For me, I decide to sparsely predict the motor in 3d structure instead of reading entire volume. The result is pretty good. Faster and stronger than any 3d model which take volume as input.
\nOne thing I can tell you is I assume the subgraphs formed from keypoints in the volume as strongly connected components.

", "rawMarkdown": "@cody11null 3d approach can work (I use the codebase from @christofhenkel ctyo solution). The biggest problem is 3d unet or ResSegNet have hard time to model a single positive point. I saw people struggling on positive weight or focal loss hyperparamter turning. So we have to put more effort on label development and prediction design. For me, I decide to sparsely predict the motor in 3d structure instead of reading entire volume. The result is pretty good. Faster and stronger than any 3d model which take volume as input. \nOne thing I can tell you is I assume the subgraphs formed from keypoints in the volume as strongly connected components.", "votes": 3 } ] }, { "id": 3188885, "postDate": "2025-04-28T11:37:40.933Z", "content": "

Looks amazing, but I have no idea what's going on here. Will wait for your solution to learn something totally orthogonal to what I know :)

", "rawMarkdown": "Looks amazing, but I have no idea what's going on here. Will wait for your solution to learn something totally orthogonal to what I know :)", "votes": 1 }, { "id": 3188980, "postDate": "2025-04-28T15:11:23.507Z", "content": "

@tom99763 it is not a simplicial complex, right? U don't have 2-simplices in it, right? Only 0- and 1-simplices?

", "rawMarkdown": "@tom99763 it is not a simplicial complex, right? U don't have 2-simplices in it, right? Only 0- and 1-simplices?", "replies": [ { "id": 3191996, "postDate": "2025-05-02T10:41:06.550Z", "content": "

@eikyou The graph construction is quite tricky :)

", "rawMarkdown": "@eikyou The graph construction is quite tricky :)\n", "votes": 1 } ] }, { "id": 3191752, "postDate": "2025-05-02T05:30:11.420Z", "content": "

are you using only kaggle gpu for training

", "rawMarkdown": "are you using only kaggle gpu for training", "replies": [ { "id": 3191993, "postDate": "2025-05-02T10:39:22.997Z", "content": "

@thisarmin I train it locally. But training on kaggle is completely OK cuz my model is really lightweight.

", "rawMarkdown": "@thisarmin I train it locally. But training on kaggle is completely OK cuz my model is really lightweight.", "votes": 1 } ] }, { "id": 3191840, "postDate": "2025-05-02T08:01:24.833Z", "content": "

I'll release my github respository after this comp ends.

", "rawMarkdown": "I'll release my github respository after this comp ends.", "votes": 4 }, { "id": 3202814, "postDate": "2025-05-16T00:35:04.337Z", "content": "

what GNN are you using or custom message passing layers??

", "rawMarkdown": "what GNN are you using or custom message passing layers??", "replies": [ { "id": 3202853, "postDate": "2025-05-16T02:55:01.240Z", "content": "

If you want to predict unseen graph, you need to use the one who is trained transductively.

", "rawMarkdown": "If you want to predict unseen graph, you need to use the one who is trained transductively.\n" } ] } ] }, { "id": 3187436, "postDate": "2025-04-26T03:18:44.717Z", "content": "

Curious how people’s solo models are doing. Currently most our best solo models are at around .8 but we expect them to be much higher in about a week.

", "rawMarkdown": "Curious how people’s solo models are doing. Currently most our best solo models are at around .8 but we expect them to be much higher in about a week. ", "votes": 12 }, { "id": 3188870, "postDate": "2025-04-28T11:04:28.183Z", "content": "

My best LB 0.818 as of today is from a single YOLO model (\"best.pt\" only)

\n", "rawMarkdown": "My best **LB 0.818** as of today is from a single YOLO model (**\"best.pt\" only**)\n* No pre-processing\n* No post-processing\n* No external data used", "votes": 7, "replies": [ { "id": 3188889, "postDate": "2025-04-28T11:47:21.870Z", "content": "

What does \"no preprocessing\" mean? Just use the given slice?

", "rawMarkdown": "What does \"no preprocessing\" mean? Just use the given slice?" }, { "id": 3189216, "postDate": "2025-04-29T00:38:02.240Z", "content": "

can you tell us your f1 and map scores?

