{ "id": 424112, "title": "Experimental results from LB: 0.515", "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/424112", "author_name": "bent1e", "post_date": "2023-07-12T15:40:01.206000", "votes": 15, "comment_count": 33, "views": 0, "content": "

LB: 0.512 based on yolov7

\n

The ones that didn't work for me:

\n

semi-supervised learning
\nMulti-scale inference
\nmerge nms
\nrank function https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/422250

\n

useful:
\nsimple TTA
\nmodel ensemble
\nfind better conf thr (seriously affects inference time)
\nDifferent number of dilate iterations

\n

need to try:
\nremove the edge effects

", "messages": [ { "id": 2342198, "postDate": "2023-07-12T15:40:01.207Z", "content": "

LB: 0.512 based on yolov7

\n

The ones that didn't work for me:

\n

semi-supervised learning
\nMulti-scale inference
\nmerge nms
\nrank function https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/422250

\n

useful:
\nsimple TTA
\nmodel ensemble
\nfind better conf thr (seriously affects inference time)
\nDifferent number of dilate iterations

\n

need to try:
\nremove the edge effects

", "rawMarkdown": "LB: 0.512 based on yolov7\n\nThe ones that didn't work for me:\n\nsemi-supervised learning\nMulti-scale inference\nmerge nms\nrank function https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/422250\n\nuseful:\nsimple TTA\nmodel ensemble\nfind better conf thr (seriously affects inference time)\nDifferent number of dilate iterations\n\nneed to try:\nremove the edge effects\n\n", "votes": 15 }, { "id": 2361675, "postDate": "2023-07-27T14:21:30.113Z", "content": "
    \n
  1. Add the unet++ segmentation header. The specific idea is to additionally train a semantic segmentation network that takes GT masks and boxes as input.
    \nIn the inference phase, the boxes generated by yolo will carry out this segmentation network to generate masks.
    \nSome issues to note: since the resolution of the vessel instances is relatively small, they are resized to a uniform size in the training phase, but a lot of information will be lost
    \nResult: score: 0 (confuses me)
    2. \n
", "rawMarkdown": "2. Add the unet++ segmentation header. The specific idea is to additionally train a semantic segmentation network that takes GT masks and boxes as input.\nIn the inference phase, the boxes generated by yolo will carry out this segmentation network to generate masks.\nSome issues to note: since the resolution of the vessel instances is relatively small, they are resized to a uniform size in the training phase, but a lot of information will be lost\nResult: score: 0 (confuses me)", "votes": 3 }, { "id": 2347879, "postDate": "2023-07-17T09:01:24.413Z", "content": "

pseudo label didn't really help me so much, there is a big difference in dataset1 and dataset3. I'm trying to use some \"normalization\" method to solve it.

", "rawMarkdown": "pseudo label didn't really help me so much, there is a big difference in dataset1 and dataset3. I'm trying to use some \"normalization\" method to solve it.", "votes": 4 }, { "id": 2343075, "postDate": "2023-07-13T11:39:12.063Z", "content": "

TTA not work for me, what the methods you used ?

", "rawMarkdown": "TTA not work for me, what the methods you used ?", "votes": 1, "replies": [ { "id": 2343154, "postDate": "2023-07-13T12:45:23.443Z", "content": "

Flip + 0.8~1.2 scaling, nms uniformly for all boxes generated. but it doesn't work for me, maybe I shouldn't nms the boxes but average the masks

", "rawMarkdown": "Flip + 0.8~1.2 scaling, nms uniformly for all boxes generated. but it doesn't work for me, maybe I shouldn't nms the boxes but average the masks" } ] }, { "id": 2345024, "postDate": "2023-07-15T02:18:00.350Z", "content": "

2, find better conf thr (seriously affects inference time) LB：0.494—>0.496

", "rawMarkdown": "2, find better conf thr (seriously affects inference time) LB：0.494—>0.496", "votes": 2, "replies": [ { "id": 2346485, "postDate": "2023-07-16T09:39:20.300Z", "content": "

Almost at 0.5, well done!

