I wander how long can we go with just one model without any ensemble method.
\nfor me:
\n| method | \nLocal | \nLB | \nTTA | \n
|---|---|---|---|
| MMDetection Swin Transformer | \n0.4747 | \n0.502 | \nNo | \n
| MMDetection HRNet | \n0.5013 | \n0.555 | \nNo | \n
| PyTorch FasterRCNN | \n- | \n0.469 | \nNo | \n
| YOLOv5 | \n- | \n- | \n- | \n
| YOLOx | \n- | \n- | \n- | \n
The local score is obtained by one model trained on video_0 and video_2, evaluate on video_1.
\nThe LB score is obtained by one model that trained on the whole datasets.
I wander how long can we go with just one model without any ensemble method.
\nfor me:
\n| method | \nLocal | \nLB | \nTTA | \n
|---|---|---|---|
| MMDetection Swin Transformer | \n0.4747 | \n0.502 | \nNo | \n
| MMDetection HRNet | \n0.5013 | \n0.555 | \nNo | \n
| PyTorch FasterRCNN | \n- | \n0.469 | \nNo | \n
| YOLOv5 | \n- | \n- | \n- | \n
| YOLOx | \n- | \n- | \n- | \n
The local score is obtained by one model trained on video_0 and video_2, evaluate on video_1.
\nThe LB score is obtained by one model that trained on the whole datasets.
Yolov5 1 fold :
\nCV: 0.62
\nLB: 0.61
That's honestly amazing. YoloX is said to perform better than Yolov5 on COCO, and yet I can't get YoloX to even come close to your score.
", "rawMarkdown": "That's honestly amazing. YoloX is said to perform better than Yolov5 on COCO, and yet I can't get YoloX to even come close to your score. ", "votes": 2 }, { "id": 1638637, "postDate": "2022-01-04T22:18:36.927Z", "content": "Thanks for sharing, that is absolutely inspiring
", "rawMarkdown": "Thanks for sharing, that is absolutely inspiring", "votes": 1 }, { "id": 1639131, "postDate": "2022-01-05T12:04:38.830Z", "content": "@matthieuplante which fold and split are you using?
", "rawMarkdown": "@matthieuplante which fold and split are you using?", "votes": 5 }, { "id": 1642378, "postDate": "2022-01-08T10:05:25.320Z", "content": "Are you sure there was no leak while cross validation? YOLO is typically giving lower CV than Public LB
", "rawMarkdown": "Are you sure there was no leak while cross validation? YOLO is typically giving lower CV than Public LB " }, { "id": 1645141, "postDate": "2022-01-10T18:06:06.970Z", "content": "@matthieuplante amazing job! is the CV score calculated with all images (labeled + unlabeled) or just unlabeled ?
", "rawMarkdown": "@matthieuplante amazing job! is the CV score calculated with all images (labeled + unlabeled) or just unlabeled ? " } ] }, { "id": 1642788, "postDate": "2022-01-08T17:05:25.547Z", "content": "YoloX 1 fold (5 fold split)
\nCV - 0.520
\nLB - 0.556
\nLB with tracking - 0.563
This is with yolox default Hyperparmeter?
\nWhat img size you used while model training?
@anshulkhadse No, this is with modified hyperparameters, optimizer and lr scheduler.
\nImage size is 960x1280.
@yovinyahathugoda would u like to team up?
", "rawMarkdown": "@yovinyahathugoda would u like to team up?" }, { "id": 1643107, "postDate": "2022-01-09T04:36:36.407Z", "content": "\n1) Which validation strategy do you use - Subseq, Sequence Group fold, video split?
\n2) Do you use background images in your training and validation set?
@anshulkhadse replied to your team up request email
", "rawMarkdown": "@anshulkhadse replied to your team up request email", "votes": 1 }, { "id": 1643405, "postDate": "2022-01-09T11:18:02.410Z", "content": "@garvitgarg train/validation split is based on subsequences. I have not used background images yet
", "rawMarkdown": "@garvitgarg train/validation split is based on subsequences. I have not used background images yet" }, { "id": 1643469, "postDate": "2022-01-09T12:39:46.450Z", "content": "@yovinyahathugoda which fold did u use? and splitting by sequence_id did u do u from urself or use from this link https://www.kaggle.com/julian3833/reef-a-cv-strategy-subsequences?scriptVersionId=80623179 CV fold?