", "rawMarkdown": "can you tell us your f1 and map scores?" }, { "id": 3190964, "postDate": "2025-05-01T07:56:45.137Z", "content": "

Hi @hideyukizushi what's the version of your YOLO model?

", "rawMarkdown": "Hi @hideyukizushi what's the version of your YOLO model?" }, { "id": 3202150, "postDate": "2025-05-14T23:44:19.470Z", "content": "

Did u use tta and what were your training augmentations or while creating yolodataset

", "rawMarkdown": "Did u use tta and what were your training augmentations or while creating yolodataset", "isDeleted": true } ] }, { "id": 3193410, "postDate": "2025-05-04T10:48:11.573Z", "content": "

0.778, I focused too much on stability, and I think i finally managed to find a stable pipeline. Changing the confidence threshold from 0.05 up to 0.5 gives lb results between 0.755 to 0.778 which is very stable i guess. Now time to stop stability and improve results :)

", "rawMarkdown": "0.778, I focused too much on stability, and I think i finally managed to find a stable pipeline. Changing the confidence threshold from 0.05 up to 0.5 gives lb results between 0.755 to 0.778 which is very stable i guess. Now time to stop stability and improve results :)", "votes": 3, "replies": [ { "id": 3195064, "postDate": "2025-05-06T15:32:02.747Z", "content": "

Tried a bit of a crazy idea for cv (based on a specific data insight), looks like i managed to get a bit of correlation:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
CV ScoreLB Score
0.02820.578
0.29210.710
0.31910.751
0.42830.766
0.46440.738
\n

Will update if it worked for new subs

", "rawMarkdown": "Tried a bit of a crazy idea for cv (based on a specific data insight), looks like i managed to get a bit of correlation:\n\n| CV Score | LB Score |\n|----------|----------|\n| 0.0282 | 0.578 |\n| 0.2921 | 0.710 |\n| 0.3191 | 0.751 |\n| 0.4283 | 0.766 |\n| 0.4644 | 0.738 |\n\n\nWill update if it worked for new subs", "votes": 3, "replies": [ { "id": 3197149, "postDate": "2025-05-07T21:04:43.643Z", "content": "

Had some bugs with the previous cv scores. Fixed now.

\n

\"\"
\nNot the best, but i can see some good trend.

", "rawMarkdown": "Had some bugs with the previous cv scores. Fixed now.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F7311816%2F4461254c35f85d2f4cfd02a15521c4dd%2Foutput.png?generation=1746651857791231&alt=media)\nNot the best, but i can see some good trend.", "votes": 2 }, { "id": 3197256, "postDate": "2025-05-08T01:16:08.210Z", "content": "

May I ask what metric you are using for your CV score? It seems a bit lower than expected—are you perhaps using the F-beta score or a similar metric?

", "rawMarkdown": "May I ask what metric you are using for your CV score? It seems a bit lower than expected—are you perhaps using the F-beta score or a similar metric?" }, { "id": 3197549, "postDate": "2025-05-08T09:18:06.447Z", "content": "

F-beta, but i changed something in the data for this split.

", "rawMarkdown": "F-beta, but i changed something in the data for this split." } ] } ] }, { "id": 3193476, "postDate": "2025-05-04T13:05:23.867Z", "content": "

LB 0.825 with the pre-trained yolo8m, without external data.

", "rawMarkdown": "LB 0.825 with the pre-trained yolo8m, without external data.", "votes": 4, "replies": [ { "id": 3196749, "postDate": "2025-05-07T11:41:52.177Z", "content": "

Good! I got 0.693 with pre-trained yolo8m

", "rawMarkdown": "Good! I got 0.693 with pre-trained yolo8m" }, { "id": 3198623, "postDate": "2025-05-09T17:16:45.180Z", "content": "

I would like to ask if the pre-trained yolo8m was downloaded from the official website of ultralytics? Or was it obtained from other channels? Is it legal to download it from the official website? Does it violate the relevant regulations of the competition?

", "rawMarkdown": "I would like to ask if the pre-trained yolo8m was downloaded from the official website of ultralytics? Or was it obtained from other channels? Is it legal to download it from the official website? Does it violate the relevant regulations of the competition?" }, { "id": 3202147, "postDate": "2025-05-14T23:38:40.800Z", "content": "

What augmentations did u use for training?? Also did u use tta ??

