{ "id": 291663, "title": "YoloX Kaggle notebook - I made it! ❤️", "url": "/competitions/tensorflow-great-barrier-reef/discussion/291663", "author_name": "Remek Kinas", "post_date": "2021-11-30T12:48:45.246000", "votes": 64, "comment_count": 31, "views": 0, "content": "

I was looking for a reference YOLOX notebook on Kaggle showing how to train custom models, but unfortunately I have not find it. It was also not in list on the great post by @usharengaraju Kaggle Starter Notebooks for Object Detection.

\n

Here you can find COTS YOLOX full pipeline for your experiments. Hope you find this notebook helpful. We still have a lot of time in this competition so there will be opportunities for many experiments usung YOLOX.

\n

Here you can find -> ❤️ YoloX full training pipeline for COTS dataset ❤️

\n

Inference part for COTS competition you can find here: https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots

\n

If you have any ideas how to make it better please let me know. Thank you for your feedback.

\n

Score log:

\n", "messages": [ { "id": 1600460, "postDate": "2021-11-30T12:48:45.247Z", "content": "

I was looking for a reference YOLOX notebook on Kaggle showing how to train custom models, but unfortunately I have not find it. It was also not in list on the great post by @usharengaraju Kaggle Starter Notebooks for Object Detection.

\n

Here you can find COTS YOLOX full pipeline for your experiments. Hope you find this notebook helpful. We still have a lot of time in this competition so there will be opportunities for many experiments usung YOLOX.

\n

Here you can find -> ❤️ YoloX full training pipeline for COTS dataset ❤️

\n

Inference part for COTS competition you can find here: https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots

\n

If you have any ideas how to make it better please let me know. Thank you for your feedback.

\n

Score log:

\n", "rawMarkdown": "I was looking for a reference YOLOX notebook on Kaggle showing how to train custom models, but unfortunately I have not find it. It was also not in list on the great post by @usharengaraju [Kaggle Starter Notebooks for Object Detection](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290016).\n\nHere you can find COTS YOLOX full pipeline for your experiments. Hope you find this notebook helpful. We still have a lot of time in this competition so there will be opportunities for many experiments usung YOLOX.\n\n\nHere you can find -> ❤️ [YoloX full training pipeline for COTS dataset](https://www.kaggle.com/remekkinas/yolox-full-training-pipeline-for-cots-dataset) ❤️\n\n\nInference part for COTS competition you can find here: https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots\n\nIf you have any ideas how to make it better please let me know. Thank you for your feedback.\n\n\nScore log:\n- 2021.12.04 - LB 0.483 💪\n- 2021.12.05 - Now even better .... LB 0.507 😍😍😍", "votes": 63 }, { "id": 1606162, "postDate": "2021-12-04T21:39:35.867Z", "content": "

The best one model on public score [0.483] … so far … and new model is learning now …. I am happy! 😄

", "rawMarkdown": "The best one model on public score [0.483] ... so far ... and new model is learning now .... I am happy! 😄 ", "votes": 3, "replies": [ { "id": 1606564, "postDate": "2021-12-05T04:43:08.453Z", "content": "

which weight? I only got 0.43

\n

never mind. I got it.

\n

how many epochs did you trained?

", "rawMarkdown": "which weight? I only got 0.43\n\nnever mind. I got it.\n\nhow many epochs did you trained? " } ] }, { "id": 1606349, "postDate": "2021-12-05T00:15:02.813Z", "content": "

Some people said we should get a new graphics card, then I'm thinking about how to compile on opencl
\nwe are not the same

", "rawMarkdown": "Some people said we should get a new graphics card, then I'm thinking about how to compile on opencl\nwe are not the same", "votes": 1 }, { "id": 1604043, "postDate": "2021-12-03T04:13:38.507Z", "content": "

thx for sharing. based my trial, large model is better.