", "rawMarkdown": "Almost at 0.5, well done!" } ] }, { "id": 2343296, "postDate": "2023-07-13T14:53:05.630Z", "content": "

Any attempts to use Yolo-NAS which is supposed to be better performing model?

", "rawMarkdown": "Any attempts to use Yolo-NAS which is supposed to be better performing model?", "votes": 2 }, { "id": 2363926, "postDate": "2023-07-29T03:30:37.787Z", "content": "

Hey @bent1e , good work. What's a good starter notebook to use and what changes is needed for 0.5?

\n

Thanks

", "rawMarkdown": "Hey @bent1e , good work. What's a good starter notebook to use and what changes is needed for 0.5?\n\nThanks" }, { "id": 2361662, "postDate": "2023-07-27T14:13:17.143Z", "content": "

Things to do lately:
\n1, Rank the instances and construct a ranking function.
\nscore= function (area of mask, confidence score of mask, texture of the region where mask is located, average pixel value of the region where mask is located)
\nDuring training, labels are labeled using iou values and regression prediction is performed using MLP.
\nDuring inference phase, the output of yolov7 is ranked
\nResult: significant drop in LB, but I think it is worth investigating, the reason it doesn't work could be that I didn't include this ranking function in the end-to-end training process

", "rawMarkdown": "Things to do lately:\n1, Rank the instances and construct a ranking function.\nscore= function (area of mask, confidence score of mask, texture of the region where mask is located, average pixel value of the region where mask is located)\nDuring training, labels are labeled using iou values and regression prediction is performed using MLP.\nDuring inference phase, the output of yolov7 is ranked\nResult: significant drop in LB, but I think it is worth investigating, the reason it doesn't work could be that I didn't include this ranking function in the end-to-end training process" }, { "id": 2349903, "postDate": "2023-07-18T19:21:44.107Z", "content": "

Hey,
\nhave you given a try to self Meta pseudo labels ? It works well on segmentation comps !

", "rawMarkdown": "Hey, \nhave you given a try to self Meta pseudo labels ? It works well on segmentation comps !" }, { "id": 2345023, "postDate": "2023-07-15T02:17:09.097Z", "content": "

5, Different number of dilate iterations（not work）

", "rawMarkdown": "5, Different number of dilate iterations（not work）", "replies": [ { "id": 2345259, "postDate": "2023-07-15T08:29:17.907Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 2345381, "postDate": "2023-07-15T10:39:23.247Z", "content": "

no,i just changed the iterations

", "rawMarkdown": "no,i just changed the iterations" }, { "id": 2345420, "postDate": "2023-07-15T11:20:28.090Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2346103, "postDate": "2023-07-16T03:11:10.413Z", "content": "

which kernel sizes have you tried?

", "rawMarkdown": "which kernel sizes have you tried?" }, { "id": 2349275, "postDate": "2023-07-18T09:25:17.657Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2350154, "postDate": "2023-07-19T03:38:40.323Z", "content": "

Thank you very much!!!

", "rawMarkdown": "Thank you very much!!!" } ] } ] }, { "id": 2343977, "postDate": "2023-07-14T06:26:16.037Z", "content": "

Thanks for sharing. What's your validation score of each fold?

", "rawMarkdown": "Thanks for sharing. What's your validation score of each fold?", "replies": [ { "id": 2343981, "postDate": "2023-07-14T06:32:32.363Z", "content": "

LB is 0.441, 0.442, 0.443, 0.466, and 0.494, respectively.
\nDoesn't seem to make sense.

", "rawMarkdown": "LB is 0.441, 0.442, 0.443, 0.466, and 0.494, respectively.\nDoesn't seem to make sense.", "replies": [ { "id": 2343997, "postDate": "2023-07-14T06:48:34.663Z", "content": "

Okay but I was asking what was your local validation score? I think yolov7 computes COCO mAP (mean of AP @ [0.5:0.95]).