", "rawMarkdown": "What augmentations did u use for training?? Also did u use tta ??", "isDeleted": true } ] }, { "id": 3187609, "postDate": "2025-04-26T09:31:30.463Z", "content": "

Mine is 0.831 lb

", "rawMarkdown": "Mine is 0.831 lb", "votes": 3, "replies": [ { "id": 3193907, "postDate": "2025-05-05T07:32:22.090Z", "content": "

Do you use Yolo?

", "rawMarkdown": "Do you use Yolo?" } ] }, { "id": 3190199, "postDate": "2025-04-30T09:29:45.633Z", "content": "

0.794 lb model, yolo11l, only positive samples

", "rawMarkdown": "0.794 lb model, yolo11l, only positive samples\n", "votes": 1, "replies": [ { "id": 3190499, "postDate": "2025-04-30T18:27:46.063Z", "content": "

what is your batch size? i am using Kaggle environment, 8 is the best I can do with yolov11L.

", "rawMarkdown": "what is your batch size? i am using Kaggle environment, 8 is the best I can do with yolov11L.", "replies": [ { "id": 3190543, "postDate": "2025-04-30T19:42:50.520Z", "content": "

@konohayui batch size 16

", "rawMarkdown": "@konohayui batch size 16", "votes": 1 } ] }, { "id": 3192669, "postDate": "2025-05-03T07:54:18.160Z", "content": "

Hey, what do you mean only positive samples?

", "rawMarkdown": "Hey, what do you mean only positive samples?", "replies": [ { "id": 3195105, "postDate": "2025-05-06T16:31:00.503Z", "content": "

@vasileioscharatsidis i use only slices, where motors are present

", "rawMarkdown": "@vasileioscharatsidis i use only slices, where motors are present" } ] }, { "id": 3198625, "postDate": "2025-05-09T17:21:03.083Z", "content": "

I would like to ask whether the yolo model was downloaded from the official website of Ultralytics? Or was it obtained from other channels? Is it legal to download it from the official website? Does it violate the relevant regulations of the competition?

", "rawMarkdown": "I would like to ask whether the yolo model was downloaded from the official website of Ultralytics? Or was it obtained from other channels? Is it legal to download it from the official website? Does it violate the relevant regulations of the competition?" }, { "id": 3202151, "postDate": "2025-05-14T23:45:12.620Z", "content": "

What were your training epochs and training time (which gpu)

", "rawMarkdown": "What were your training epochs and training time (which gpu)", "isDeleted": true } ] }, { "id": 3190125, "postDate": "2025-04-30T06:32:55.957Z", "content": "

0.785lb RTDETR Model . Trained on images with only one motor and negative samples.

", "rawMarkdown": "0.785lb RTDETR Model . Trained on images with only one motor and negative samples.", "votes": 1 }, { "id": 3187498, "postDate": "2025-04-26T05:36:31.903Z", "content": "

Hi, @cody11null
\nWhat do you mean by solo model? A single checkpoint? Average of several checkpoints from a single fold? Average of several folds of a single model?

", "rawMarkdown": "Hi, @cody11null \nWhat do you mean by solo model? A single checkpoint? Average of several checkpoints from a single fold? Average of several folds of a single model?", "votes": 1, "replies": [ { "id": 3187509, "postDate": "2025-04-26T06:07:03.023Z", "content": "

I was meaning single checkpoints but I guess both are interesting!

", "rawMarkdown": "I was meaning single checkpoints but I guess both are interesting!", "replies": [ { "id": 3187570, "postDate": "2025-04-26T08:11:55.580Z", "content": "

I didn't do single chkp inference. My score is from a single fold multiple checkpoints. 0.818

", "rawMarkdown": "I didn't do single chkp inference. My score is from a single fold multiple checkpoints. 0.818", "votes": 1 }, { "id": 3187776, "postDate": "2025-04-26T14:28:34.227Z", "content": "

Are you just doing simple ensemble of averaging the last 3 sets of weights or something then or are you doing multiple full trains?

