{ "id": 220473, "title": "Can't wait to see 1st team's solution ", "url": "/competitions/cassava-leaf-disease-classification/discussion/220473", "author_name": "", "post_date": "2021-02-18T13:03:14.280332100Z", "votes": 32, "comment_count": 29, "views": 0, "content": "

How they got such a huge leading +0.005 above the second, while the rest of us stuck together for months.

\n

Which is too excellent for me to understand how.

\n

Good luck guys, BTW

", "messages": [ { "id": "1208720", "postDate": "02/18/2021 13:03:14", "content": "

How they got such a huge leading +0.005 above the second, while the rest of us stuck together for months.

\n

Which is too excellent for me to understand how.

\n

Good luck guys, BTW

", "rawMarkdown": "How they got such a huge leading +0.005 above the second, while the rest of us stuck together for months.\n\nWhich is too excellent for me to understand how.\n\nGood luck guys, BTW", "votes": null }, { "id": "1208725", "postDate": "02/18/2021 13:07:31", "content": "

It will be nice to see the First place solution after the competition ends, How they managed to get past the impasse of 0.90 scores.

", "rawMarkdown": "It will be nice to see the First place solution after the competition ends, How they managed to get past the impasse of 0.90 scores.", "votes": null }, { "id": "1208777", "postDate": "02/18/2021 13:34:54", "content": "

I also can't wait to see the solution your team submitted~

", "rawMarkdown": "I also can't wait to see the solution your team submitted~", "votes": null }, { "id": "1208786", "postDate": "02/18/2021 13:40:32", "content": "

I am more interested in if any team managed to have reliable CV/LB. If so, how?

", "rawMarkdown": "I am more interested in if any team managed to have reliable CV/LB. If so, how?", "votes": null }, { "id": "1208863", "postDate": "02/18/2021 14:17:58", "content": "

A shake up is coming!

", "rawMarkdown": "A shake up is coming!", "votes": null }, { "id": "1209025", "postDate": "02/18/2021 16:30:09", "content": "

We are failed to get such high score even in one fold, once one fold achieve 0.910+, other fold get much worse.
\nLooking forward to read their and your write up!

", "rawMarkdown": "We are failed to get such high score even in one fold, once one fold achieve 0.910+, other fold get much worse.\nLooking forward to read their and your write up!", "votes": null }, { "id": "1209033", "postDate": "02/18/2021 16:38:44", "content": "

Really interesting to see the choice of models and augmentation as to what led to 0.915 whereas improving by .001 is too difficult after .90

", "rawMarkdown": "Really interesting to see the choice of models and augmentation as to what led to 0.915 whereas improving by .001 is too difficult after .90", "votes": null }, { "id": "1209070", "postDate": "02/18/2021 17:01:30", "content": "

shake-up is coming and 1st solution maybe yours :))

", "rawMarkdown": "shake-up is coming and 1st solution maybe yours :))", "votes": null }, { "id": "1209082", "postDate": "02/18/2021 17:11:58", "content": "

Very unlikely. I believe they got something fundamentally vital. Most shakes shall under +- 0.004

", "rawMarkdown": "Very unlikely. I believe they got something fundamentally vital. Most shakes shall under +- 0.004", "votes": null }, { "id": "1209086", "postDate": "02/18/2021 17:16:22", "content": "

The highest one fold single model score we got is 0.9051, did not try ensemble test.

", "rawMarkdown": "The highest one fold single model score we got is 0.9051, did not try ensemble test.", "votes": null }, { "id": "1209104", "postDate": "02/18/2021 17:23:19", "content": "

@zlanan Any tips for 2 final submission or let the system choose 2 best lbs?

", "rawMarkdown": "zlanan Any tips for 2 final submission or let the system choose 2 best lbs?", "votes": null }, { "id": "1209110", "postDate": "02/18/2021 17:30:02", "content": "

This competition is really tricky, failed to confirm many models due to the week correlation between CV and LB.
\nIt would be a common practice that: one focus on public LB only, and the other focus on the CV only regardless of it's LB score.

