{ "id": 306125, "title": "Is there a way to beat 0.665 LB notebook?", "url": "/competitions/tensorflow-great-barrier-reef/discussion/306125", "author_name": "", "post_date": "2022-02-08T08:40:38.423550400Z", "votes": 6, "comment_count": 18, "views": 0, "content": "

I originally had a good place on the LB (0.654), but after the 0.665 LB notebook was published, I dropped by many places.

\n

I've been trying to fine-tune IOU or CONF in the 0.665 notebook, but I can't get higher than 0.665. How do I find the optimum CONF and IOU?

\n

Also, I noticed that the notebook doesn't use tracking. Would tracking boost LB?

", "messages": [ { "id": "1681106", "postDate": "02/08/2022 08:40:38", "content": "

I originally had a good place on the LB (0.654), but after the 0.665 LB notebook was published, I dropped by many places.

\n

I've been trying to fine-tune IOU or CONF in the 0.665 notebook, but I can't get higher than 0.665. How do I find the optimum CONF and IOU?

\n

Also, I noticed that the notebook doesn't use tracking. Would tracking boost LB?

", "rawMarkdown": "I originally had a good place on the LB (0.654), but after the 0.665 LB notebook was published, I dropped by many places.\n\nI've been trying to fine-tune IOU or CONF in the 0.665 notebook, but I can't get higher than 0.665. How do I find the optimum CONF and IOU?\n\nAlso, I noticed that the notebook doesn't use tracking. Would tracking boost LB?", "votes": null }, { "id": "1681114", "postDate": "02/08/2022 08:47:18", "content": "

too many high scores on the LB.

\n

either better models or better parameters or both.

", "rawMarkdown": "too many high scores on the LB.\n\neither better models or better parameters or both.", "votes": null }, { "id": "1681119", "postDate": "02/08/2022 08:50:12", "content": "

@dragonzhang thank you for replying - how do you find better parameters (CONF and IOU)?

", "rawMarkdown": "dragonzhang thank you for replying - how do you find better parameters (CONF and IOU)?", "votes": null }, { "id": "1681169", "postDate": "02/08/2022 09:31:38", "content": "

I tried track but it does not work

", "rawMarkdown": "I tried track but it does not work", "votes": null }, { "id": "1681183", "postDate": "02/08/2022 09:47:03", "content": "

just try out your ideas and see what happen.

", "rawMarkdown": "just try out your ideas and see what happen.", "votes": null }, { "id": "1682138", "postDate": "02/09/2022 00:29:48", "content": "

tracking actually works, the right tracking boosts the score to 0.672

", "rawMarkdown": "tracking actually works, the right tracking boosts the score to 0.672", "votes": null }, { "id": "1682536", "postDate": "02/09/2022 08:27:42", "content": "

You did it with original parameters in the original notebook?

", "rawMarkdown": "You did it with original parameters in the original notebook?", "votes": null }, { "id": "1682577", "postDate": "02/09/2022 08:53:50", "content": "

Yes, we can publish our 0.710 …. 😍😂😆

", "rawMarkdown": "Yes, we can publish our 0.710 .... 😍😂😆", "votes": null }, { "id": "1682596", "postDate": "02/09/2022 09:06:45", "content": "

I think It's very easy
\nIncrease image size as much as you can

", "rawMarkdown": "I think It's very easy\nIncrease image size as much as you can", "votes": null }, { "id": "1682609", "postDate": "02/09/2022 09:25:59", "content": "

It is not too late to publish. lol.
\nBut I think overfit in CV is not as that much as in NLP.

", "rawMarkdown": "It is not too late to publish. lol.\nBut I think overfit in CV is not as that much as in NLP.", "votes": null }, { "id": "1682618", "postDate": "02/09/2022 09:38:04", "content": "

We will see …. 😄

", "rawMarkdown": "We will see .... 😄", "votes": null }, { "id": "1682625", "postDate": "02/09/2022 09:45:09", "content": "

really? Experiments show that It is not linear.