", "rawMarkdown": "@yovinyahathugoda which fold did u use? and splitting by sequence_id did u do u from urself or use from this link [https://www.kaggle.com/julian3833/reef-a-cv-strategy-subsequences?scriptVersionId=80623179](url) CV fold?" }, { "id": 1643481, "postDate": "2022-01-09T12:55:39.460Z", "content": "@anshulkhadse the subsequence ids were from the notebook in the link you posted, but I split the data later in the training notebook separately. I only used fold 0 for validation for now.
", "rawMarkdown": "@anshulkhadse the subsequence ids were from the notebook in the link you posted, but I split the data later in the training notebook separately. I only used fold 0 for validation for now.", "votes": 1 }, { "id": 1643560, "postDate": "2022-01-09T14:30:39.877Z", "content": "@yovinyahathugoda Understand ,thanks for sharing
", "rawMarkdown": "@yovinyahathugoda Understand ,thanks for sharing" }, { "id": 1644195, "postDate": "2022-01-10T06:02:06.840Z", "content": "@yovinyahathugoda In CV, What metrics do you use f2score? If you don't mind, can you share code snippet for f2 score? My CV/LB gap is so high. 😄
", "rawMarkdown": "@yovinyahathugoda In CV, What metrics do you use f2score? If you don't mind, can you share code snippet for f2 score? My CV/LB gap is so high. 😄" }, { "id": 1645210, "postDate": "2022-01-10T18:51:24.280Z", "content": "@tanjirooo I use f2 metric for validation at the end of training. I used the code in the discussion here
", "rawMarkdown": "@tanjirooo I use f2 metric for validation at the end of training. I used the code in the discussion [here](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290757#1595988)", "votes": 1 } ] }, { "id": 1638002, "postDate": "2022-01-04T11:12:34.303Z", "content": "The maximum LB for my single model is 0.547 (Yolov5, No ensemble No knowledge distillation, No TTA).
\nBut there is no certainty that this is compatible with Private LB.
use TTA actually given lower LB score
", "rawMarkdown": "use TTA actually given lower LB score", "votes": 1 }, { "id": 1638344, "postDate": "2022-01-04T16:27:50.510Z", "content": "me too.
\nI'm wondering why TTA doesn't work.
@syurenuko, which folding method and split are you using?
", "rawMarkdown": "@syurenuko, which folding method and split are you using?", "votes": 1 }, { "id": 1639425, "postDate": "2022-01-05T16:12:24.463Z", "content": "@garvitgarg
\nI used GroupKFold (n = 5, groups = sequence).
which split did you used to get that
", "rawMarkdown": "which split did you used to get that", "votes": 1 }, { "id": 1642593, "postDate": "2022-01-08T14:13:28.307Z", "content": "@dineshmanikanta
\nThere are 5 folds from 0 to 4.
\nUse fold = 4 for validation.
@shivanigarg32
\nbackground images are not used.
@syurenuko, can tell which hyperparameter file u used from the available yolov5 repo in the data folder?
", "rawMarkdown": "@syurenuko, can tell which hyperparameter file u used from the available yolov5 repo in the data folder?" } ] }, { "id": 1659336, "postDate": "2022-01-21T18:02:15.103Z", "content": "FYI, there is a transformer version of HRNet: https://github.com/HRNet/HRFormer
", "rawMarkdown": "FYI, there is a transformer version of HRNet: https://github.com/HRNet/HRFormer" }, { "id": 1645776, "postDate": "2022-01-11T08:05:32.190Z", "content": "Hi, @snaker . What do you mean for the \"knowledge distillation\" in this problem?
", "rawMarkdown": "Hi, @snaker . What do you mean for the \"knowledge distillation\" in this problem?", "replies": [ { "id": 1645795, "postDate": "2022-01-11T08:19:23.263Z", "content": "Not much we can do about this cause we are lack of test set, i just mention it because it's still a ensemling method.
", "rawMarkdown": "Not much we can do about this cause we are lack of test set, i just mention it because it's still a ensemling method." } ] }, { "id": 1644022, "postDate": "2022-01-10T01:12:05.210Z", "content": "cas r50 1 fold(5 fold split)
\nCV - 66.0
\nLB - 0.579
\nno tracking no tta
is this using mmdet cascade or detectron2?