", "rawMarkdown": "This competition is really tricky, failed to confirm many models due to the week correlation between CV and LB.\nIt would be a common practice that: one focus on public LB only, and the other focus on the CV only regardless of it's LB score.", "votes": null }, { "id": "1209112", "postDate": "02/18/2021 17:31:16", "content": "

Make sure that your 2 submissions are significantly different yet have good CV/LB. That'll increase your chances of surviving shakeup. This strategy has helped me in multiple past competitions.

", "rawMarkdown": "Make sure that your 2 submissions are significantly different yet have good CV/LB. That'll increase your chances of surviving shakeup. This strategy has helped me in multiple past competitions.", "votes": null }, { "id": "1209114", "postDate": "02/18/2021 17:33:08", "content": "

KS is right, you shall select a quite different version, not a second version only with minor changes based on the first.

", "rawMarkdown": "KS is right, you shall select a quite different version, not a second version only with minor changes based on the first.", "votes": null }, { "id": "1209115", "postDate": "02/18/2021 17:34:34", "content": "

Thanks for the info, the final score will be best of the two? right

", "rawMarkdown": "Thanks for the info, the final score will be best of the two? right", "votes": null }, { "id": "1209123", "postDate": "02/18/2021 17:37:21", "content": "

Yes, the higher one is your final score.

", "rawMarkdown": "Yes, the higher one is your final score.", "votes": null }, { "id": "1209124", "postDate": "02/18/2021 17:37:39", "content": "

Shake up may convert a gold into silver and a silver into bronze or vice versa

", "rawMarkdown": "Shake up may convert a gold into silver and a silver into bronze or vice versa", "votes": null }, { "id": "1209139", "postDate": "02/18/2021 17:52:52", "content": "

Same here. Really looking forward to see first person solution.

", "rawMarkdown": "Same here. Really looking forward to see first person solution.", "votes": null }, { "id": "1209159", "postDate": "02/18/2021 18:11:59", "content": "

I don't think Kaggle rerun the inference notebooks on all data again after submission deadline but instead it run only once on all data after we submit before the deadline and store our 2 type of results but display our result on a percentage of data called as public test data before the deadline and display our result on private test data after the deadline. Right?

", "rawMarkdown": "I don't think Kaggle rerun the inference notebooks on all data again after submission deadline but instead it run only once on all data after we submit before the deadline and store our 2 type of results but display our result on a percentage of data called as public test data before the deadline and display our result on private test data after the deadline. Right?", "votes": null }, { "id": "1209199", "postDate": "02/18/2021 18:48:36", "content": "

Yes @vickygoyal Kaggle already knows our public and private LB score for each submission. So when the competition ends at 11:59pm tonight, the final private LB results will be displayed immediately.

", "rawMarkdown": "Yes @vickygoyal Kaggle already knows our public and private LB score for each submission. So when the competition ends at 11:59pm tonight, the final private LB results will be displayed immediately.", "votes": null }, { "id": "1209240", "postDate": "02/18/2021 19:12:04", "content": "

Usually what you do is:

\n
    \n
  1. Submission: Highest CV
    2. \n
  2. Submission: Highest LB
    3. \n
\n

In my experience it always went out well for me.

", "rawMarkdown": "Usually what you do is:\n\n1. Submission: Highest CV \n2. Submission: Highest LB\n\nIn my experience it always went out well for me.", "votes": null }, { "id": "1209242", "postDate": "02/18/2021 19:12:23", "content": "

@cdeotte Thanks for the information, have my fingers crossed for the results :).

", "rawMarkdown": "cdeotte Thanks for the information, have my fingers crossed for the results :).", "votes": null }, { "id": "1209365", "postDate": "02/18/2021 21:12:34", "content": "

Me too, i asking 🤒

", "rawMarkdown": "Me too, i asking 🤒", "votes": null }, { "id": "1209518", "postDate": "02/18/2021 23:59:48", "content": "

I'm counting down to learning your excellent solutions.