", "rawMarkdown": "Are you just doing simple ensemble of averaging the last 3 sets of weights or something then or are you doing multiple full trains?" }, { "id": 3187810, "postDate": "2025-04-26T15:02:43.827Z", "content": "

Just simple average of predicts from three best checkpoints from the sinle train run

", "rawMarkdown": "Just simple average of predicts from three best checkpoints from the sinle train run", "votes": 4 }, { "id": 3190302, "postDate": "2025-04-30T12:58:20.407Z", "content": "

The submission of a single checkpoint is about six hours for me.
\nI tried two checkpoints and got a timeout.
\nDid I miss something?

", "rawMarkdown": "The submission of a single checkpoint is about six hours for me.\nI tried two checkpoints and got a timeout.\nDid I miss something?" } ] } ] }, { "id": 3188027, "postDate": "2025-04-26T23:38:01.007Z", "content": "

Hi everyone! Thanks for the insights
\nI'm currently training a solo YOLOv10s model. My single checkpoint performance is decent, but I'm thinking about applying simple ensembling — like averaging predictions from the top saved checkpoints (not multiple models yet). Does anyone have advice whether checkpoint ensembling helped significantly? Or is model diversity (like combining yolov10s + yolov10n) more important in this case? Thanks

", "rawMarkdown": "Hi everyone! Thanks for the insights \nI'm currently training a solo YOLOv10s model. My single checkpoint performance is decent, but I'm thinking about applying simple ensembling — like averaging predictions from the top saved checkpoints (not multiple models yet). Does anyone have advice whether checkpoint ensembling helped significantly? Or is model diversity (like combining yolov10s + yolov10n) more important in this case? Thanks" }, { "id": 3187921, "postDate": "2025-04-26T17:48:12.957Z", "content": "

kaggle notebook only (training and inference): 0.776 single yolov11

\n

i am struggling to reach 0.8

", "rawMarkdown": "kaggle notebook only (training and inference): 0.776 single yolov11\n\ni am struggling to reach 0.8" }, { "id": 3187495, "postDate": "2025-04-26T05:28:25.060Z", "content": "

It more depends on computation power and ensemble for model robustness.

", "rawMarkdown": "It more depends on computation power and ensemble for model robustness." } ], "comments": [ { "id": 3188253, "author_name": "Tom", "author_url": "", "post_date": "2025-04-27T09:45:53.073000", "content": "

GNN solution 0.83+ without setting any confidence threshold, can be better with ensemble. I expect to achieve 0.87+.

\n

Basically it is a 3d approach.

\n

\"\"

", "votes": 20, "replies": [ { "id": 3188314, "author_name": "Andrei Zamfir", "author_url": "", "post_date": "2025-04-27T11:49:05.317000", "content": "

This is one of the more creative solutions of this competition, I really applaud your out of the box approach to this problem. I think I am not the only one who can't wait to read the solutions once this finishes!

", "votes": 2, "replies": [ { "id": 3188584, "author_name": "Cody_Null", "author_url": "", "post_date": "2025-04-27T21:56:37.620000", "content": "

This is super interesting, yes I cant wait to hear more about it too! I tried a 3D approach at the beginning but wasn't working out the best. I think your solution is easily the most interesting and I also have 0 faith in myself to be able to replicate it with GNN so this will be a super cool learning opportunity!

", "votes": 2, "replies": [] }, { "id": 3191999, "author_name": "Tom", "author_url": "", "post_date": "2025-05-02T10:44:54.867000", "content": "

@cody11null 3d approach can work (I use the codebase from @christofhenkel ctyo solution). The biggest problem is 3d unet or ResSegNet have hard time to model a single positive point. I saw people struggling on positive weight or focal loss hyperparamter turning. So we have to put more effort on label development and prediction design. For me, I decide to sparsely predict the motor in 3d structure instead of reading entire volume. The result is pretty good. Faster and stronger than any 3d model which take volume as input.
\nOne thing I can tell you is I assume the subgraphs formed from keypoints in the volume as strongly connected components.

", "votes": 3, "replies": [] } ] }, { "id": 3188885, "author_name": "DennisSakva", "author_url": "", "post_date": "2025-04-28T11:37:40.933000", "content": "

Looks amazing, but I have no idea what's going on here. Will wait for your solution to learn something totally orthogonal to what I know :)

", "votes": 1, "replies": [] }, { "id": 3188980, "author_name": "eikyou", "author_url": "", "post_date": "2025-04-28T15:11:23.507000", "content": "

@tom99763 it is not a simplicial complex, right? U don't have 2-simplices in it, right? Only 0- and 1-simplices?