", "rawMarkdown": "is this using mmdet cascade or detectron2?\n" } ] }, { "id": 1641440, "postDate": "2022-01-07T12:51:15.330Z", "content": "model: yolov5-l
\nsplit: video_id
\nCV: 0.642
\nLB: 0.522
Amazing results for hrnet, which backbone did you use, HR-W18, HR-W32 or W40?
\nFor mmdetection, I only get 0.47 in LB with Faster R-CNN and 0.46 with Cascade R-CNN.
HR-W18 ，the smallest one
", "rawMarkdown": "HR-W18 ，the smallest one" }, { "id": 1641689, "postDate": "2022-01-07T16:28:54.900Z", "content": "Do you use single model cascade HR-W18 to get highest score? Thanks! Now I use yolov5, but the result is unstable in both CV and LB.
", "rawMarkdown": "Do you use single model cascade HR-W18 to get highest score? Thanks! Now I use yolov5, but the result is unstable in both CV and LB." }, { "id": 1642089, "postDate": "2022-01-08T03:29:05.903Z", "content": "And do you use F2 metric for Local ?
", "rawMarkdown": "And do you use F2 metric for Local ?" }, { "id": 1643138, "postDate": "2022-01-09T05:13:31.713Z", "content": "yes, all the metric is F2, not mAP
", "rawMarkdown": "yes, all the metric is F2, not mAP" }, { "id": 1643985, "postDate": "2022-01-09T23:15:06.247Z", "content": "did you ever encounter loss_rpn_cls: nan, loss_rpn_bbox: nan, loss_cls: nan, acc: 100.0000, loss_bbox: nan, loss: nan while setting up your training?
figured it out, torch 1.9.0 is REQUIRED anything higher will give issues during training ig
", "rawMarkdown": "figured it out, torch 1.9.0 is REQUIRED anything higher will give issues during training ig" } ] }, { "id": 1639893, "postDate": "2022-01-06T03:10:17.090Z", "content": "Hi snaker, As you said you finally trained your model on the whole datasets, I wonder how you select the conf_threshold and nms_threshold. Thankyou.
", "rawMarkdown": "Hi snaker, As you said you finally trained your model on the whole datasets, I wonder how you select the conf_threshold and nms_threshold. Thankyou.", "replies": [ { "id": 1640037, "postDate": "2022-01-06T06:11:50.700Z", "content": "The dataset is hard to split, so i just split it by video_id, video_1 for validation and video0, video2 for training, but this is just for experiments.
\nAfter I get a higher score in local i will use the same training set and strategy for the whole dataset, the thresholds is get from the experiment, which is video1 actually.
\" i just split it by video_id, video_1 for validation and video0, video2 for training,\"
\nThese results may be biased.
\nif you use video_0 for validation, your LB score will be different
yes, i have tested on eighter the three of them as validation set, the gap is large(so is the LB), and the video_1 is the lowest so i chose it finally
", "rawMarkdown": "yes, i have tested on eighter the three of them as validation set, the gap is large(so is the LB), and the video_1 is the lowest so i chose it finally" }, { "id": 1642344, "postDate": "2022-01-08T09:37:07.903Z", "content": "\"the gap is large(so is the LB),\"
\nif i apply different augmentation to different dataset, I get better results. but I am not sure if it is because of the difference in the augmented raining set or the difference in the validation set.
", "rawMarkdown": "\"the gap is large(so is the LB),\"\n\nif i apply different augmentation to different dataset, I get better results. but I am not sure if it is because of the difference in the augmented raining set or the difference in the validation set.", "votes": 1 }, { "id": 1642457, "postDate": "2022-01-08T12:01:52.837Z", "content": "\n\nand the video_1 is the lowest so i chose it finally
\n
I don’t think this is correct strategy. It might cause over fitting to public LB.
", "rawMarkdown": "> and the video_1 is the lowest so i chose it finally\n\nI don’t think this is correct strategy. It might cause over fitting to public LB." }, { "id": 1643139, "postDate": "2022-01-09T05:14:47.627Z", "content": "we need a solid method to evaluate model in local, this is i think the best way to do it. cause the LB and private LB could be a totally another video sequence.
\nI'm not using the split to do model selection BTW, just for local experiment evaluation.