", "rawMarkdown": "thx for sharing. based my trial, large model is better.", "votes": 1, "replies": [ { "id": 1604074, "postDate": "2021-12-03T04:49:17.703Z", "content": "

Yes. I agree. Changing training dataset strategy will be beneficial as well. I will do it when I got new GPU resources:)

", "rawMarkdown": "Yes. I agree. Changing training dataset strategy will be beneficial as well. I will do it when I got new GPU resources:)", "votes": 2 } ] }, { "id": 1603599, "postDate": "2021-12-02T16:06:51.387Z", "content": "

Good job for learning new thing Remek! and nice workaround with offline installing package

", "rawMarkdown": "Good job for learning new thing Remek! and nice workaround with offline installing package", "votes": 1, "replies": [ { "id": 1603605, "postDate": "2021-12-02T16:10:47.550Z", "content": "

Thank you! Now it is time to play with input data (in notebook is far from optimal …) and build strong model. I have to wait … kaggle GPU time for this week = 0 :)

", "rawMarkdown": "Thank you! Now it is time to play with input data (in notebook is far from optimal ...) and build strong model. I have to wait ... kaggle GPU time for this week = 0 :)", "votes": 1 }, { "id": 1603618, "postDate": "2021-12-02T16:19:47.450Z", "content": "

Try colab with kaggle API 😀

", "rawMarkdown": "Try colab with kaggle API 😀" }, { "id": 1603621, "postDate": "2021-12-02T16:22:38.253Z", "content": "

Yes, I work such way but to commit changes (even for inference) I need at least 5 minutes of available GPU on Kaggle 😄😆

", "rawMarkdown": "Yes, I work such way but to commit changes (even for inference) I need at least 5 minutes of available GPU on Kaggle 😄😆", "votes": 1 }, { "id": 1603624, "postDate": "2021-12-02T16:26:24.567Z", "content": "

So now it is a problem 😀

", "rawMarkdown": "So now it is a problem 😀" }, { "id": 1603637, "postDate": "2021-12-02T16:44:58.460Z", "content": "

You know Lukasz … in TPS-12 where I temporarily switched I do not need GPU … :) Currently I use standard ML tools ….

", "rawMarkdown": "You know Lukasz ... in TPS-12 where I temporarily switched I do not need GPU ... :) Currently I use standard ML tools .... ", "votes": 1 } ] }, { "id": 1608018, "postDate": "2021-12-06T07:21:06.903Z", "content": "

Thanks for your share. What is your CV during training👀

", "rawMarkdown": "Thanks for your share. What is your CV during training👀", "votes": 2, "replies": [ { "id": 1608085, "postDate": "2021-12-06T08:13:47.080Z", "content": "

You can get his CV using

\n
\n

python tools/eval.py -f cots_config.py -n yolox-l -c yx_l_003.pth -b 8 -d 1

\n
", "rawMarkdown": "You can get his CV using \n> python tools/eval.py -f cots_config.py -n yolox-l -c yx_l_003.pth -b 8 -d 1", "votes": 3 }, { "id": 1608092, "postDate": "2021-12-06T08:16:11.067Z", "content": "

okay, thanks

", "rawMarkdown": "okay, thanks" }, { "id": 1616263, "postDate": "2021-12-13T09:31:00.360Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1618601, "postDate": "2021-12-15T06:25:14.457Z", "content": "

Cross Validation I guess

", "rawMarkdown": "Cross Validation I guess" } ] }, { "id": 1607418, "postDate": "2021-12-05T22:33:21.737Z", "content": "

Score log:

\n", "rawMarkdown": "Score log:\n- 2021.12.04 - LB 0.483 💪\n- 2021.12.05 - Now even better .... LB 0.507 😍😍😍", "votes": 2, "replies": [ { "id": 1622668, "postDate": "2021-12-19T03:03:51.587Z", "content": "

Thanks for sharing!
\nWhat's the difference between both score logs?

", "rawMarkdown": "Thanks for sharing!\nWhat's the difference between both score logs?" } ] }, { "id": 1657140, "postDate": "2022-01-20T00:15:16.367Z", "content": "

Thank you for sharing the notebook and LB. I'll try the notebook!! :)

", "rawMarkdown": "Thank you for sharing the notebook and LB. I'll try the notebook!! :)" }, { "id": 1612138, "postDate": "2021-12-08T15:32:02.147Z", "content": "

how many epochs did you train for 003.pth? I want to continue training with it.