", "rawMarkdown": "Okay but I was asking what was your **local validation** score? I think yolov7 computes COCO mAP (mean of AP @ [0.5:0.95])." }, { "id": 2344014, "postDate": "2023-07-14T06:55:22.467Z", "content": "

LB is 0.441, 0.442, 0.443, 0.466, and 0.494, respectively.
\nValidation set mAP: 0.373, 0.365, 0.379, 0.369, 0.371
\nBecause of the problem of randomly splitting the dataset, the validation set mAP doesn't have much correlation with the LB, or maybe I didn't split it in the right way

", "rawMarkdown": "LB is 0.441, 0.442, 0.443, 0.466, and 0.494, respectively.\nValidation set mAP: 0.373, 0.365, 0.379, 0.369, 0.371\nBecause of the problem of randomly splitting the dataset, the validation set mAP doesn't have much correlation with the LB, or maybe I didn't split it in the right way", "votes": 2 }, { "id": 2344036, "postDate": "2023-07-14T07:14:43.897Z", "content": "

I'm doing splits by leaving one WSI out so the validation sets are WSI 1, 2, 3 and 4 respectively. I'll try yolov7 and compare it with your scores.

", "rawMarkdown": "I'm doing splits by leaving one WSI out so the validation sets are WSI 1, 2, 3 and 4 respectively. I'll try yolov7 and compare it with your scores." }, { "id": 2344038, "postDate": "2023-07-14T07:23:51.993Z", "content": "

Do you mean you have one WSL in each fold that isn't involved in training?

", "rawMarkdown": "Do you mean you have one WSL in each fold that isn't involved in training?" }, { "id": 2344044, "postDate": "2023-07-14T07:30:08.690Z", "content": "

Yes, folds are created like this.

\n
 fold  (df[].unique()):\n    train = df[] != fold\n    val = df[] == fold\n
", "rawMarkdown": "Yes, folds are created like this.\n```python\nfor fold in sorted(df['source_wsi'].unique()):\n train = df['source_wsi'] != fold\n val = df['source_wsi'] == fold\n```", "votes": 1 }, { "id": 2344078, "postDate": "2023-07-14T08:04:50.883Z", "content": "

Thanks for the advice, I'll give it a try

", "rawMarkdown": "Thanks for the advice, I'll give it a try" }, { "id": 2345026, "postDate": "2023-07-15T02:19:20.520Z", "content": "

Not work for me

", "rawMarkdown": "Not work for me" }, { "id": 2349198, "postDate": "2023-07-18T08:21:40.077Z", "content": "

Thanks for sharing! My map50-95 is stuck at 0.325. If you could give me some hints, I would greatly appreciate it.

", "rawMarkdown": "Thanks for sharing! My map50-95 is stuck at 0.325. If you could give me some hints, I would greatly appreciate it." }, { "id": 2349208, "postDate": "2023-07-18T08:25:52.723Z", "content": "

data augmentation

", "rawMarkdown": "data augmentation" }, { "id": 2349220, "postDate": "2023-07-18T08:35:33.990Z", "content": "

Thanks! I'll try than.

", "rawMarkdown": "Thanks! I'll try than." } ] } ] }, { "id": 2343871, "postDate": "2023-07-14T04:23:31.280Z", "content": "

1, merge nms under single model not work

", "rawMarkdown": "1, merge nms under single model not work" }, { "id": 2342963, "postDate": "2023-07-13T09:49:51.527Z", "content": "

The semi-supervised algorithm doesn't seem to work: i used the model with LB: 0.494 to generate blood vessel labels for dataset3, after that i pre-trained the model on dataset3, the validation set was used with dataset1+dataset2, validation set mAP: 0.160. after pre-training, the model was re-trained with ds1+ds2, in order to make sure that the variables are the same. The hyperparameters of the model are kept the same as LB: 0.494, validation set mAP: 0.390 (validation set mAP for LB: 0.494: 0.371)