", "rawMarkdown": "we need a solid method to evaluate model in local, this is i think **the best** way to do it. cause the LB and private LB could be a totally another video sequence.\n\nI'm not using the split to do model selection BTW, just for local experiment evaluation." }, { "id": 1644258, "postDate": "2022-01-10T07:07:57.813Z", "content": "@hengck23 yes, same here. I saw your comments in another thread. And I think what you suggest is correct and we are doing it now: \"to find the augmentation that minimize the gap between the 3 videos\".
", "rawMarkdown": "@hengck23 yes, same here. I saw your comments in another thread. And I think what you suggest is correct and we are doing it now: \"to find the augmentation that minimize the gap between the 3 videos\"." } ] } ], "comments": [ { "id": 1638333, "author_name": "Matthieu Planté", "author_url": "", "post_date": "2022-01-04T16:16:52.753000", "content": "Yolov5 1 fold :
\nCV: 0.62
\nLB: 0.61
That's honestly amazing. YoloX is said to perform better than Yolov5 on COCO, and yet I can't get YoloX to even come close to your score.
", "votes": 2, "replies": [] }, { "id": 1638637, "author_name": "Chenglu", "author_url": "", "post_date": "2022-01-04T22:18:36.927000", "content": "Thanks for sharing, that is absolutely inspiring
", "votes": 1, "replies": [] }, { "id": 1639131, "author_name": "Awsaf", "author_url": "", "post_date": "2022-01-05T12:04:38.830000", "content": "@matthieuplante which fold and split are you using?
", "votes": 5, "replies": [] }, { "id": 1642378, "author_name": "Hannah B", "author_url": "", "post_date": "2022-01-08T10:05:25.320000", "content": "Are you sure there was no leak while cross validation? YOLO is typically giving lower CV than Public LB
", "votes": 0, "replies": [] }, { "id": 1645141, "author_name": "Ioannis M", "author_url": "", "post_date": "2022-01-10T18:06:06.970000", "content": "@matthieuplante amazing job! is the CV score calculated with all images (labeled + unlabeled) or just unlabeled ?
", "votes": 0, "replies": [] } ] }, { "id": 1642788, "author_name": "Yovin Yahathugoda", "author_url": "", "post_date": "2022-01-08T17:05:25.547000", "content": "YoloX 1 fold (5 fold split)
\nCV - 0.520
\nLB - 0.556
\nLB with tracking - 0.563
This is with yolox default Hyperparmeter?
\nWhat img size you used while model training?
@anshulkhadse No, this is with modified hyperparameters, optimizer and lr scheduler.
\nImage size is 960x1280.
@yovinyahathugoda would u like to team up?
", "votes": 0, "replies": [] }, { "id": 1643107, "author_name": "Garvit Garg", "author_url": "", "post_date": "2022-01-09T04:36:36.407000", "content": "\n1) Which validation strategy do you use - Subseq, Sequence Group fold, video split?
\n2) Do you use background images in your training and validation set?
@anshulkhadse replied to your team up request email
", "votes": 1, "replies": [] }, { "id": 1643405, "author_name": "Yovin Yahathugoda", "author_url": "", "post_date": "2022-01-09T11:18:02.410000", "content": "@garvitgarg train/validation split is based on subsequences. I have not used background images yet
", "votes": 0, "replies": [] }, { "id": 1643469, "author_name": "Anshul Khadse", "author_url": "", "post_date": "2022-01-09T12:39:46.450000", "content": "@yovinyahathugoda which fold did u use? and splitting by sequence_id did u do u from urself or use from this link https://www.kaggle.com/julian3833/reef-a-cv-strategy-subsequences?scriptVersionId=80623179 CV fold?
", "votes": 0, "replies": [] }, { "id": 1643481, "author_name": "Yovin Yahathugoda", "author_url": "", "post_date": "2022-01-09T12:55:39.460000", "content": "@anshulkhadse the subsequence ids were from the notebook in the link you posted, but I split the data later in the training notebook separately. I only used fold 0 for validation for now.
", "votes": 1, "replies": [] }, { "id": 1643560, "author_name": "Anshul Khadse", "author_url": "", "post_date": "2022-01-09T14:30:39.877000", "content": "@yovinyahathugoda Understand ,thanks for sharing
", "votes": 0, "replies": [] }, { "id": 1644195, "author_name": "Tanjirooo", "author_url": "", "post_date": "2022-01-10T06:02:06.840000", "content": "@yovinyahathugoda In CV, What metrics do you use f2score? If you don't mind, can you share code snippet for f2 score? My CV/LB gap is so high. 😄
", "votes": 0, "replies": [] }, { "id": 1645210, "author_name": "Yovin Yahathugoda", "author_url": "", "post_date": "2022-01-10T18:51:24.280000", "content": "@tanjirooo I use f2 metric for validation at the end of training. I used the code in the discussion here
", "votes": 1, "replies": [] } ] }, { "id": 1638002, "author_name": "syurenuko", "author_url": "", "post_date": "2022-01-04T11:12:34.303000", "content": "The maximum LB for my single model is 0.547 (Yolov5, No ensemble No knowledge distillation, No TTA).
