{ "id": 299409, "title": "Yolov5 model performance is not up to mark.", "url": "/competitions/tensorflow-great-barrier-reef/discussion/299409", "author_name": "", "post_date": "2022-01-07T18:32:34.139991Z", "votes": 2, "comment_count": 15, "views": 0, "content": "

I have been trying to train yolov5 model for the challenge Tensorflow Great Barrier Reef,
\nI saw in discussion that 0.48 was yolov5 got for this competition however no matter what I do I am not able to cross 0.425 barrier. So far I have:

\n
    \n
  1. Trained model on preprocessed Underwater images using tf addons.
    2. \n
  2. Implemented Tracking which actually increased score a bit. [https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539]
    3. \n
  3. Trained yoloL6 model.
    \n0.42 is what I got for yoloL6 with Tracking. Tracking clearly improves the result as is evident from the video I am attaching here.
    \nhttps://drive.google.com/file/d/11gp6IDav0X7fNxBYG3C0uL6qOZ6pj0eD/view?usp=sharing
    4. \n
\n

I just want to know what more can be done to improve this model. Right now I am trying hyper-parameter tuning. Is there any other suggestions?

", "messages": [ { "id": "1641834", "postDate": "01/07/2022 18:32:34", "content": "

I have been trying to train yolov5 model for the challenge Tensorflow Great Barrier Reef,
\nI saw in discussion that 0.48 was yolov5 got for this competition however no matter what I do I am not able to cross 0.425 barrier. So far I have:

\n
    \n
  1. Trained model on preprocessed Underwater images using tf addons.
    2. \n
  2. Implemented Tracking which actually increased score a bit. [https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539]
    3. \n
  3. Trained yoloL6 model.
    \n0.42 is what I got for yoloL6 with Tracking. Tracking clearly improves the result as is evident from the video I am attaching here.
    \nhttps://drive.google.com/file/d/11gp6IDav0X7fNxBYG3C0uL6qOZ6pj0eD/view?usp=sharing
    4. \n
\n

I just want to know what more can be done to improve this model. Right now I am trying hyper-parameter tuning. Is there any other suggestions?

", "rawMarkdown": "I have been trying to train yolov5 model for the challenge Tensorflow Great Barrier Reef, \nI saw in discussion that 0.48 was yolov5 got for this competition however no matter what I do I am not able to cross 0.425 barrier. So far I have:\n1. Trained model on preprocessed Underwater images using tf addons.\n2. Implemented Tracking which actually increased score a bit. [https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539]\n3. Trained yoloL6 model.\n0.42 is what I got for yoloL6 with Tracking. Tracking clearly improves the result as is evident from the video I am attaching here.\n[https://drive.google.com/file/d/11gp6IDav0X7fNxBYG3C0uL6qOZ6pj0eD/view?usp=sharing](url)\n\nI just want to know what more can be done to improve this model. Right now I am trying hyper-parameter tuning. Is there any other suggestions?", "votes": null }, { "id": "1641881", "postDate": "01/07/2022 19:38:06", "content": "

@justforgags I am also currently stuck with underperforming yolov5 models. The highest I've gotten without shared notebooks is 0.437. Moreover, I believe the highest lb with a yolov5 model is 0.61, not 0.48.

\n

Currently, I am experimenting with different hyperparameters (increasing augmentation) and changing my val split (to 3-fold video split) to tune the model for higher recall. I hope to break 0.5 in the upcoming week. I suggest a sequence-based split if you aren't using one already.

\n

(edit) Can you tell me what your CV score is for your best-performing (0.42) model?

", "rawMarkdown": "justforgags I am also currently stuck with underperforming yolov5 models. The highest I've gotten without shared notebooks is 0.437. Moreover, I believe the highest lb with a [yolov5 model is 0.61](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/298514), not 0.48.\n\nCurrently, I am experimenting with different hyperparameters (increasing augmentation) and changing my val split (to 3-fold video split) to tune the model for higher recall. I hope to break 0.5 in the upcoming week. I suggest a sequence-based split if you aren't using one already.\n\n(edit) Can you tell me what your CV score is for your best-performing (0.42) model?", "votes": null }, { "id": "1641941", "postDate": "01/07/2022 20:43:15", "content": "

Oh I am now switching back to yolov5m and trying on increasing accuracy in this and then move to yolov5L6 since heavier model is taking a lot of time to train. It is very reassuring to hear that yolov5 notebooks are performing so well.

\n

I am currently training yolov5m with hyper parameters optimization. I am using 4-fold video split for my validation. I will definitely try sequence based split approach.

\n

My current CV score - Precision, Recall, mAP 0.5, mAP 0.5:0.95 is 0.95646, 0.47135, 0.64296, 0.3754 respectively.

