{ "id": 159883, "title": "What worked for me", "url": "/competitions/global-wheat-detection/discussion/159883", "author_name": "", "post_date": "2020-06-19T04:49:39.640841800Z", "votes": 85, "comment_count": 34, "views": 0, "content": "

What worked for me (frcnn, centernet)

\n\n\n\n

gives a 0.02 boost. awsome!

\n\n\n\n

horizontal, vertical, rotate.\nboosts +0.01-0.02

\n\n\n\n

what did not work (for me)

\n\n\n\n

Any other ideas? Feel free to share!

", "messages": [ { "id": "892687", "postDate": "06/19/2020 04:49:39", "content": "

What worked for me (frcnn, centernet)

\n\n\n\n

gives a 0.02 boost. awsome!

\n\n\n\n

horizontal, vertical, rotate.\nboosts +0.01-0.02

\n\n\n\n

what did not work (for me)

\n\n\n\n

Any other ideas? Feel free to share!

", "rawMarkdown": "# What worked for me (frcnn, centernet)\n* Normal augmentations\nflip, rotate, adding noise to images..\n\n* Cutmix like augmentations\nhttps://www.kaggle.com/nvnnghia/awesome-augmentation\n\ngives a 0.02 boost. awsome!\n\n* Psuedo labeling\n2 rounds > 1 rounds > none\n\n* Weighted box fusion\nhttps://www.kaggle.com/shonenkov/wbf-approach-for-ensemble\n\n* TTA \nhttps://www.kaggle.com/kyoshioka47/pytorch-centernet-inference-with-tta\nhttps://www.kaggle.com/shonenkov/wbf-over-tta-single-model-efficientdet\n\nhorizontal, vertical, rotate.\nboosts +0.01-0.02\n\n* Ensambling over folds, models\nlittle boost. not significant.\n\n* Training with higher resolution\n1024 > 512\n\n* toggling with thresholds.. (by toggling the score threshold of nms or wbf)\nchanges the LB 0.01. Due to the metrics, this happen :(\n\n## what did not work (for me)\n* multi-scale TTA\n* Larger networks\nnot always better.. res50~101 seemed to be quite good.\n* BiFPNs\nnot much difference\n\nAny other ideas? Feel free to share!", "votes": null }, { "id": "892688", "postDate": "06/19/2020 04:51:35", "content": "

I agree with most of what you posted. What do you mean by \"toggling with thresholds\"

", "rawMarkdown": "I agree with most of what you posted. What do you mean by \"toggling with thresholds\"", "votes": null }, { "id": "892690", "postDate": "06/19/2020 04:52:54", "content": "

by toggling the score threshold of nms or wbf. \nIt's quite sensitive because FP are penalized in this competition.

", "rawMarkdown": "by toggling the score threshold of nms or wbf. \nIt's quite sensitive because FP are penalized in this competition.", "votes": null }, { "id": "892694", "postDate": "06/19/2020 05:04:18", "content": "

Ah I see. Yes, that's a good point. It is also important not to tune them too much to the LB score as you may be overfitting :)

", "rawMarkdown": "Ah I see. Yes, that's a good point. It is also important not to tune them too much to the LB score as you may be overfitting :)", "votes": null }, { "id": "892811", "postDate": "06/19/2020 07:16:32", "content": "

Hi @kyoshioka47! Thanks for your insights. I am quite new to Kaggle, and Psuedo labeling is very new to me. I do understand the basic concept that we predict test images and use those predictions to retrain our model. So my question is what do you mean by 2 rounds > 1 rounds > none ? Do you mean number of epochs? Sorry if this sound stupid.

", "rawMarkdown": "Hi @kyoshioka47! Thanks for your insights. I am quite new to Kaggle, and Psuedo labeling is very new to me. I do understand the basic concept that we predict test images and use those predictions to retrain our model. So my question is what do you mean by `2 rounds > 1 rounds > none` ? Do you mean number of epochs? Sorry if this sound stupid.", "votes": null }, { "id": "892900", "postDate": "06/19/2020 08:41:11", "content": "

1 rounds mean that we label the test set once and train the model.

\n\n

2 rounds mean that we we label the test set once and train the model. THEN, we the psuedo-label-trained model is used to relabel the test set, then trained again.

\n\n

multi-round psuedo labeling was effective in some competitions.