", "rawMarkdown": "I'm counting down to learning your excellent solutions.", "votes": null }, { "id": "1209526", "postDate": "02/19/2021 00:09:36", "content": "

I remembered in another thread you said your best cv is 0.919, that is the guide of my experiment…

", "rawMarkdown": "I remembered in another thread you said your best cv is 0.919, that is the guide of my experiment...", "votes": null }, { "id": "1209528", "postDate": "02/19/2021 00:12:23", "content": "

I have a 903 score in private leaderboard but not selected……

", "rawMarkdown": "I have a 903 score in private leaderboard but not selected......", "votes": null }, { "id": "1209568", "postDate": "02/19/2021 00:35:50", "content": "

Hi @aerdem4

\n

Resnet200d with Cutmix and Fmix worked well for me in this competition. The CV/ LB scores were reliable.

\n

Notebook: https://www.kaggle.com/jagadish13/pytorch-resnet200d-baseline-inference-tta/output

", "rawMarkdown": "Hi @aerdem4 \n\nResnet200d with Cutmix and Fmix worked well for me in this competition. The CV/ LB scores were reliable.\n\nNotebook: https://www.kaggle.com/jagadish13/pytorch-resnet200d-baseline-inference-tta/output", "votes": null }, { "id": "1209598", "postDate": "02/19/2021 00:55:53", "content": "

That is same fold self-ensemble + TTA, here is a single model without tta.

", "rawMarkdown": "That is same fold self-ensemble + TTA, here is a single model without tta.", "votes": null }, { "id": "1210139", "postDate": "02/19/2021 08:16:46", "content": "

Bro, sorry to hear that

", "rawMarkdown": "Bro, sorry to hear that", "votes": null }, { "id": "1210982", "postDate": "02/19/2021 21:11:28", "content": "

Me to. I suspect it was some kind of relabelling or data cleaning, but I really want to know!

", "rawMarkdown": "Me to. I suspect it was some kind of relabelling or data cleaning, but I really want to know!", "votes": null } ], "comments": [ { "id": 1208725, "author_name": "mohneesh7", "author_url": "", "post_date": "02/18/2021 13:07:31", "content": "

It will be nice to see the First place solution after the competition ends, How they managed to get past the impasse of 0.90 scores.

", "votes": null, "replies": [] }, { "id": 1208777, "author_name": "deepkim", "author_url": "", "post_date": "02/18/2021 13:34:54", "content": "

I also can't wait to see the solution your team submitted~

", "votes": null, "replies": [] }, { "id": 1208786, "author_name": "aerdem4", "author_url": "", "post_date": "02/18/2021 13:40:32", "content": "

I am more interested in if any team managed to have reliable CV/LB. If so, how?

", "votes": null, "replies": [ { "id": 1209568, "author_name": "jagadish13", "author_url": "", "post_date": "02/19/2021 00:35:50", "content": "

Hi @aerdem4

\n

Resnet200d with Cutmix and Fmix worked well for me in this competition. The CV/ LB scores were reliable.

\n

Notebook: https://www.kaggle.com/jagadish13/pytorch-resnet200d-baseline-inference-tta/output

", "votes": null, "replies": [] } ] }, { "id": 1208863, "author_name": "cuteffff", "author_url": "", "post_date": "02/18/2021 14:17:58", "content": "

A shake up is coming!

", "votes": null, "replies": [] }, { "id": 1209025, "author_name": "steamedsheep", "author_url": "", "post_date": "02/18/2021 16:30:09", "content": "

We are failed to get such high score even in one fold, once one fold achieve 0.910+, other fold get much worse.
\nLooking forward to read their and your write up!

", "votes": null, "replies": [ { "id": 1209086, "author_name": "zlanan", "author_url": "", "post_date": "02/18/2021 17:16:22", "content": "

The highest one fold single model score we got is 0.9051, did not try ensemble test.

", "votes": null, "replies": [] }, { "id": 1209526, "author_name": "mrxuehb", "author_url": "", "post_date": "02/19/2021 00:09:36", "content": "

I remembered in another thread you said your best cv is 0.919, that is the guide of my experiment…

", "votes": null, "replies": [] }, { "id": 1209598, "author_name": "zlanan", "author_url": "", "post_date": "02/19/2021 00:55:53", "content": "

That is same fold self-ensemble + TTA, here is a single model without tta.