\nBut there is no certainty that this is compatible with Private LB.
use TTA actually given lower LB score
", "votes": 1, "replies": [] }, { "id": 1638344, "author_name": "syurenuko", "author_url": "", "post_date": "2022-01-04T16:27:50.510000", "content": "me too.
\nI'm wondering why TTA doesn't work.
@syurenuko, which folding method and split are you using?
", "votes": 1, "replies": [] }, { "id": 1639425, "author_name": "syurenuko", "author_url": "", "post_date": "2022-01-05T16:12:24.463000", "content": "@garvitgarg
\nI used GroupKFold (n = 5, groups = sequence).
which split did you used to get that
", "votes": 1, "replies": [] }, { "id": 1642593, "author_name": "syurenuko", "author_url": "", "post_date": "2022-01-08T14:13:28.307000", "content": "@dineshmanikanta
\nThere are 5 folds from 0 to 4.
\nUse fold = 4 for validation.
@shivanigarg32
\nbackground images are not used.
@syurenuko, can tell which hyperparameter file u used from the available yolov5 repo in the data folder?
", "votes": 0, "replies": [] } ] }, { "id": 1659336, "author_name": "hengck23", "author_url": "", "post_date": "2022-01-21T18:02:15.103000", "content": "FYI, there is a transformer version of HRNet: https://github.com/HRNet/HRFormer
", "votes": 0, "replies": [] }, { "id": 1645776, "author_name": "Dapeng", "author_url": "", "post_date": "2022-01-11T08:05:32.190000", "content": "Hi, @snaker . What do you mean for the \"knowledge distillation\" in this problem?
", "votes": 0, "replies": [ { "id": 1645795, "author_name": "Chenglu", "author_url": "", "post_date": "2022-01-11T08:19:23.263000", "content": "Not much we can do about this cause we are lack of test set, i just mention it because it's still a ensemling method.
", "votes": 0, "replies": [] } ] }, { "id": 1644022, "author_name": "CarryHJR", "author_url": "", "post_date": "2022-01-10T01:12:05.210000", "content": "cas r50 1 fold(5 fold split)
\nCV - 66.0
\nLB - 0.579
\nno tracking no tta
is this using mmdet cascade or detectron2?
", "votes": 0, "replies": [] } ] }, { "id": 1641440, "author_name": "Garvit Garg", "author_url": "", "post_date": "2022-01-07T12:51:15.330000", "content": "model: yolov5-l
\nsplit: video_id
\nCV: 0.642
\nLB: 0.522
Amazing results for hrnet, which backbone did you use, HR-W18, HR-W32 or W40?
\nFor mmdetection, I only get 0.47 in LB with Faster R-CNN and 0.46 with Cascade R-CNN.
HR-W18 ，the smallest one
", "votes": 0, "replies": [] }, { "id": 1641689, "author_name": "clwclw", "author_url": "", "post_date": "2022-01-07T16:28:54.900000", "content": "Do you use single model cascade HR-W18 to get highest score? Thanks! Now I use yolov5, but the result is unstable in both CV and LB.
", "votes": 0, "replies": [] }, { "id": 1642089, "author_name": "clwclw", "author_url": "", "post_date": "2022-01-08T03:29:05.903000", "content": "And do you use F2 metric for Local ?
", "votes": 0, "replies": [] }, { "id": 1643138, "author_name": "Chenglu", "author_url": "", "post_date": "2022-01-09T05:13:31.713000", "content": "yes, all the metric is F2, not mAP
", "votes": 0, "replies": [] }, { "id": 1643985, "author_name": "Hannah B", "author_url": "", "post_date": "2022-01-09T23:15:06.247000", "content": "did you ever encounter loss_rpn_cls: nan, loss_rpn_bbox: nan, loss_cls: nan, acc: 100.0000, loss_bbox: nan, loss: nan while setting up your training?
figured it out, torch 1.9.0 is REQUIRED anything higher will give issues during training ig
", "votes": 0, "replies": [] } ] }, { "id": 1639893, "author_name": "Kevin", "author_url": "", "post_date": "2022-01-06T03:10:17.090000", "content": "Hi snaker, As you said you finally trained your model on the whole datasets, I wonder how you select the conf_threshold and nms_threshold. Thankyou.