", "rawMarkdown": "1 rounds mean that we label the test set once and train the model.\n\n2 rounds mean that we we label the test set once and train the model. THEN, we the psuedo-label-trained model is used to relabel the test set, then trained again.\n\nmulti-round psuedo labeling was effective in some competitions.", "votes": null }, { "id": "892903", "postDate": "06/19/2020 08:41:51", "content": "

I agree on that. PB is a quite different domain and I should trust my CV on it..

", "rawMarkdown": "I agree on that. PB is a quite different domain and I should trust my CV on it..", "votes": null }, { "id": "894339", "postDate": "06/20/2020 10:53:11", "content": "

Oh got it. Thanks!\n I have seen others retraining with pseudo labels with a very high LR and to me, it seems that it might overfit the model. I know that there is a trade between the computation time and the model performance on pseudo labeling. What are your thoughts on this? What's a good approach to train for pseudo-labels?

", "rawMarkdown": "Oh got it. Thanks!\n I have seen others retraining with pseudo labels with a very high LR and to me, it seems that it might overfit the model. I know that there is a trade between the computation time and the model performance on pseudo labeling. What are your thoughts on this? What's a good approach to train for pseudo-labels?", "votes": null }, { "id": "895350", "postDate": "06/21/2020 09:21:26", "content": "

@kyoshioka47 It would be great if you could also tell me how did you solve the submission scoring error

", "rawMarkdown": "kyoshioka47 It would be great if you could also tell me how did you solve the submission scoring error", "votes": null }, { "id": "896307", "postDate": "06/22/2020 04:19:30", "content": "

I'm not sure what you're doing wrong but the imagesize in the test set includes other than 1024.\nYou need to take care of that.

", "rawMarkdown": "I'm not sure what you're doing wrong but the imagesize in the test set includes other than 1024.\nYou need to take care of that.", "votes": null }, { "id": "897722", "postDate": "06/23/2020 04:01:57", "content": "

okay..thanks for the info

", "rawMarkdown": "okay..thanks for the info", "votes": null }, { "id": "899269", "postDate": "06/24/2020 06:05:01", "content": "

I tried something similar but with resnext101 backbone, single fold, no TTA, no pseudo labeling. Maximum i could get was 0.6957 :/ I hope ensembling over folds and pseudo labeling helps.

", "rawMarkdown": "I tried something similar but with resnext101 backbone, single fold, no TTA, no pseudo labeling. Maximum i could get was 0.6957 :/ I hope ensembling over folds and pseudo labeling helps.", "votes": null }, { "id": "899331", "postDate": "06/24/2020 07:03:28", "content": "

I get \n0.71 wo/TTA\n0.72 w/TTA\n0.74 w/Psuedo Labeling\nThinking that I need to train better base detection model to get 0.75+.

", "rawMarkdown": "I get \n0.71 wo/TTA\n0.72 w/TTA\n0.74 w/Psuedo Labeling\nThinking that I need to train better base detection model to get 0.75+.", "votes": null }, { "id": "899381", "postDate": "06/24/2020 07:46:06", "content": "

Is this single fold results or wbf over folds ? I am trying n region out CV. Currently best LB scores I got was from using only 3 regions. Adding more regions is lowering my LB score.

", "rawMarkdown": "Is this single fold results or wbf over folds ? I am trying n region out CV. Currently best LB scores I got was from using only 3 regions. Adding more regions is lowering my LB score.", "votes": null }, { "id": "899415", "postDate": "06/24/2020 08:18:27", "content": "

almost same.

\n\n

Psuedo labeling did not worked for me :(\nI would retry it later

", "rawMarkdown": "almost same.\n\nPsuedo labeling did not worked for me :(\nI would retry it later", "votes": null }, { "id": "900088", "postDate": "06/24/2020 16:00:42", "content": "

@kyoshioka47 Hi, thanks for sharing. I would like to know how do you keep high resolution(1024) without running out of memory?

", "rawMarkdown": "kyoshioka47 Hi, thanks for sharing. I would like to know how do you keep high resolution(1024) without running out of memory?", "votes": null }, { "id": "900490", "postDate": "06/24/2020 20:48:10", "content": "

Try to use smaller batches

", "rawMarkdown": "Try to use smaller batches", "votes": null }, { "id": "900629", "postDate": "06/25/2020 00:28:54", "content": "

this is single fold.