", "rawMarkdown": "how many epochs did you train for 003.pth? I want to continue training with it.", "replies": [ { "id": 1612140, "postDate": "2021-12-08T15:33:34.213Z", "content": "

25 epochs only … as I can see best models are trained after 15-18 epochs

", "rawMarkdown": "25 epochs only ... as I can see best models are trained after 15-18 epochs ", "votes": 1 }, { "id": 1612565, "postDate": "2021-12-09T05:14:39.200Z", "content": "

Thanks for your reply. I tried to train on free Colab, however only around 12 epochs with limited gpu, It stopped.

\n

Half year ago, I could run about 10 hours even with better GPU.

", "rawMarkdown": "Thanks for your reply. I tried to train on free Colab, however only around 12 epochs with limited gpu, It stopped.\n\nHalf year ago, I could run about 10 hours even with better GPU.\n\n" } ] }, { "id": 1608632, "postDate": "2021-12-06T14:05:07.677Z", "content": "

thanks for sharing, but have you trained with yolox-x yet? I am trying with yolox-x but cuda is always out of memory on kaggle :(((

", "rawMarkdown": "thanks for sharing, but have you trained with yolox-x yet? I am trying with yolox-x but cuda is always out of memory on kaggle :(((", "replies": [ { "id": 1612139, "postDate": "2021-12-08T15:33:10.667Z", "content": "

I think that He trains offline. If you want to train on kaggle, you have to reduce batch size.

", "rawMarkdown": "I think that He trains offline. If you want to train on kaggle, you have to reduce batch size." }, { "id": 1612142, "postDate": "2021-12-08T15:35:13.023Z", "content": "

Yes. I train all models offline on my machine. There is not enough GPU for me on Kaggle (I start on only in this competition).

", "rawMarkdown": "Yes. I train all models offline on my machine. There is not enough GPU for me on Kaggle (I start on only in this competition). ", "votes": 2 }, { "id": 1617524, "postDate": "2021-12-14T06:30:17.877Z", "content": "

Well, that sounds interesting, but how do you train all models offline on your machine?

", "rawMarkdown": "Well, that sounds interesting, but how do you train all models offline on your machine?" }, { "id": 1617556, "postDate": "2021-12-14T07:02:56.080Z", "content": "

Yes, my local machine

", "rawMarkdown": "Yes, my local machine", "votes": 2 }, { "id": 1618419, "postDate": "2021-12-15T01:00:24.130Z", "content": "

I mean not on Kaggle platform. not referring to internet.

", "rawMarkdown": "I mean not on Kaggle platform. not referring to internet." } ] }, { "id": 1602662, "postDate": "2021-12-02T03:51:36.463Z", "content": "

refer to yolov5, put installation package to a dataset for offline submission

", "rawMarkdown": "refer to yolov5, put installation package to a dataset for offline submission", "replies": [ { "id": 1602824, "postDate": "2021-12-02T06:48:47.567Z", "content": "

Hi, yes I am doing that (here is dataset -> https://www.kaggle.com/remekkinas/yolox-cots-models) …. Today shoud be finished :)

", "rawMarkdown": "Hi, yes I am doing that (here is dataset -> https://www.kaggle.com/remekkinas/yolox-cots-models) .... Today shoud be finished :)", "votes": 1 }, { "id": 1603171, "postDate": "2021-12-02T11:35:23.383Z", "content": "

You can find final inference script here …. https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots

\n

It was … like a hacking :)

", "rawMarkdown": "You can find final inference script here .... https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots\n\nIt was ... like a hacking :)", "votes": 2 } ] } ], "comments": [ { "id": 1606162, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-04T21:39:35.867000", "content": "

The best one model on public score [0.483] … so far … and new model is learning now …. I am happy! 😄

", "votes": 3, "replies": [ { "id": 1606564, "author_name": "dragon zhang", "author_url": "", "post_date": "2021-12-05T04:43:08.453000", "content": "

which weight? I only got 0.43

\n

never mind. I got it.

\n

how many epochs did you trained?

", "votes": 0, "replies": [] } ] }, { "id": 1606349, "author_name": "Mahdi_asdzd", "author_url": "", "post_date": "2021-12-05T00:15:02.813000", "content": "

Some people said we should get a new graphics card, then I'm thinking about how to compile on opencl
\nwe are not the same

", "votes": 1, "replies": [] }, { "id": 1604043, "author_name": "dragon zhang", "author_url": "", "post_date": "2021-12-03T04:13:38.507000", "content": "

thx for sharing. based my trial, large model is better.