", "votes": null, "replies": [] } ] }, { "id": 1209033, "author_name": "vickygoyal", "author_url": "", "post_date": "02/18/2021 16:38:44", "content": "

Really interesting to see the choice of models and augmentation as to what led to 0.915 whereas improving by .001 is too difficult after .90

", "votes": null, "replies": [] }, { "id": 1209070, "author_name": "hungkhoi", "author_url": "", "post_date": "02/18/2021 17:01:30", "content": "

shake-up is coming and 1st solution maybe yours :))

", "votes": null, "replies": [ { "id": 1209082, "author_name": "zlanan", "author_url": "", "post_date": "02/18/2021 17:11:58", "content": "

Very unlikely. I believe they got something fundamentally vital. Most shakes shall under +- 0.004

", "votes": null, "replies": [] }, { "id": 1209124, "author_name": "vickygoyal", "author_url": "", "post_date": "02/18/2021 17:37:39", "content": "

Shake up may convert a gold into silver and a silver into bronze or vice versa

", "votes": null, "replies": [] } ] }, { "id": 1209104, "author_name": "vickygoyal", "author_url": "", "post_date": "02/18/2021 17:23:19", "content": "

@zlanan Any tips for 2 final submission or let the system choose 2 best lbs?

", "votes": null, "replies": [ { "id": 1209110, "author_name": "zlanan", "author_url": "", "post_date": "02/18/2021 17:30:02", "content": "

This competition is really tricky, failed to confirm many models due to the week correlation between CV and LB.
\nIt would be a common practice that: one focus on public LB only, and the other focus on the CV only regardless of it's LB score.

", "votes": null, "replies": [] }, { "id": 1209112, "author_name": "ks2019", "author_url": "", "post_date": "02/18/2021 17:31:16", "content": "

Make sure that your 2 submissions are significantly different yet have good CV/LB. That'll increase your chances of surviving shakeup. This strategy has helped me in multiple past competitions.

", "votes": null, "replies": [] }, { "id": 1209114, "author_name": "zlanan", "author_url": "", "post_date": "02/18/2021 17:33:08", "content": "

KS is right, you shall select a quite different version, not a second version only with minor changes based on the first.

", "votes": null, "replies": [] }, { "id": 1209115, "author_name": "vickygoyal", "author_url": "", "post_date": "02/18/2021 17:34:34", "content": "

Thanks for the info, the final score will be best of the two? right

", "votes": null, "replies": [] }, { "id": 1209123, "author_name": "zlanan", "author_url": "", "post_date": "02/18/2021 17:37:21", "content": "

Yes, the higher one is your final score.

", "votes": null, "replies": [] }, { "id": 1209159, "author_name": "vickygoyal", "author_url": "", "post_date": "02/18/2021 18:11:59", "content": "

I don't think Kaggle rerun the inference notebooks on all data again after submission deadline but instead it run only once on all data after we submit before the deadline and store our 2 type of results but display our result on a percentage of data called as public test data before the deadline and display our result on private test data after the deadline. Right?

", "votes": null, "replies": [] }, { "id": 1209199, "author_name": "cdeotte", "author_url": "", "post_date": "02/18/2021 18:48:36", "content": "

Yes @vickygoyal Kaggle already knows our public and private LB score for each submission. So when the competition ends at 11:59pm tonight, the final private LB results will be displayed immediately.

", "votes": null, "replies": [] }, { "id": 1209240, "author_name": "aliabdin1", "author_url": "", "post_date": "02/18/2021 19:12:04", "content": "

Usually what you do is:

\n
    \n
  1. Submission: Highest CV
    2. \n
  2. Submission: Highest LB
    3. \n
\n

In my experience it always went out well for me.

", "votes": null, "replies": [] }, { "id": 1209242, "author_name": "vickygoyal", "author_url": "", "post_date": "02/18/2021 19:12:23", "content": "

@cdeotte Thanks for the information, have my fingers crossed for the results :).

", "votes": null, "replies": [] } ] }, { "id": 1209139, "author_name": "derivativefno", "author_url": "", "post_date": "02/18/2021 17:52:52", "content": "

Same here. Really looking forward to see first person solution.