", "rawMarkdown": "this is single fold.", "votes": null }, { "id": "904009", "postDate": "06/27/2020 09:17:06", "content": "

As @pestipeti said you can use smaller batches or you can use gradient accumalation or NVIDIA apex\nHope this helps

", "rawMarkdown": "As @pestipeti said you can use smaller batches or you can use gradient accumalation or NVIDIA apex\nHope this helps", "votes": null }, { "id": "910494", "postDate": "07/01/2020 07:10:26", "content": "

Set w and h >=10, it improves my LB 1%.

", "rawMarkdown": "Set w and h >=10, it improves my LB 1%.", "votes": null }, { "id": "910701", "postDate": "07/01/2020 10:10:24", "content": "

What's the meaning of w and h?

", "rawMarkdown": "What's the meaning of w and h?", "votes": null }, { "id": "911179", "postDate": "07/01/2020 15:47:38", "content": "

the width and height of bounding boxes. He is suggesting you remove bounding boxes with width or height less than 10 pixels

", "rawMarkdown": "the width and height of bounding boxes. He is suggesting you remove bounding boxes with width or height less than 10 pixels", "votes": null }, { "id": "911203", "postDate": "07/01/2020 15:59:40", "content": "

For me, I tried to remove too huge boxes and boxes under area of 30 and score seemed to fall....\nTried it on effdet though have to try it on yolo.....\nOn which did you try @m1nglei

", "rawMarkdown": "For me, I tried to remove too huge boxes and boxes under area of 30 and score seemed to fall....\nTried it on effdet though have to try it on yolo.....\nOn which did you try @m1nglei", "votes": null }, { "id": "913073", "postDate": "07/03/2020 00:06:08", "content": "

@nitindatta I haven't try it on yolov5, because I am lack of GPU hours. I tried it on DetectoRS.

", "rawMarkdown": "nitindatta I haven't try it on yolov5, because I am lack of GPU hours. I tried it on DetectoRS.", "votes": null }, { "id": "914675", "postDate": "07/04/2020 06:19:45", "content": "

I have tried the following\nEfficientDet D5, multi-scale TTA didnt work, the best ones based on my tests are combinations of flip lr, flip vertical and rotate.

\n\n

And here are some interesting fact, I have also tried yolov5 as many suggested, surprisingly, \nEff-D5 TTA = 0.738 LB\nyolo_v5 original TTA = 0.735 LB \nBoth of them are on the same local cv and the scores are consistent, however, adding suggested multi scale TTA on yolo_v5 boosted LB to 0.74x, but Eff-D5 with same multi scale dropped to 0.730.

", "rawMarkdown": "I have tried the following\nEfficientDet D5, multi-scale TTA didnt work, the best ones based on my tests are combinations of flip lr, flip vertical and rotate. \n\nAnd here are some interesting fact, I have also tried yolov5 as many suggested, surprisingly, \nEff-D5 TTA = 0.738 LB\nyolo_v5 original TTA = 0.735 LB \nBoth of them are on the same local cv and the scores are consistent, however, adding suggested multi scale TTA on yolo_v5 boosted LB to 0.74x, but Eff-D5 with same multi scale dropped to 0.730.", "votes": null }, { "id": "914783", "postDate": "07/04/2020 08:18:06", "content": "

Interesting!

", "rawMarkdown": "Interesting!", "votes": null }, { "id": "914854", "postDate": "07/04/2020 09:44:02", "content": "

I wonder that how many epochs pseudo label should run? Does pseudo label need data augmentation as training?

", "rawMarkdown": "I wonder that how many epochs pseudo label should run? Does pseudo label need data augmentation as training?", "votes": null }, { "id": "916258", "postDate": "07/05/2020 14:11:36", "content": "

Train with 512 size didn't work for me.\nSo sad..

", "rawMarkdown": "Train with 512 size didn't work for me.\nSo sad..", "votes": null }, { "id": "918136", "postDate": "07/07/2020 02:49:04", "content": "

Hi @kyoshioka47 When you refers to \"re-train\", you mean finetuning the model for some epochs, right? I haven't applied such trick. Could you let me know how many epochs are needed? Thanks!