", "rawMarkdown": "really? Experiments show that It is not linear.", "votes": null }, { "id": "1682818", "postDate": "02/09/2022 11:54:01", "content": "

FYI no late sharing during quiet period

", "rawMarkdown": "[FYI](https://www.kaggle.com/general/291540) no late sharing during quiet period", "votes": null }, { "id": "1683030", "postDate": "02/09/2022 14:16:42", "content": "

but it increases until reaching some points

", "rawMarkdown": "but it increases until reaching some points", "votes": null }, { "id": "1683062", "postDate": "02/09/2022 14:47:55", "content": "

You can use the published public notebook weights and apply tricks and it will score above LB0.700.

", "rawMarkdown": "You can use the published public notebook weights and apply tricks and it will score above LB0.700.", "votes": null }, { "id": "1683078", "postDate": "02/09/2022 14:56:57", "content": "

And are you sure that these publicly available weights \"on steroids\" to fit the public LB will be equally useful for final results on private LB? I would be quite careful to give such advises.

", "rawMarkdown": "And are you sure that these publicly available weights \"on steroids\" to fit the public LB will be equally useful for final results on private LB? I would be quite careful to give such advises.", "votes": null }, { "id": "1683087", "postDate": "02/09/2022 15:07:47", "content": "

I didn't say anything whether it will be usefull for the private LB. I answered only the question of the topics author: -> you can tune it and get a higher LB score.

", "rawMarkdown": "I didn't say anything whether it will be usefull for the private LB. I answered only the question of the topics author: -> you can tune it and get a higher LB score.", "votes": null }, { "id": "1683091", "postDate": "02/09/2022 15:14:10", "content": "

OK, it is an appropriate clarification. I believe that you really don't want to confuse anybody.

", "rawMarkdown": "OK, it is an appropriate clarification. I believe that you really don't want to confuse anybody.", "votes": null }, { "id": "1691079", "postDate": "02/15/2022 08:21:40", "content": "

Beat it by itself. Haha！

", "rawMarkdown": "Beat it by itself. Haha！", "votes": null } ], "comments": [ { "id": 1681114, "author_name": "dragonzhang", "author_url": "", "post_date": "02/08/2022 08:47:18", "content": "

too many high scores on the LB.

\n

either better models or better parameters or both.

", "votes": null, "replies": [ { "id": 1681119, "author_name": "kagglemaniak", "author_url": "", "post_date": "02/08/2022 08:50:12", "content": "

@dragonzhang thank you for replying - how do you find better parameters (CONF and IOU)?

", "votes": null, "replies": [] }, { "id": 1681183, "author_name": "dragonzhang", "author_url": "", "post_date": "02/08/2022 09:47:03", "content": "

just try out your ideas and see what happen.

", "votes": null, "replies": [] } ] }, { "id": 1681169, "author_name": "liyilin14", "author_url": "", "post_date": "02/08/2022 09:31:38", "content": "

I tried track but it does not work

", "votes": null, "replies": [ { "id": 1682138, "author_name": "gjhlove", "author_url": "", "post_date": "02/09/2022 00:29:48", "content": "

tracking actually works, the right tracking boosts the score to 0.672

", "votes": null, "replies": [] }, { "id": 1682536, "author_name": "liyilin14", "author_url": "", "post_date": "02/09/2022 08:27:42", "content": "

You did it with original parameters in the original notebook?

", "votes": null, "replies": [] } ] }, { "id": 1682577, "author_name": "remekkinas", "author_url": "", "post_date": "02/09/2022 08:53:50", "content": "

Yes, we can publish our 0.710 …. 😍😂😆

", "votes": null, "replies": [ { "id": 1682609, "author_name": "dragonzhang", "author_url": "", "post_date": "02/09/2022 09:25:59", "content": "

It is not too late to publish. lol.
\nBut I think overfit in CV is not as that much as in NLP.