", "votes": 0, "replies": [ { "id": 3191996, "author_name": "Tom", "author_url": "", "post_date": "2025-05-02T10:41:06.550000", "content": "

@eikyou The graph construction is quite tricky :)

", "votes": 1, "replies": [] } ] }, { "id": 3191752, "author_name": "thisArmin", "author_url": "", "post_date": "2025-05-02T05:30:11.420000", "content": "

are you using only kaggle gpu for training

", "votes": 0, "replies": [ { "id": 3191993, "author_name": "Tom", "author_url": "", "post_date": "2025-05-02T10:39:22.997000", "content": "

@thisarmin I train it locally. But training on kaggle is completely OK cuz my model is really lightweight.

", "votes": 1, "replies": [] } ] }, { "id": 3191840, "author_name": "Tom", "author_url": "", "post_date": "2025-05-02T08:01:24.833000", "content": "

I'll release my github respository after this comp ends.

", "votes": 4, "replies": [] }, { "id": 3202814, "author_name": "Surya_trainer", "author_url": "", "post_date": "2025-05-16T00:35:04.337000", "content": "

what GNN are you using or custom message passing layers??

", "votes": 0, "replies": [ { "id": 3202853, "author_name": "Tom", "author_url": "", "post_date": "2025-05-16T02:55:01.240000", "content": "

If you want to predict unseen graph, you need to use the one who is trained transductively.

", "votes": 0, "replies": [] } ] } ] }, { "id": 3188870, "author_name": "yukiZ", "author_url": "", "post_date": "2025-04-28T11:04:28.183000", "content": "

My best LB 0.818 as of today is from a single YOLO model (\"best.pt\" only)

\n", "votes": 7, "replies": [ { "id": 3188889, "author_name": "fan yanbing", "author_url": "", "post_date": "2025-04-28T11:47:21.870000", "content": "

What does \"no preprocessing\" mean? Just use the given slice?

", "votes": 0, "replies": [] }, { "id": 3189216, "author_name": "MengYe", "author_url": "", "post_date": "2025-04-29T00:38:02.240000", "content": "

can you tell us your f1 and map scores?

", "votes": 0, "replies": [] }, { "id": 3190964, "author_name": "ironrro", "author_url": "", "post_date": "2025-05-01T07:56:45.137000", "content": "

Hi @hideyukizushi what's the version of your YOLO model?

", "votes": 0, "replies": [] }, { "id": 3202150, "author_name": "", "author_url": "", "post_date": "2025-05-14T23:44:19.470000", "content": "

Did u use tta and what were your training augmentations or while creating yolodataset

", "votes": 0, "replies": [] } ] }, { "id": 3193410, "author_name": "Mohamed Eltayeb", "author_url": "", "post_date": "2025-05-04T10:48:11.573000", "content": "

0.778, I focused too much on stability, and I think i finally managed to find a stable pipeline. Changing the confidence threshold from 0.05 up to 0.5 gives lb results between 0.755 to 0.778 which is very stable i guess. Now time to stop stability and improve results :)

", "votes": 3, "replies": [ { "id": 3195064, "author_name": "Mohamed Eltayeb", "author_url": "", "post_date": "2025-05-06T15:32:02.747000", "content": "

Tried a bit of a crazy idea for cv (based on a specific data insight), looks like i managed to get a bit of correlation:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
CV ScoreLB Score
0.02820.578
0.29210.710
0.31910.751
0.42830.766
0.46440.738
\n

Will update if it worked for new subs

", "votes": 3, "replies": [ { "id": 3197149, "author_name": "Mohamed Eltayeb", "author_url": "", "post_date": "2025-05-07T21:04:43.643000", "content": "

Had some bugs with the previous cv scores. Fixed now.

\n

\"\"
\nNot the best, but i can see some good trend.

", "votes": 2, "replies": [] }, { "id": 3197256, "author_name": "Fangsion Fang", "author_url": "", "post_date": "2025-05-08T01:16:08.210000", "content": "

May I ask what metric you are using for your CV score? It seems a bit lower than expected—are you perhaps using the F-beta score or a similar metric?