", "rawMarkdown": "Hi @kyoshioka47 When you refers to \"re-train\", you mean finetuning the model for some epochs, right? I haven't applied such trick. Could you let me know how many epochs are needed? Thanks!", "votes": null }, { "id": "918147", "postDate": "07/07/2020 02:59:14", "content": "

Any benchmark you compare it to? Models trained with 1024 images?

", "rawMarkdown": "Any benchmark you compare it to? Models trained with 1024 images?", "votes": null }, { "id": "918149", "postDate": "07/07/2020 02:59:50", "content": "

By the way, are you from China? Do you mind if we discuss the competition vai Wechat?

", "rawMarkdown": "By the way, are you from China? Do you mind if we discuss the competition vai Wechat?", "votes": null }, { "id": "918205", "postDate": "07/07/2020 04:39:00", "content": "

Be mindful, you cannot discuss competition details outside of Kaggle if you are not on the same team...

", "rawMarkdown": "Be mindful, you cannot discuss competition details outside of Kaggle if you are not on the same team...", "votes": null }, { "id": "919694", "postDate": "07/08/2020 04:15:12", "content": "

Hi @matthewmasters, thanks for the reminder. Just out of curiosity to make friends. No intension to break the rule.

", "rawMarkdown": "Hi @matthewmasters, thanks for the reminder. Just out of curiosity to make friends. No intension to break the rule.", "votes": null }, { "id": "953956", "postDate": "08/01/2020 09:14:38", "content": "

Pseudo label really needs strong augmentation.

", "rawMarkdown": "Pseudo label really needs strong augmentation.", "votes": null }, { "id": "953972", "postDate": "08/01/2020 09:30:09", "content": "

Could you please elaborate how you came to that conclusion...\nAlso, you are speaking about Effdet isnt it??

", "rawMarkdown": "Could you please elaborate how you came to that conclusion...\nAlso, you are speaking about Effdet isnt it??", "votes": null } ], "comments": [ { "id": 892688, "author_name": "matthewmasters", "author_url": "", "post_date": "06/19/2020 04:51:35", "content": "

I agree with most of what you posted. What do you mean by \"toggling with thresholds\"

", "votes": null, "replies": [ { "id": 892690, "author_name": "kyoshioka47", "author_url": "", "post_date": "06/19/2020 04:52:54", "content": "

by toggling the score threshold of nms or wbf. \nIt's quite sensitive because FP are penalized in this competition.

", "votes": null, "replies": [] }, { "id": 892694, "author_name": "matthewmasters", "author_url": "", "post_date": "06/19/2020 05:04:18", "content": "

Ah I see. Yes, that's a good point. It is also important not to tune them too much to the LB score as you may be overfitting :)

", "votes": null, "replies": [] }, { "id": 892903, "author_name": "kyoshioka47", "author_url": "", "post_date": "06/19/2020 08:41:51", "content": "

I agree on that. PB is a quite different domain and I should trust my CV on it..

", "votes": null, "replies": [] } ] }, { "id": 892811, "author_name": "bendang", "author_url": "", "post_date": "06/19/2020 07:16:32", "content": "

Hi @kyoshioka47! Thanks for your insights. I am quite new to Kaggle, and Psuedo labeling is very new to me. I do understand the basic concept that we predict test images and use those predictions to retrain our model. So my question is what do you mean by 2 rounds > 1 rounds > none ? Do you mean number of epochs? Sorry if this sound stupid.

", "votes": null, "replies": [ { "id": 892900, "author_name": "kyoshioka47", "author_url": "", "post_date": "06/19/2020 08:41:11", "content": "

1 rounds mean that we label the test set once and train the model.

\n\n

2 rounds mean that we we label the test set once and train the model. THEN, we the psuedo-label-trained model is used to relabel the test set, then trained again.

\n\n

multi-round psuedo labeling was effective in some competitions.

", "votes": null, "replies": [] }, { "id": 894339, "author_name": "bendang", "author_url": "", "post_date": "06/20/2020 10:53:11", "content": "

Oh got it. Thanks!\n I have seen others retraining with pseudo labels with a very high LR and to me, it seems that it might overfit the model. I know that there is a trade between the computation time and the model performance on pseudo labeling. What are your thoughts on this? What's a good approach to train for pseudo-labels?

", "votes": null, "replies": [] }, { "id": 918136, "author_name": "", "author_url": "", "post_date": "07/07/2020 02:49:04", "content": "

Hi @kyoshioka47 When you refers to \"re-train\", you mean finetuning the model for some epochs, right? I haven't applied such trick. Could you let me know how many epochs are needed? Thanks!