", "votes": null, "replies": [] }, { "id": 1209365, "author_name": "milobele", "author_url": "", "post_date": "02/18/2021 21:12:34", "content": "

Me too, i asking 🤒

", "votes": null, "replies": [] }, { "id": 1209518, "author_name": "komakizzz", "author_url": "", "post_date": "02/18/2021 23:59:48", "content": "

I'm counting down to learning your excellent solutions.

", "votes": null, "replies": [] }, { "id": 1209528, "author_name": "qixinyan", "author_url": "", "post_date": "02/19/2021 00:12:23", "content": "

I have a 903 score in private leaderboard but not selected……

", "votes": null, "replies": [ { "id": 1210139, "author_name": "jsrdcht", "author_url": "", "post_date": "02/19/2021 08:16:46", "content": "

Bro, sorry to hear that

", "votes": null, "replies": [] } ] }, { "id": 1210982, "author_name": "eschibli", "author_url": "", "post_date": "02/19/2021 21:11:28", "content": "

Me to. I suspect it was some kind of relabelling or data cleaning, but I really want to know!

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1208720": "How they got such a huge leading +0.005 above the second, while the rest of us stuck together for months.\n\nWhich is too excellent for me to understand how.\n\nGood luck guys, BTW", "1208725": "It will be nice to see the First place solution after the competition ends, How they managed to get past the impasse of 0.90 scores.", "1208777": "I also can't wait to see the solution your team submitted~", "1208786": "I am more interested in if any team managed to have reliable CV/LB. If so, how?", "1208863": "A shake up is coming!", "1209025": "We are failed to get such high score even in one fold, once one fold achieve 0.910+, other fold get much worse.\nLooking forward to read their and your write up!", "1209033": "Really interesting to see the choice of models and augmentation as to what led to 0.915 whereas improving by .001 is too difficult after .90", "1209070": "shake-up is coming and 1st solution maybe yours :))", "1209082": "Very unlikely. I believe they got something fundamentally vital. Most shakes shall under +- 0.004", "1209086": "The highest one fold single model score we got is 0.9051, did not try ensemble test.", "1209104": "zlanan Any tips for 2 final submission or let the system choose 2 best lbs?", "1209110": "This competition is really tricky, failed to confirm many models due to the week correlation between CV and LB.\nIt would be a common practice that: one focus on public LB only, and the other focus on the CV only regardless of it's LB score.", "1209112": "Make sure that your 2 submissions are significantly different yet have good CV/LB. That'll increase your chances of surviving shakeup. This strategy has helped me in multiple past competitions.", "1209114": "KS is right, you shall select a quite different version, not a second version only with minor changes based on the first.", "1209115": "Thanks for the info, the final score will be best of the two? right", "1209123": "Yes, the higher one is your final score.", "1209124": "Shake up may convert a gold into silver and a silver into bronze or vice versa", "1209139": "Same here. Really looking forward to see first person solution.", "1209159": "I don't think Kaggle rerun the inference notebooks on all data again after submission deadline but instead it run only once on all data after we submit before the deadline and store our 2 type of results but display our result on a percentage of data called as public test data before the deadline and display our result on private test data after the deadline. Right?", "1209199": "Yes @vickygoyal Kaggle already knows our public and private LB score for each submission. So when the competition ends at 11:59pm tonight, the final private LB results will be displayed immediately.", "1209240": "Usually what you do is:\n\n1. Submission: Highest CV \n2. Submission: Highest LB\n\nIn my experience it always went out well for me.", "1209242": "cdeotte Thanks for the information, have my fingers crossed for the results :).", "1209365": "Me too, i asking 🤒", "1209518": "I'm counting down to learning your excellent solutions.", "1209526": "I remembered in another thread you said your best cv is 0.919, that is the guide of my experiment...", "1209528": "I have a 903 score in private leaderboard but not selected......", "1209568": "Hi @aerdem4 \n\nResnet200d with Cutmix and Fmix worked well for me in this competition. The CV/ LB scores were reliable.\n\nNotebook: https://www.kaggle.com/jagadish13/pytorch-resnet200d-baseline-inference-tta/output", "1209598": "That is same fold self-ensemble + TTA, here is a single model without tta.", "1210139": "Bro, sorry to hear that", "1210982": "Me to. I suspect it was some kind of relabelling or data cleaning, but I really want to know!" }, "source": "meta" }