\n

The final submission result LB: 0.494 ---> 0.473 This result confuses me

", "rawMarkdown": "The semi-supervised algorithm doesn't seem to work: i used the model with LB: 0.494 to generate blood vessel labels for dataset3, after that i pre-trained the model on dataset3, the validation set was used with dataset1+dataset2, validation set mAP: 0.160. after pre-training, the model was re-trained with ds1+ds2, in order to make sure that the variables are the same. The hyperparameters of the model are kept the same as LB: 0.494, validation set mAP: 0.390 (validation set mAP for LB: 0.494: 0.371)\n\nThe final submission result LB: 0.494 ---> 0.473 This result confuses me", "replies": [ { "id": 2344947, "postDate": "2023-07-14T23:49:39.117Z", "content": "

Did you dilate your pseudo label?
\nDid you apply heavy augmentation on your student model ?

", "rawMarkdown": "Did you dilate your pseudo label?\nDid you apply heavy augmentation on your student model ?", "replies": [ { "id": 2346745, "postDate": "2023-07-16T13:38:57.430Z", "content": "

I didn't do anything with the pseudo label.In the student model, I did not add data augmentation

", "rawMarkdown": "I didn't do anything with the pseudo label.In the student model, I did not add data augmentation" } ] } ] } ], "comments": [ { "id": 2361675, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-27T14:21:30.113000", "content": "
    \n
  1. Add the unet++ segmentation header. The specific idea is to additionally train a semantic segmentation network that takes GT masks and boxes as input.
    \nIn the inference phase, the boxes generated by yolo will carry out this segmentation network to generate masks.
    \nSome issues to note: since the resolution of the vessel instances is relatively small, they are resized to a uniform size in the training phase, but a lot of information will be lost
    \nResult: score: 0 (confuses me)
    2. \n
", "votes": 3, "replies": [] }, { "id": 2347879, "author_name": "Aurora Rabbit", "author_url": "", "post_date": "2023-07-17T09:01:24.413000", "content": "

pseudo label didn't really help me so much, there is a big difference in dataset1 and dataset3. I'm trying to use some \"normalization\" method to solve it.

", "votes": 4, "replies": [] }, { "id": 2343075, "author_name": "LLM_BOT", "author_url": "", "post_date": "2023-07-13T11:39:12.063000", "content": "

TTA not work for me, what the methods you used ?

", "votes": 1, "replies": [ { "id": 2343154, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-13T12:45:23.443000", "content": "

Flip + 0.8~1.2 scaling, nms uniformly for all boxes generated. but it doesn't work for me, maybe I shouldn't nms the boxes but average the masks

", "votes": 0, "replies": [] } ] }, { "id": 2345024, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-15T02:18:00.350000", "content": "

2, find better conf thr (seriously affects inference time) LB：0.494—>0.496

", "votes": 2, "replies": [ { "id": 2346485, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2023-07-16T09:39:20.300000", "content": "

Almost at 0.5, well done!

", "votes": 0, "replies": [] } ] }, { "id": 2343296, "author_name": "C R Suthikshn Kumar", "author_url": "", "post_date": "2023-07-13T14:53:05.630000", "content": "

Any attempts to use Yolo-NAS which is supposed to be better performing model?

", "votes": 2, "replies": [] }, { "id": 2363926, "author_name": "Bo Peng", "author_url": "", "post_date": "2023-07-29T03:30:37.787000", "content": "

Hey @bent1e , good work. What's a good starter notebook to use and what changes is needed for 0.5?

\n

Thanks

", "votes": 0, "replies": [] }, { "id": 2361662, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-27T14:13:17.143000", "content": "

Things to do lately:
\n1, Rank the instances and construct a ranking function.
\nscore= function (area of mask, confidence score of mask, texture of the region where mask is located, average pixel value of the region where mask is located)
\nDuring training, labels are labeled using iou values and regression prediction is performed using MLP.
\nDuring inference phase, the output of yolov7 is ranked
\nResult: significant drop in LB, but I think it is worth investigating, the reason it doesn't work could be that I didn't include this ranking function in the end-to-end training process

", "votes": 0, "replies": [] }, { "id": 2349903, "author_name": "JEANMPIA", "author_url": "", "post_date": "2023-07-18T19:21:44.107000", "content": "

Hey,
\nhave you given a try to self Meta pseudo labels ? It works well on segmentation comps !