", "votes": null, "replies": [] } ] }, { "id": 895350, "author_name": "kalyanpichuka", "author_url": "", "post_date": "06/21/2020 09:21:26", "content": "

@kyoshioka47 It would be great if you could also tell me how did you solve the submission scoring error

", "votes": null, "replies": [ { "id": 896307, "author_name": "kyoshioka47", "author_url": "", "post_date": "06/22/2020 04:19:30", "content": "

I'm not sure what you're doing wrong but the imagesize in the test set includes other than 1024.\nYou need to take care of that.

", "votes": null, "replies": [] }, { "id": 897722, "author_name": "kalyanpichuka", "author_url": "", "post_date": "06/23/2020 04:01:57", "content": "

okay..thanks for the info

", "votes": null, "replies": [] } ] }, { "id": 899269, "author_name": "utsavnandi", "author_url": "", "post_date": "06/24/2020 06:05:01", "content": "

I tried something similar but with resnext101 backbone, single fold, no TTA, no pseudo labeling. Maximum i could get was 0.6957 :/ I hope ensembling over folds and pseudo labeling helps.

", "votes": null, "replies": [ { "id": 899331, "author_name": "kyoshioka47", "author_url": "", "post_date": "06/24/2020 07:03:28", "content": "

I get \n0.71 wo/TTA\n0.72 w/TTA\n0.74 w/Psuedo Labeling\nThinking that I need to train better base detection model to get 0.75+.

", "votes": null, "replies": [] }, { "id": 899381, "author_name": "utsavnandi", "author_url": "", "post_date": "06/24/2020 07:46:06", "content": "

Is this single fold results or wbf over folds ? I am trying n region out CV. Currently best LB scores I got was from using only 3 regions. Adding more regions is lowering my LB score.

", "votes": null, "replies": [] }, { "id": 900629, "author_name": "kyoshioka47", "author_url": "", "post_date": "06/25/2020 00:28:54", "content": "

this is single fold.

", "votes": null, "replies": [] } ] }, { "id": 899415, "author_name": "lisosia", "author_url": "", "post_date": "06/24/2020 08:18:27", "content": "

almost same.

\n\n

Psuedo labeling did not worked for me :(\nI would retry it later

", "votes": null, "replies": [] }, { "id": 900088, "author_name": "xhe071", "author_url": "", "post_date": "06/24/2020 16:00:42", "content": "

@kyoshioka47 Hi, thanks for sharing. I would like to know how do you keep high resolution(1024) without running out of memory?

", "votes": null, "replies": [ { "id": 900490, "author_name": "pestipeti", "author_url": "", "post_date": "06/24/2020 20:48:10", "content": "

Try to use smaller batches

", "votes": null, "replies": [] }, { "id": 904009, "author_name": "nitindatta", "author_url": "", "post_date": "06/27/2020 09:17:06", "content": "

As @pestipeti said you can use smaller batches or you can use gradient accumalation or NVIDIA apex\nHope this helps

", "votes": null, "replies": [] } ] }, { "id": 910494, "author_name": "m1nglei", "author_url": "", "post_date": "07/01/2020 07:10:26", "content": "

Set w and h >=10, it improves my LB 1%.

", "votes": null, "replies": [ { "id": 910701, "author_name": "thefatcat", "author_url": "", "post_date": "07/01/2020 10:10:24", "content": "

What's the meaning of w and h?

", "votes": null, "replies": [] }, { "id": 911179, "author_name": "matthewmasters", "author_url": "", "post_date": "07/01/2020 15:47:38", "content": "

the width and height of bounding boxes. He is suggesting you remove bounding boxes with width or height less than 10 pixels

", "votes": null, "replies": [] }, { "id": 911203, "author_name": "nitindatta", "author_url": "", "post_date": "07/01/2020 15:59:40", "content": "

For me, I tried to remove too huge boxes and boxes under area of 30 and score seemed to fall....\nTried it on effdet though have to try it on yolo.....\nOn which did you try @m1nglei

", "votes": null, "replies": [] } ] }, { "id": 913073, "author_name": "m1nglei", "author_url": "", "post_date": "07/03/2020 00:06:08", "content": "

@nitindatta I haven't try it on yolov5, because I am lack of GPU hours. I tried it on DetectoRS.