", "rawMarkdown": "Oh I am now switching back to yolov5m and trying on increasing accuracy in this and then move to yolov5L6 since heavier model is taking a lot of time to train. It is very reassuring to hear that yolov5 notebooks are performing so well.\n\nI am currently training yolov5m with hyper parameters optimization. I am using 4-fold video split for my validation. I will definitely try sequence based split approach.\n\nMy current CV score - Precision, Recall, mAP 0.5, mAP 0.5:0.95 is 0.95646, 0.47135, 0.64296, 0.3754 respectively.", "votes": null }, { "id": "1642025", "postDate": "01/08/2022 00:28:13", "content": "

I am also working with YOLOv5. The LB for my best model was 0.476, using yolov5l and changing some augmentations hyps as hsv_s, hsv_v and mosaic. There are still a lot of hyps to check.

", "rawMarkdown": "I am also working with YOLOv5. The LB for my best model was 0.476, using yolov5l and changing some augmentations hyps as hsv_s, hsv_v and mosaic. There are still a lot of hyps to check.", "votes": null }, { "id": "1642242", "postDate": "01/08/2022 07:14:30", "content": "

Did you try yolov5l6 model too? I think its recommended for 1280 image size

", "rawMarkdown": "Did you try yolov5l6 model too? I think its recommended for 1280 image size", "votes": null }, { "id": "1642264", "postDate": "01/08/2022 08:07:08", "content": "

Great! Let me know if you have success with any single change.

\n

Also, you might want to implement the competition metric for CV.

", "rawMarkdown": "Great! Let me know if you have success with any single change.\n\nAlso, you might want to implement the competition metric for CV.", "votes": null }, { "id": "1642601", "postDate": "01/08/2022 14:25:39", "content": "

Currently, I have Yolov5 score (Public 0.561) without post-processing and ensemble.
\nYolov5's performance is good enough. In order for you to get high performance with Yolov5 in this contest, you need to change the parameters etc. of the Yolov5 library.

", "rawMarkdown": "Currently, I have Yolov5 score (Public 0.561) without post-processing and ensemble.\nYolov5's performance is good enough. In order for you to get high performance with Yolov5 in this contest, you need to change the parameters etc. of the Yolov5 library.", "votes": null }, { "id": "1642639", "postDate": "01/08/2022 14:48:15", "content": "

What do you mean by parameters? Can I get a high performance just by changing the parameters from hyp.scratch.yaml? Thanks!

", "rawMarkdown": "What do you mean by parameters? Can I get a high performance just by changing the parameters from hyp.scratch.yaml? Thanks!", "votes": null }, { "id": "1642686", "postDate": "01/08/2022 15:20:36", "content": "

Oh okay should I form a function for calculating F2 score for CV?

", "rawMarkdown": "Oh okay should I form a function for calculating F2 score for CV?", "votes": null }, { "id": "1642691", "postDate": "01/08/2022 15:24:20", "content": "

Hi ,
\nI am currently trying to make changes to the dataset itself, I was using kfold split but now I am shifting to sub-sequence splits. I will try tuning hyper parameter after I get benchmark improvement in changing dataset. Also regarding ensemble are you using weighted box fusion? What are the models you are using for ensemble?

", "rawMarkdown": "Hi ,\nI am currently trying to make changes to the dataset itself, I was using kfold split but now I am shifting to sub-sequence splits. I will try tuning hyper parameter after I get benchmark improvement in changing dataset. Also regarding ensemble are you using weighted box fusion? What are the models you are using for ensemble?", "votes": null }, { "id": "1642702", "postDate": "01/08/2022 15:33:26", "content": "

I haven't done the ensemble yet.

", "rawMarkdown": "I haven't done the ensemble yet.", "votes": null }, { "id": "1642883", "postDate": "01/08/2022 19:06:50", "content": "

Yeah, you can find an implementation in the comments of this discussion.

", "rawMarkdown": "Yeah, you can find an implementation in the comments of [this discussion](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290757).", "votes": null }, { "id": "1644893", "postDate": "01/10/2022 15:29:06", "content": "

Hi, anyone succeed in ensemble? Really struggled in ensemble. Cannot get a reasonable score…

", "rawMarkdown": "Hi, anyone succeed in ensemble? Really struggled in ensemble. Cannot get a reasonable score...", "votes": null }, { "id": "1646748", "postDate": "01/12/2022 02:23:04", "content": "

Well, \"I was using kfold split but now I am shifting to sub-sequence splits. \", will this improve the performance of the model?