", "votes": 0, "replies": [ { "id": 1640037, "author_name": "Chenglu", "author_url": "", "post_date": "2022-01-06T06:11:50.700000", "content": "The dataset is hard to split, so i just split it by video_id, video_1 for validation and video0, video2 for training, but this is just for experiments.
\nAfter I get a higher score in local i will use the same training set and strategy for the whole dataset, the thresholds is get from the experiment, which is video1 actually.
\" i just split it by video_id, video_1 for validation and video0, video2 for training,\"
\nThese results may be biased.
\nif you use video_0 for validation, your LB score will be different
yes, i have tested on eighter the three of them as validation set, the gap is large(so is the LB), and the video_1 is the lowest so i chose it finally
", "votes": 0, "replies": [] }, { "id": 1642344, "author_name": "hengck23", "author_url": "", "post_date": "2022-01-08T09:37:07.903000", "content": "\"the gap is large(so is the LB),\"
\nif i apply different augmentation to different dataset, I get better results. but I am not sure if it is because of the difference in the augmented raining set or the difference in the validation set.
", "votes": 1, "replies": [] }, { "id": 1642457, "author_name": "Bilzard", "author_url": "", "post_date": "2022-01-08T12:01:52.837000", "content": "\n\nand the video_1 is the lowest so i chose it finally
\n
I don’t think this is correct strategy. It might cause over fitting to public LB.
", "votes": 0, "replies": [] }, { "id": 1643139, "author_name": "Chenglu", "author_url": "", "post_date": "2022-01-09T05:14:47.627000", "content": "we need a solid method to evaluate model in local, this is i think the best way to do it. cause the LB and private LB could be a totally another video sequence.
\nI'm not using the split to do model selection BTW, just for local experiment evaluation.
", "votes": 0, "replies": [] }, { "id": 1644258, "author_name": "Chenglu", "author_url": "", "post_date": "2022-01-10T07:07:57.813000", "content": "@hengck23 yes, same here. I saw your comments in another thread. And I think what you suggest is correct and we are doing it now: \"to find the augmentation that minimize the gap between the 3 videos\".
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1636801": "I wander how long can we go with just one model without any ensemble method.\n\nfor me:\n\n| method | Local | LB | TTA |\n| --- | --- | --- | --- |\n| MMDetection Swin Transformer | 0.4747 | 0.502 | No |\n| MMDetection HRNet | 0.5013 | 0.555 | No |\n| PyTorch FasterRCNN | - | 0.469 | No |\n| YOLOv5 | -| - | - | \n| YOLOx | - | - | - |\n\nThe local score is obtained by one model trained on video_0 and video_2, evaluate on video_1.\nThe LB score is obtained by one model that trained on the whole datasets.\n\n", "1638333": "Yolov5 1 fold :\nCV: 0.62\nLB: 0.61", "1642788": "YoloX 1 fold (5 fold split)\nCV - 0.520\nLB - 0.556\nLB with [tracking](https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539) - 0.563", "1638002": "The maximum LB for my single model is 0.547 (Yolov5, No ensemble No knowledge distillation, No TTA).\nBut there is no certainty that this is compatible with Private LB.", "1659336": "FYI, there is a transformer version of HRNet: https://github.com/HRNet/HRFormer", "1645776": "Hi, @snaker . What do you mean for the \"knowledge distillation\" in this problem?", "1644022": "cas r50 1 fold(5 fold split)\nCV - 66.0\nLB - 0.579\nno tracking no tta", "1641440": "model: yolov5-l\nsplit: video_id\nCV: 0.642\nLB: 0.522", "1639914": "Amazing results for hrnet, which backbone did you use, HR-W18, HR-W32 or W40? \nFor mmdetection, I only get 0.47 in LB with Faster R-CNN and 0.46 with Cascade R-CNN. ", "1639893": "Hi snaker, As you said you finally trained your model on the whole datasets, I wonder how you select the conf_threshold and nms_threshold. Thankyou." } }