", "votes": null, "replies": [] }, { "id": 914675, "author_name": "heye0507", "author_url": "", "post_date": "07/04/2020 06:19:45", "content": "

I have tried the following\nEfficientDet D5, multi-scale TTA didnt work, the best ones based on my tests are combinations of flip lr, flip vertical and rotate.

\n\n

And here are some interesting fact, I have also tried yolov5 as many suggested, surprisingly, \nEff-D5 TTA = 0.738 LB\nyolo_v5 original TTA = 0.735 LB \nBoth of them are on the same local cv and the scores are consistent, however, adding suggested multi scale TTA on yolo_v5 boosted LB to 0.74x, but Eff-D5 with same multi scale dropped to 0.730.

", "votes": null, "replies": [] }, { "id": 914783, "author_name": "rutviksurani", "author_url": "", "post_date": "07/04/2020 08:18:06", "content": "

Interesting!

", "votes": null, "replies": [] }, { "id": 914854, "author_name": "fangcheng", "author_url": "", "post_date": "07/04/2020 09:44:02", "content": "

I wonder that how many epochs pseudo label should run? Does pseudo label need data augmentation as training?

", "votes": null, "replies": [ { "id": 953956, "author_name": "fangcheng", "author_url": "", "post_date": "08/01/2020 09:14:38", "content": "

Pseudo label really needs strong augmentation.

", "votes": null, "replies": [] }, { "id": 953972, "author_name": "nitindatta", "author_url": "", "post_date": "08/01/2020 09:30:09", "content": "

Could you please elaborate how you came to that conclusion...\nAlso, you are speaking about Effdet isnt it??

", "votes": null, "replies": [] } ] }, { "id": 916258, "author_name": "chrisstan", "author_url": "", "post_date": "07/05/2020 14:11:36", "content": "

Train with 512 size didn't work for me.\nSo sad..

", "votes": null, "replies": [ { "id": 918147, "author_name": "", "author_url": "", "post_date": "07/07/2020 02:59:14", "content": "

Any benchmark you compare it to? Models trained with 1024 images?

", "votes": null, "replies": [] }, { "id": 918149, "author_name": "", "author_url": "", "post_date": "07/07/2020 02:59:50", "content": "

By the way, are you from China? Do you mind if we discuss the competition vai Wechat?

", "votes": null, "replies": [] }, { "id": 918205, "author_name": "matthewmasters", "author_url": "", "post_date": "07/07/2020 04:39:00", "content": "

Be mindful, you cannot discuss competition details outside of Kaggle if you are not on the same team...

", "votes": null, "replies": [] }, { "id": 919694, "author_name": "", "author_url": "", "post_date": "07/08/2020 04:15:12", "content": "