", "rawMarkdown": "Well, \"I was using kfold split but now I am shifting to sub-sequence splits. \", will this improve the performance of the model?", "votes": null }, { "id": "1646874", "postDate": "01/12/2022 06:00:50", "content": "

Not yet, I think the key to succeed seems to be in dataset. I find my model facing difficulties in detecting very small COTS and COTS which are something like on side of some rock. I think augmentation will be key. :\"(

", "rawMarkdown": "Not yet, I think the key to succeed seems to be in dataset. I find my model facing difficulties in detecting very small COTS and COTS which are something like on side of some rock. I think augmentation will be key. :\"(", "votes": null }, { "id": "1647485", "postDate": "01/12/2022 16:35:34", "content": "

Yes, but i got worse results.

", "rawMarkdown": "Yes, but i got worse results.", "votes": null } ], "comments": [ { "id": 1641881, "author_name": "ther3venant", "author_url": "", "post_date": "01/07/2022 19:38:06", "content": "

@justforgags I am also currently stuck with underperforming yolov5 models. The highest I've gotten without shared notebooks is 0.437. Moreover, I believe the highest lb with a yolov5 model is 0.61, not 0.48.

\n

Currently, I am experimenting with different hyperparameters (increasing augmentation) and changing my val split (to 3-fold video split) to tune the model for higher recall. I hope to break 0.5 in the upcoming week. I suggest a sequence-based split if you aren't using one already.

\n

(edit) Can you tell me what your CV score is for your best-performing (0.42) model?

", "votes": null, "replies": [ { "id": 1641941, "author_name": "justforgags", "author_url": "", "post_date": "01/07/2022 20:43:15", "content": "

Oh I am now switching back to yolov5m and trying on increasing accuracy in this and then move to yolov5L6 since heavier model is taking a lot of time to train. It is very reassuring to hear that yolov5 notebooks are performing so well.

\n

I am currently training yolov5m with hyper parameters optimization. I am using 4-fold video split for my validation. I will definitely try sequence based split approach.

\n

My current CV score - Precision, Recall, mAP 0.5, mAP 0.5:0.95 is 0.95646, 0.47135, 0.64296, 0.3754 respectively.

", "votes": null, "replies": [] }, { "id": 1642264, "author_name": "ther3venant", "author_url": "", "post_date": "01/08/2022 08:07:08", "content": "

Great! Let me know if you have success with any single change.

\n

Also, you might want to implement the competition metric for CV.

", "votes": null, "replies": [] }, { "id": 1642686, "author_name": "justforgags", "author_url": "", "post_date": "01/08/2022 15:20:36", "content": "

Oh okay should I form a function for calculating F2 score for CV?

", "votes": null, "replies": [] }, { "id": 1642883, "author_name": "ther3venant", "author_url": "", "post_date": "01/08/2022 19:06:50", "content": "

Yeah, you can find an implementation in the comments of this discussion.

", "votes": null, "replies": [] } ] }, { "id": 1642025, "author_name": "lucaspillapimentel", "author_url": "", "post_date": "01/08/2022 00:28:13", "content": "

I am also working with YOLOv5. The LB for my best model was 0.476, using yolov5l and changing some augmentations hyps as hsv_s, hsv_v and mosaic. There are still a lot of hyps to check.

", "votes": null, "replies": [ { "id": 1642242, "author_name": "justforgags", "author_url": "", "post_date": "01/08/2022 07:14:30", "content": "

Did you try yolov5l6 model too? I think its recommended for 1280 image size

", "votes": null, "replies": [] }, { "id": 1647485, "author_name": "lucaspillapimentel", "author_url": "", "post_date": "01/12/2022 16:35:34", "content": "

Yes, but i got worse results.

", "votes": null, "replies": [] } ] }, { "id": 1642601, "author_name": "syurenuko", "author_url": "", "post_date": "01/08/2022 14:25:39", "content": "

Currently, I have Yolov5 score (Public 0.561) without post-processing and ensemble.
\nYolov5's performance is good enough. In order for you to get high performance with Yolov5 in this contest, you need to change the parameters etc. of the Yolov5 library.

", "votes": null, "replies": [ { "id": 1642639, "author_name": "lucaspillapimentel", "author_url": "", "post_date": "01/08/2022 14:48:15", "content": "

What do you mean by parameters? Can I get a high performance just by changing the parameters from hyp.scratch.yaml? Thanks!

", "votes": null, "replies": [] }, { "id": 1642691, "author_name": "justforgags", "author_url": "", "post_date": "01/08/2022 15:24:20", "content": "

Hi ,
\nI am currently trying to make changes to the dataset itself, I was using kfold split but now I am shifting to sub-sequence splits. I will try tuning hyper parameter after I get benchmark improvement in changing dataset. Also regarding ensemble are you using weighted box fusion? What are the models you are using for ensemble?

", "votes": null, "replies": [] }, { "id": 1642702, "author_name": "syurenuko", "author_url": "", "post_date": "01/08/2022 15:33:26", "content": "

I haven't done the ensemble yet.