", "votes": 0, "replies": [] }, { "id": 2345023, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-15T02:17:09.097000", "content": "

5, Different number of dilate iterations（not work）

", "votes": 0, "replies": [ { "id": 2345259, "author_name": "", "author_url": "", "post_date": "2023-07-15T08:29:17.907000", "content": "", "votes": 0, "replies": [ { "id": 2345381, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-15T10:39:23.247000", "content": "

no,i just changed the iterations

", "votes": 0, "replies": [] }, { "id": 2345420, "author_name": "", "author_url": "", "post_date": "2023-07-15T11:20:28.090000", "content": "", "votes": 0, "replies": [] }, { "id": 2346103, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-16T03:11:10.413000", "content": "

which kernel sizes have you tried?

", "votes": 0, "replies": [] }, { "id": 2349275, "author_name": "", "author_url": "", "post_date": "2023-07-18T09:25:17.657000", "content": "", "votes": 0, "replies": [] }, { "id": 2350154, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-19T03:38:40.323000", "content": "

Thank you very much!!!

", "votes": 0, "replies": [] } ] } ] }, { "id": 2343977, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-07-14T06:26:16.037000", "content": "

Thanks for sharing. What's your validation score of each fold?

", "votes": 0, "replies": [ { "id": 2343981, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-14T06:32:32.363000", "content": "

LB is 0.441, 0.442, 0.443, 0.466, and 0.494, respectively.
\nDoesn't seem to make sense.

", "votes": 0, "replies": [ { "id": 2343997, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-07-14T06:48:34.663000", "content": "

Okay but I was asking what was your local validation score? I think yolov7 computes COCO mAP (mean of AP @ [0.5:0.95]).

", "votes": 0, "replies": [] }, { "id": 2344014, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-14T06:55:22.467000", "content": "

LB is 0.441, 0.442, 0.443, 0.466, and 0.494, respectively.
\nValidation set mAP: 0.373, 0.365, 0.379, 0.369, 0.371
\nBecause of the problem of randomly splitting the dataset, the validation set mAP doesn't have much correlation with the LB, or maybe I didn't split it in the right way

", "votes": 2, "replies": [] }, { "id": 2344036, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-07-14T07:14:43.897000", "content": "

I'm doing splits by leaving one WSI out so the validation sets are WSI 1, 2, 3 and 4 respectively. I'll try yolov7 and compare it with your scores.

", "votes": 0, "replies": [] }, { "id": 2344038, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-14T07:23:51.993000", "content": "

Do you mean you have one WSL in each fold that isn't involved in training?

", "votes": 0, "replies": [] }, { "id": 2344044, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-07-14T07:30:08.690000", "content": "

Yes, folds are created like this.

\n
 fold  (df[].unique()):\n    train = df[] != fold\n    val = df[] == fold\n
", "votes": 1, "replies": [] }, { "id": 2344078, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-14T08:04:50.883000", "content": "

Thanks for the advice, I'll give it a try

", "votes": 0, "replies": [] }, { "id": 2345026, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-15T02:19:20.520000", "content": "

Not work for me

", "votes": 0, "replies": [] }, { "id": 2349198, "author_name": "FreddyHu", "author_url": "", "post_date": "2023-07-18T08:21:40.077000", "content": "

Thanks for sharing! My map50-95 is stuck at 0.325. If you could give me some hints, I would greatly appreciate it.

", "votes": 0, "replies": [] }, { "id": 2349208, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-18T08:25:52.723000", "content": "

data augmentation

", "votes": 0, "replies": [] }, { "id": 2349220, "author_name": "FreddyHu", "author_url": "", "post_date": "2023-07-18T08:35:33.990000", "content": "

Thanks! I'll try than.