Hi @matthewmasters, thanks for the reminder. Just out of curiosity to make friends. No intension to break the rule.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "892687": "# What worked for me (frcnn, centernet)\n* Normal augmentations\nflip, rotate, adding noise to images..\n\n* Cutmix like augmentations\nhttps://www.kaggle.com/nvnnghia/awesome-augmentation\n\ngives a 0.02 boost. awsome!\n\n* Psuedo labeling\n2 rounds > 1 rounds > none\n\n* Weighted box fusion\nhttps://www.kaggle.com/shonenkov/wbf-approach-for-ensemble\n\n* TTA \nhttps://www.kaggle.com/kyoshioka47/pytorch-centernet-inference-with-tta\nhttps://www.kaggle.com/shonenkov/wbf-over-tta-single-model-efficientdet\n\nhorizontal, vertical, rotate.\nboosts +0.01-0.02\n\n* Ensambling over folds, models\nlittle boost. not significant.\n\n* Training with higher resolution\n1024 > 512\n\n* toggling with thresholds.. (by toggling the score threshold of nms or wbf)\nchanges the LB 0.01. Due to the metrics, this happen :(\n\n## what did not work (for me)\n* multi-scale TTA\n* Larger networks\nnot always better.. res50~101 seemed to be quite good.\n* BiFPNs\nnot much difference\n\nAny other ideas? Feel free to share!", "892688": "I agree with most of what you posted. What do you mean by \"toggling with thresholds\"", "892690": "by toggling the score threshold of nms or wbf. \nIt's quite sensitive because FP are penalized in this competition.", "892694": "Ah I see. Yes, that's a good point. It is also important not to tune them too much to the LB score as you may be overfitting :)", "892811": "Hi @kyoshioka47! Thanks for your insights. I am quite new to Kaggle, and Psuedo labeling is very new to me. I do understand the basic concept that we predict test images and use those predictions to retrain our model. So my question is what do you mean by `2 rounds > 1 rounds > none` ? Do you mean number of epochs? Sorry if this sound stupid.", "892900": "1 rounds mean that we label the test set once and train the model.\n\n2 rounds mean that we we label the test set once and train the model. THEN, we the psuedo-label-trained model is used to relabel the test set, then trained again.\n\nmulti-round psuedo labeling was effective in some competitions.", "892903": "I agree on that. PB is a quite different domain and I should trust my CV on it..", "894339": "Oh got it. Thanks!\n I have seen others retraining with pseudo labels with a very high LR and to me, it seems that it might overfit the model. I know that there is a trade between the computation time and the model performance on pseudo labeling. What are your thoughts on this? What's a good approach to train for pseudo-labels?", "895350": "kyoshioka47 It would be great if you could also tell me how did you solve the submission scoring error", "896307": "I'm not sure what you're doing wrong but the imagesize in the test set includes other than 1024.\nYou need to take care of that.", "897722": "okay..thanks for the info", "899269": "I tried something similar but with resnext101 backbone, single fold, no TTA, no pseudo labeling. Maximum i could get was 0.6957 :/ I hope ensembling over folds and pseudo labeling helps.", "899331": "I get \n0.71 wo/TTA\n0.72 w/TTA\n0.74 w/Psuedo Labeling\nThinking that I need to train better base detection model to get 0.75+.", "899381": "Is this single fold results or wbf over folds ? I am trying n region out CV. Currently best LB scores I got was from using only 3 regions. Adding more regions is lowering my LB score.", "899415": "almost same.\n\nPsuedo labeling did not worked for me :(\nI would retry it later", "900088": "kyoshioka47 Hi, thanks for sharing. I would like to know how do you keep high resolution(1024) without running out of memory?", "900490": "Try to use smaller batches", "900629": "this is single fold.", "904009": "As @pestipeti said you can use smaller batches or you can use gradient accumalation or NVIDIA apex\nHope this helps", "910494": "Set w and h >=10, it improves my LB 1%.", "910701": "What's the meaning of w and h?", "911179": "the width and height of bounding boxes. He is suggesting you remove bounding boxes with width or height less than 10 pixels", "911203": "For me, I tried to remove too huge boxes and boxes under area of 30 and score seemed to fall....\nTried it on effdet though have to try it on yolo.....\nOn which did you try @m1nglei", "913073": "nitindatta I haven't try it on yolov5, because I am lack of GPU hours. I tried it on DetectoRS.", "914675": "I have tried the following\nEfficientDet D5, multi-scale TTA didnt work, the best ones based on my tests are combinations of flip lr, flip vertical and rotate. \n\nAnd here are some interesting fact, I have also tried yolov5 as many suggested, surprisingly, \nEff-D5 TTA = 0.738 LB\nyolo_v5 original TTA = 0.735 LB \nBoth of them are on the same local cv and the scores are consistent, however, adding suggested multi scale TTA on yolo_v5 boosted LB to 0.74x, but Eff-D5 with same multi scale dropped to 0.730.", "914783": "Interesting!", "914854": "I wonder that how many epochs pseudo label should run? Does pseudo label need data augmentation as training?", "916258": "Train with 512 size didn't work for me.\nSo sad..", "918136": "Hi @kyoshioka47 When you refers to \"re-train\", you mean finetuning the model for some epochs, right? I haven't applied such trick. Could you let me know how many epochs are needed? Thanks!", "918147": "Any benchmark you compare it to? Models trained with 1024 images?", "918149": "By the way, are you from China? Do you mind if we discuss the competition vai Wechat?", "918205": "Be mindful, you cannot discuss competition details outside of Kaggle if you are not on the same team...", "919694": "Hi @matthewmasters, thanks for the reminder. Just out of curiosity to make friends. No intension to break the rule.", "953956": "Pseudo label really needs strong augmentation.", "953972": "Could you please elaborate how you came to that conclusion...\nAlso, you are speaking about Effdet isnt it??" }, "source": "meta" }