", "votes": 1, "replies": [ { "id": 1604074, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-03T04:49:17.703000", "content": "

Yes. I agree. Changing training dataset strategy will be beneficial as well. I will do it when I got new GPU resources:)

", "votes": 2, "replies": [] } ] }, { "id": 1603599, "author_name": "Lukasz Borecki", "author_url": "", "post_date": "2021-12-02T16:06:51.387000", "content": "

Good job for learning new thing Remek! and nice workaround with offline installing package

", "votes": 1, "replies": [ { "id": 1603605, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-02T16:10:47.550000", "content": "

Thank you! Now it is time to play with input data (in notebook is far from optimal …) and build strong model. I have to wait … kaggle GPU time for this week = 0 :)

", "votes": 1, "replies": [] }, { "id": 1603618, "author_name": "Lukasz Borecki", "author_url": "", "post_date": "2021-12-02T16:19:47.450000", "content": "

Try colab with kaggle API 😀

", "votes": 0, "replies": [] }, { "id": 1603621, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-02T16:22:38.253000", "content": "

Yes, I work such way but to commit changes (even for inference) I need at least 5 minutes of available GPU on Kaggle 😄😆

", "votes": 1, "replies": [] }, { "id": 1603624, "author_name": "Lukasz Borecki", "author_url": "", "post_date": "2021-12-02T16:26:24.567000", "content": "

So now it is a problem 😀

", "votes": 0, "replies": [] }, { "id": 1603637, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-02T16:44:58.460000", "content": "

You know Lukasz … in TPS-12 where I temporarily switched I do not need GPU … :) Currently I use standard ML tools ….

", "votes": 1, "replies": [] } ] }, { "id": 1608018, "author_name": "xuxu_sky", "author_url": "", "post_date": "2021-12-06T07:21:06.903000", "content": "

Thanks for your share. What is your CV during training👀

", "votes": 2, "replies": [ { "id": 1608085, "author_name": "Leon", "author_url": "", "post_date": "2021-12-06T08:13:47.080000", "content": "

You can get his CV using

\n
\n

python tools/eval.py -f cots_config.py -n yolox-l -c yx_l_003.pth -b 8 -d 1

\n
", "votes": 3, "replies": [] }, { "id": 1608092, "author_name": "xuxu_sky", "author_url": "", "post_date": "2021-12-06T08:16:11.067000", "content": "

okay, thanks

", "votes": 0, "replies": [] }, { "id": 1616263, "author_name": "", "author_url": "", "post_date": "2021-12-13T09:31:00.360000", "content": "", "votes": 0, "replies": [] }, { "id": 1618601, "author_name": "Kevin", "author_url": "", "post_date": "2021-12-15T06:25:14.457000", "content": "

Cross Validation I guess

", "votes": 0, "replies": [] } ] }, { "id": 1607418, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-05T22:33:21.737000", "content": "

Score log:

\n", "votes": 2, "replies": [ { "id": 1622668, "author_name": "Diego Gomez", "author_url": "", "post_date": "2021-12-19T03:03:51.587000", "content": "

Thanks for sharing!
\nWhat's the difference between both score logs?

", "votes": 0, "replies": [] } ] }, { "id": 1657140, "author_name": "Eugene J. Ryu", "author_url": "", "post_date": "2022-01-20T00:15:16.367000", "content": "

Thank you for sharing the notebook and LB. I'll try the notebook!! :)

", "votes": 0, "replies": [] }, { "id": 1612138, "author_name": "dragon zhang", "author_url": "", "post_date": "2021-12-08T15:32:02.147000", "content": "

how many epochs did you train for 003.pth? I want to continue training with it.

", "votes": 0, "replies": [ { "id": 1612140, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-08T15:33:34.213000", "content": "

25 epochs only … as I can see best models are trained after 15-18 epochs

", "votes": 1, "replies": [] }, { "id": 1612565, "author_name": "dragon zhang", "author_url": "", "post_date": "2021-12-09T05:14:39.200000", "content": "

Thanks for your reply. I tried to train on free Colab, however only around 12 epochs with limited gpu, It stopped.