", "votes": 0, "replies": [] } ] } ] }, { "id": 2343871, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-14T04:23:31.280000", "content": "

1, merge nms under single model not work

", "votes": 0, "replies": [] }, { "id": 2342963, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-13T09:49:51.527000", "content": "

The semi-supervised algorithm doesn't seem to work: i used the model with LB: 0.494 to generate blood vessel labels for dataset3, after that i pre-trained the model on dataset3, the validation set was used with dataset1+dataset2, validation set mAP: 0.160. after pre-training, the model was re-trained with ds1+ds2, in order to make sure that the variables are the same. The hyperparameters of the model are kept the same as LB: 0.494, validation set mAP: 0.390 (validation set mAP for LB: 0.494: 0.371)

\n

The final submission result LB: 0.494 ---> 0.473 This result confuses me

", "votes": 0, "replies": [ { "id": 2344947, "author_name": "atom1231", "author_url": "", "post_date": "2023-07-14T23:49:39.117000", "content": "

Did you dilate your pseudo label?
\nDid you apply heavy augmentation on your student model ?

", "votes": 0, "replies": [ { "id": 2346745, "author_name": "bent1e", "author_url": "", "post_date": "2023-07-16T13:38:57.430000", "content": "

I didn't do anything with the pseudo label.In the student model, I did not add data augmentation

", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2342198": "LB: 0.512 based on yolov7\n\nThe ones that didn't work for me:\n\nsemi-supervised learning\nMulti-scale inference\nmerge nms\nrank function https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/422250\n\nuseful:\nsimple TTA\nmodel ensemble\nfind better conf thr (seriously affects inference time)\nDifferent number of dilate iterations\n\nneed to try:\nremove the edge effects\n\n", "2361675": "2. Add the unet++ segmentation header. The specific idea is to additionally train a semantic segmentation network that takes GT masks and boxes as input.\nIn the inference phase, the boxes generated by yolo will carry out this segmentation network to generate masks.\nSome issues to note: since the resolution of the vessel instances is relatively small, they are resized to a uniform size in the training phase, but a lot of information will be lost\nResult: score: 0 (confuses me)", "2347879": "pseudo label didn't really help me so much, there is a big difference in dataset1 and dataset3. I'm trying to use some \"normalization\" method to solve it.", "2343075": "TTA not work for me, what the methods you used ?", "2345024": "2, find better conf thr (seriously affects inference time) LB：0.494—>0.496", "2343296": "Any attempts to use Yolo-NAS which is supposed to be better performing model?", "2363926": "Hey @bent1e , good work. What's a good starter notebook to use and what changes is needed for 0.5?\n\nThanks", "2361662": "Things to do lately:\n1, Rank the instances and construct a ranking function.\nscore= function (area of mask, confidence score of mask, texture of the region where mask is located, average pixel value of the region where mask is located)\nDuring training, labels are labeled using iou values and regression prediction is performed using MLP.\nDuring inference phase, the output of yolov7 is ranked\nResult: significant drop in LB, but I think it is worth investigating, the reason it doesn't work could be that I didn't include this ranking function in the end-to-end training process", "2349903": "Hey, \nhave you given a try to self Meta pseudo labels ? It works well on segmentation comps !", "2345023": "5, Different number of dilate iterations（not work）", "2343977": "Thanks for sharing. What's your validation score of each fold?", "2343871": "1, merge nms under single model not work", "2342963": "The semi-supervised algorithm doesn't seem to work: i used the model with LB: 0.494 to generate blood vessel labels for dataset3, after that i pre-trained the model on dataset3, the validation set was used with dataset1+dataset2, validation set mAP: 0.160. after pre-training, the model was re-trained with ds1+ds2, in order to make sure that the variables are the same. The hyperparameters of the model are kept the same as LB: 0.494, validation set mAP: 0.390 (validation set mAP for LB: 0.494: 0.371)\n\nThe final submission result LB: 0.494 ---> 0.473 This result confuses me" } }