", "votes": 0, "replies": [] }, { "id": 3197549, "author_name": "Mohamed Eltayeb", "author_url": "", "post_date": "2025-05-08T09:18:06.447000", "content": "

F-beta, but i changed something in the data for this split.

", "votes": 0, "replies": [] } ] } ] }, { "id": 3193476, "author_name": "MathieuD", "author_url": "", "post_date": "2025-05-04T13:05:23.867000", "content": "

LB 0.825 with the pre-trained yolo8m, without external data.

", "votes": 4, "replies": [ { "id": 3196749, "author_name": "Rustam Bazarbayev", "author_url": "", "post_date": "2025-05-07T11:41:52.177000", "content": "

Good! I got 0.693 with pre-trained yolo8m

", "votes": 0, "replies": [] }, { "id": 3198623, "author_name": "UIC Yuheng Ye", "author_url": "", "post_date": "2025-05-09T17:16:45.180000", "content": "

I would like to ask if the pre-trained yolo8m was downloaded from the official website of ultralytics? Or was it obtained from other channels? Is it legal to download it from the official website? Does it violate the relevant regulations of the competition?

", "votes": 0, "replies": [] }, { "id": 3202147, "author_name": "", "author_url": "", "post_date": "2025-05-14T23:38:40.800000", "content": "

What augmentations did u use for training?? Also did u use tta ??

", "votes": 0, "replies": [] } ] }, { "id": 3187609, "author_name": "Yann Majewski", "author_url": "", "post_date": "2025-04-26T09:31:30.463000", "content": "

Mine is 0.831 lb

", "votes": 3, "replies": [ { "id": 3193907, "author_name": "MathieuD", "author_url": "", "post_date": "2025-05-05T07:32:22.090000", "content": "

Do you use Yolo?

", "votes": 0, "replies": [] } ] }, { "id": 3190199, "author_name": "eikyou", "author_url": "", "post_date": "2025-04-30T09:29:45.633000", "content": "

0.794 lb model, yolo11l, only positive samples

", "votes": 1, "replies": [ { "id": 3190499, "author_name": "MengYe", "author_url": "", "post_date": "2025-04-30T18:27:46.063000", "content": "

what is your batch size? i am using Kaggle environment, 8 is the best I can do with yolov11L.

", "votes": 0, "replies": [ { "id": 3190543, "author_name": "eikyou", "author_url": "", "post_date": "2025-04-30T19:42:50.520000", "content": "

@konohayui batch size 16

", "votes": 1, "replies": [] } ] }, { "id": 3192669, "author_name": "Vasilis", "author_url": "", "post_date": "2025-05-03T07:54:18.160000", "content": "

Hey, what do you mean only positive samples?

", "votes": 0, "replies": [ { "id": 3195105, "author_name": "eikyou", "author_url": "", "post_date": "2025-05-06T16:31:00.503000", "content": "

@vasileioscharatsidis i use only slices, where motors are present

", "votes": 0, "replies": [] } ] }, { "id": 3198625, "author_name": "UIC Yuheng Ye", "author_url": "", "post_date": "2025-05-09T17:21:03.083000", "content": "

I would like to ask whether the yolo model was downloaded from the official website of Ultralytics? Or was it obtained from other channels? Is it legal to download it from the official website? Does it violate the relevant regulations of the competition?

", "votes": 0, "replies": [] }, { "id": 3202151, "author_name": "", "author_url": "", "post_date": "2025-05-14T23:45:12.620000", "content": "

What were your training epochs and training time (which gpu)

", "votes": 0, "replies": [] } ] }, { "id": 3190125, "author_name": "I2nfinit3y", "author_url": "", "post_date": "2025-04-30T06:32:55.957000", "content": "

0.785lb RTDETR Model . Trained on images with only one motor and negative samples.

", "votes": 1, "replies": [] }, { "id": 3187498, "author_name": "DennisSakva", "author_url": "", "post_date": "2025-04-26T05:36:31.903000", "content": "

Hi, @cody11null
\nWhat do you mean by solo model? A single checkpoint? Average of several checkpoints from a single fold? Average of several folds of a single model?