", "votes": null, "replies": [] }, { "id": 1682618, "author_name": "remekkinas", "author_url": "", "post_date": "02/09/2022 09:38:04", "content": "

We will see …. 😄

", "votes": null, "replies": [] }, { "id": 1682818, "author_name": "something4kag", "author_url": "", "post_date": "02/09/2022 11:54:01", "content": "

FYI no late sharing during quiet period

", "votes": null, "replies": [] } ] }, { "id": 1682596, "author_name": "deepkim", "author_url": "", "post_date": "02/09/2022 09:06:45", "content": "

I think It's very easy
\nIncrease image size as much as you can

", "votes": null, "replies": [ { "id": 1682625, "author_name": "dragonzhang", "author_url": "", "post_date": "02/09/2022 09:45:09", "content": "

really? Experiments show that It is not linear.

", "votes": null, "replies": [] }, { "id": 1683030, "author_name": "deepkim", "author_url": "", "post_date": "02/09/2022 14:16:42", "content": "

but it increases until reaching some points

", "votes": null, "replies": [] } ] }, { "id": 1683062, "author_name": "tugstugi", "author_url": "", "post_date": "02/09/2022 14:47:55", "content": "

You can use the published public notebook weights and apply tricks and it will score above LB0.700.

", "votes": null, "replies": [ { "id": 1683078, "author_name": "blankaf", "author_url": "", "post_date": "02/09/2022 14:56:57", "content": "

And are you sure that these publicly available weights \"on steroids\" to fit the public LB will be equally useful for final results on private LB? I would be quite careful to give such advises.

", "votes": null, "replies": [] }, { "id": 1683087, "author_name": "tugstugi", "author_url": "", "post_date": "02/09/2022 15:07:47", "content": "

I didn't say anything whether it will be usefull for the private LB. I answered only the question of the topics author: -> you can tune it and get a higher LB score.

", "votes": null, "replies": [] }, { "id": 1683091, "author_name": "blankaf", "author_url": "", "post_date": "02/09/2022 15:14:10", "content": "

OK, it is an appropriate clarification. I believe that you really don't want to confuse anybody.

", "votes": null, "replies": [] } ] }, { "id": 1691079, "author_name": "freshair1996", "author_url": "", "post_date": "02/15/2022 08:21:40", "content": "

Beat it by itself. Haha！

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1681106": "I originally had a good place on the LB (0.654), but after the 0.665 LB notebook was published, I dropped by many places.\n\nI've been trying to fine-tune IOU or CONF in the 0.665 notebook, but I can't get higher than 0.665. How do I find the optimum CONF and IOU?\n\nAlso, I noticed that the notebook doesn't use tracking. Would tracking boost LB?", "1681114": "too many high scores on the LB.\n\neither better models or better parameters or both.", "1681119": "dragonzhang thank you for replying - how do you find better parameters (CONF and IOU)?", "1681169": "I tried track but it does not work", "1681183": "just try out your ideas and see what happen.", "1682138": "tracking actually works, the right tracking boosts the score to 0.672", "1682536": "You did it with original parameters in the original notebook?", "1682577": "Yes, we can publish our 0.710 .... 😍😂😆", "1682596": "I think It's very easy\nIncrease image size as much as you can", "1682609": "It is not too late to publish. lol.\nBut I think overfit in CV is not as that much as in NLP.", "1682618": "We will see .... 😄", "1682625": "really? Experiments show that It is not linear.", "1682818": "[FYI](https://www.kaggle.com/general/291540) no late sharing during quiet period", "1683030": "but it increases until reaching some points", "1683062": "You can use the published public notebook weights and apply tricks and it will score above LB0.700.", "1683078": "And are you sure that these publicly available weights \"on steroids\" to fit the public LB will be equally useful for final results on private LB? I would be quite careful to give such advises.", "1683087": "I didn't say anything whether it will be usefull for the private LB. I answered only the question of the topics author: -> you can tune it and get a higher LB score.", "1683091": "OK, it is an appropriate clarification. I believe that you really don't want to confuse anybody.", "1691079": "Beat it by itself. Haha！" }, "source": "meta" }