", "votes": null, "replies": [] }, { "id": 1644893, "author_name": "xiaojiu1414", "author_url": "", "post_date": "01/10/2022 15:29:06", "content": "

Hi, anyone succeed in ensemble? Really struggled in ensemble. Cannot get a reasonable score…

", "votes": null, "replies": [] }, { "id": 1646874, "author_name": "justforgags", "author_url": "", "post_date": "01/12/2022 06:00:50", "content": "

Not yet, I think the key to succeed seems to be in dataset. I find my model facing difficulties in detecting very small COTS and COTS which are something like on side of some rock. I think augmentation will be key. :\"(

", "votes": null, "replies": [] } ] }, { "id": 1646748, "author_name": "hkinchina", "author_url": "", "post_date": "01/12/2022 02:23:04", "content": "

Well, \"I was using kfold split but now I am shifting to sub-sequence splits. \", will this improve the performance of the model?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1641834": "I have been trying to train yolov5 model for the challenge Tensorflow Great Barrier Reef, \nI saw in discussion that 0.48 was yolov5 got for this competition however no matter what I do I am not able to cross 0.425 barrier. So far I have:\n1. Trained model on preprocessed Underwater images using tf addons.\n2. Implemented Tracking which actually increased score a bit. [https://www.kaggle.com/parapapapam/yolox-inference-tracking-on-cots-lb-0-539]\n3. Trained yoloL6 model.\n0.42 is what I got for yoloL6 with Tracking. Tracking clearly improves the result as is evident from the video I am attaching here.\n[https://drive.google.com/file/d/11gp6IDav0X7fNxBYG3C0uL6qOZ6pj0eD/view?usp=sharing](url)\n\nI just want to know what more can be done to improve this model. Right now I am trying hyper-parameter tuning. Is there any other suggestions?", "1641881": "justforgags I am also currently stuck with underperforming yolov5 models. The highest I've gotten without shared notebooks is 0.437. Moreover, I believe the highest lb with a [yolov5 model is 0.61](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/298514), not 0.48.\n\nCurrently, I am experimenting with different hyperparameters (increasing augmentation) and changing my val split (to 3-fold video split) to tune the model for higher recall. I hope to break 0.5 in the upcoming week. I suggest a sequence-based split if you aren't using one already.\n\n(edit) Can you tell me what your CV score is for your best-performing (0.42) model?", "1641941": "Oh I am now switching back to yolov5m and trying on increasing accuracy in this and then move to yolov5L6 since heavier model is taking a lot of time to train. It is very reassuring to hear that yolov5 notebooks are performing so well.\n\nI am currently training yolov5m with hyper parameters optimization. I am using 4-fold video split for my validation. I will definitely try sequence based split approach.\n\nMy current CV score - Precision, Recall, mAP 0.5, mAP 0.5:0.95 is 0.95646, 0.47135, 0.64296, 0.3754 respectively.", "1642025": "I am also working with YOLOv5. The LB for my best model was 0.476, using yolov5l and changing some augmentations hyps as hsv_s, hsv_v and mosaic. There are still a lot of hyps to check.", "1642242": "Did you try yolov5l6 model too? I think its recommended for 1280 image size", "1642264": "Great! Let me know if you have success with any single change.\n\nAlso, you might want to implement the competition metric for CV.", "1642601": "Currently, I have Yolov5 score (Public 0.561) without post-processing and ensemble.\nYolov5's performance is good enough. In order for you to get high performance with Yolov5 in this contest, you need to change the parameters etc. of the Yolov5 library.", "1642639": "What do you mean by parameters? Can I get a high performance just by changing the parameters from hyp.scratch.yaml? Thanks!", "1642686": "Oh okay should I form a function for calculating F2 score for CV?", "1642691": "Hi ,\nI am currently trying to make changes to the dataset itself, I was using kfold split but now I am shifting to sub-sequence splits. I will try tuning hyper parameter after I get benchmark improvement in changing dataset. Also regarding ensemble are you using weighted box fusion? What are the models you are using for ensemble?", "1642702": "I haven't done the ensemble yet.", "1642883": "Yeah, you can find an implementation in the comments of [this discussion](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290757).", "1644893": "Hi, anyone succeed in ensemble? Really struggled in ensemble. Cannot get a reasonable score...", "1646748": "Well, \"I was using kfold split but now I am shifting to sub-sequence splits. \", will this improve the performance of the model?", "1646874": "Not yet, I think the key to succeed seems to be in dataset. I find my model facing difficulties in detecting very small COTS and COTS which are something like on side of some rock. I think augmentation will be key. :\"(", "1647485": "Yes, but i got worse results." }, "source": "meta" }