", "votes": 1, "replies": [ { "id": 3187509, "author_name": "Cody_Null", "author_url": "", "post_date": "2025-04-26T06:07:03.023000", "content": "

I was meaning single checkpoints but I guess both are interesting!

", "votes": 0, "replies": [ { "id": 3187570, "author_name": "DennisSakva", "author_url": "", "post_date": "2025-04-26T08:11:55.580000", "content": "

I didn't do single chkp inference. My score is from a single fold multiple checkpoints. 0.818

", "votes": 1, "replies": [] }, { "id": 3187776, "author_name": "Cody_Null", "author_url": "", "post_date": "2025-04-26T14:28:34.227000", "content": "

Are you just doing simple ensemble of averaging the last 3 sets of weights or something then or are you doing multiple full trains?

", "votes": 0, "replies": [] }, { "id": 3187810, "author_name": "DennisSakva", "author_url": "", "post_date": "2025-04-26T15:02:43.827000", "content": "

Just simple average of predicts from three best checkpoints from the sinle train run

", "votes": 4, "replies": [] }, { "id": 3190302, "author_name": "MathieuD", "author_url": "", "post_date": "2025-04-30T12:58:20.407000", "content": "

The submission of a single checkpoint is about six hours for me.
\nI tried two checkpoints and got a timeout.
\nDid I miss something?

", "votes": 0, "replies": [] } ] } ] }, { "id": 3188027, "author_name": "Fae Gaze", "author_url": "", "post_date": "2025-04-26T23:38:01.007000", "content": "

Hi everyone! Thanks for the insights
\nI'm currently training a solo YOLOv10s model. My single checkpoint performance is decent, but I'm thinking about applying simple ensembling — like averaging predictions from the top saved checkpoints (not multiple models yet). Does anyone have advice whether checkpoint ensembling helped significantly? Or is model diversity (like combining yolov10s + yolov10n) more important in this case? Thanks

", "votes": 0, "replies": [] }, { "id": 3187921, "author_name": "MengYe", "author_url": "", "post_date": "2025-04-26T17:48:12.957000", "content": "

kaggle notebook only (training and inference): 0.776 single yolov11

\n

i am struggling to reach 0.8

", "votes": 0, "replies": [] }, { "id": 3187495, "author_name": "dragon zhang", "author_url": "", "post_date": "2025-04-26T05:28:25.060000", "content": "

It more depends on computation power and ensemble for model robustness.

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3188253": "GNN solution 0.83+ without setting any confidence threshold, can be better with ensemble. I expect to achieve 0.87+. \n\nBasically it is a 3d approach.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4310004%2Fd9df241f2a0e54edf238323a962033cd%2Fyolo_pred_structure.png?generation=1745747696576552&alt=media)", "3187436": "Curious how people’s solo models are doing. Currently most our best solo models are at around .8 but we expect them to be much higher in about a week. ", "3188870": "My best **LB 0.818** as of today is from a single YOLO model (**\"best.pt\" only**)\n* No pre-processing\n* No post-processing\n* No external data used", "3193410": "0.778, I focused too much on stability, and I think i finally managed to find a stable pipeline. Changing the confidence threshold from 0.05 up to 0.5 gives lb results between 0.755 to 0.778 which is very stable i guess. Now time to stop stability and improve results :)", "3193476": "LB 0.825 with the pre-trained yolo8m, without external data.", "3187609": "Mine is 0.831 lb", "3190199": "0.794 lb model, yolo11l, only positive samples\n", "3190125": "0.785lb RTDETR Model . Trained on images with only one motor and negative samples.", "3187498": "Hi, @cody11null \nWhat do you mean by solo model? A single checkpoint? Average of several checkpoints from a single fold? Average of several folds of a single model?", "3188027": "Hi everyone! Thanks for the insights \nI'm currently training a solo YOLOv10s model. My single checkpoint performance is decent, but I'm thinking about applying simple ensembling — like averaging predictions from the top saved checkpoints (not multiple models yet). Does anyone have advice whether checkpoint ensembling helped significantly? Or is model diversity (like combining yolov10s + yolov10n) more important in this case? Thanks", "3187921": "kaggle notebook only (training and inference): 0.776 single yolov11\n\ni am struggling to reach 0.8", "3187495": "It more depends on computation power and ensemble for model robustness." } }