\n

Half year ago, I could run about 10 hours even with better GPU.

", "votes": 0, "replies": [] } ] }, { "id": 1608632, "author_name": "letrongminh", "author_url": "", "post_date": "2021-12-06T14:05:07.677000", "content": "

thanks for sharing, but have you trained with yolox-x yet? I am trying with yolox-x but cuda is always out of memory on kaggle :(((

", "votes": 0, "replies": [ { "id": 1612139, "author_name": "dragon zhang", "author_url": "", "post_date": "2021-12-08T15:33:10.667000", "content": "

I think that He trains offline. If you want to train on kaggle, you have to reduce batch size.

", "votes": 0, "replies": [] }, { "id": 1612142, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-08T15:35:13.023000", "content": "

Yes. I train all models offline on my machine. There is not enough GPU for me on Kaggle (I start on only in this competition).

", "votes": 2, "replies": [] }, { "id": 1617524, "author_name": "Dino Wun", "author_url": "", "post_date": "2021-12-14T06:30:17.877000", "content": "

Well, that sounds interesting, but how do you train all models offline on your machine?

", "votes": 0, "replies": [] }, { "id": 1617556, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-14T07:02:56.080000", "content": "

Yes, my local machine

", "votes": 2, "replies": [] }, { "id": 1618419, "author_name": "dragon zhang", "author_url": "", "post_date": "2021-12-15T01:00:24.130000", "content": "

I mean not on Kaggle platform. not referring to internet.

", "votes": 0, "replies": [] } ] }, { "id": 1602662, "author_name": "dragon zhang", "author_url": "", "post_date": "2021-12-02T03:51:36.463000", "content": "

refer to yolov5, put installation package to a dataset for offline submission

", "votes": 0, "replies": [ { "id": 1602824, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-02T06:48:47.567000", "content": "

Hi, yes I am doing that (here is dataset -> https://www.kaggle.com/remekkinas/yolox-cots-models) …. Today shoud be finished :)

", "votes": 1, "replies": [] }, { "id": 1603171, "author_name": "Remek Kinas", "author_url": "", "post_date": "2021-12-02T11:35:23.383000", "content": "

You can find final inference script here …. https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots

\n

It was … like a hacking :)

", "votes": 2, "replies": [] } ] } ], "raw_markdown_by_id": { "1600460": "I was looking for a reference YOLOX notebook on Kaggle showing how to train custom models, but unfortunately I have not find it. It was also not in list on the great post by @usharengaraju [Kaggle Starter Notebooks for Object Detection](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/290016).\n\nHere you can find COTS YOLOX full pipeline for your experiments. Hope you find this notebook helpful. We still have a lot of time in this competition so there will be opportunities for many experiments usung YOLOX.\n\n\nHere you can find -> ❤️ [YoloX full training pipeline for COTS dataset](https://www.kaggle.com/remekkinas/yolox-full-training-pipeline-for-cots-dataset) ❤️\n\n\nInference part for COTS competition you can find here: https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots\n\nIf you have any ideas how to make it better please let me know. Thank you for your feedback.\n\n\nScore log:\n- 2021.12.04 - LB 0.483 💪\n- 2021.12.05 - Now even better .... LB 0.507 😍😍😍", "1606162": "The best one model on public score [0.483] ... so far ... and new model is learning now .... I am happy! 😄 ", "1606349": "Some people said we should get a new graphics card, then I'm thinking about how to compile on opencl\nwe are not the same", "1604043": "thx for sharing. based my trial, large model is better.", "1603599": "Good job for learning new thing Remek! and nice workaround with offline installing package", "1608018": "Thanks for your share. What is your CV during training👀", "1607418": "Score log:\n- 2021.12.04 - LB 0.483 💪\n- 2021.12.05 - Now even better .... LB 0.507 😍😍😍", "1657140": "Thank you for sharing the notebook and LB. I'll try the notebook!! :)", "1612138": "how many epochs did you train for 003.pth? I want to continue training with it.", "1608632": "thanks for sharing, but have you trained with yolox-x yet? I am trying with yolox-x but cuda is always out of memory on kaggle :(((", "1602662": "refer to yolov5, put installation package to a dataset for offline submission" } }