{ "id": 229578, "title": "Normalize your predictions", "url": "/competitions/bms-molecular-translation/discussion/229578", "author_name": "", "post_date": "2021-03-30T21:58:33.666400900Z", "votes": 21, "comment_count": 6, "views": 0, "content": "

I have no idea how to share the code in the \"Code\" section, so here it is:
\nhttps://www.kaggle.com/nofreewill/normalize-your-predictions

", "messages": [ { "id": "1257465", "postDate": "03/30/2021 21:58:33", "content": "

I have no idea how to share the code in the \"Code\" section, so here it is:
\nhttps://www.kaggle.com/nofreewill/normalize-your-predictions

", "rawMarkdown": "I have no idea how to share the code in the \"Code\" section, so here it is:\nhttps://www.kaggle.com/nofreewill/normalize-your-predictions", "votes": null }, { "id": "1257696", "postDate": "03/31/2021 03:28:27", "content": "

Thanks @nofreewill. Can confirm that whatever it says is the difference at the end of the code is approximately the improvement you'll see on LB.

", "rawMarkdown": "Thanks @nofreewill. Can confirm that whatever it says is the difference at the end of the code is approximately the improvement you'll see on LB.", "votes": null }, { "id": "1257806", "postDate": "03/31/2021 05:37:50", "content": "

TLDR - START;
\nThe same molecule can be described with multiple \"valid\" InChIs, but there is only one standard one, that's most probably the label.
\nIf you have predicted a correct valid inchi other than the standard one, then you can normalize it back.
\nTLDR - END;

\n

(Well standardize, I guess. Should I change the title?)

\n

So the improvement cannot be more than this number. And I don't know if it is possible that this makes the score worse.

", "rawMarkdown": "TLDR - START;\nThe same molecule can be described with multiple \"valid\" InChIs, but there is only one standard one, that's most probably the label.\nIf you have predicted a correct valid inchi other than the standard one, then you can normalize it back.\nTLDR - END;\n\n(Well standardize, I guess. Should I change the title?)\n\nSo the improvement cannot be more than this number. And I don't know if it is possible that this makes the score worse.", "votes": null }, { "id": "1258194", "postDate": "03/31/2021 12:37:21", "content": "

i wonder is it a good idea to train different models using different representations and use them for ensemble?
\ne.g. train another model on canoical SMILES

", "rawMarkdown": "i wonder is it a good idea to train different models using different representations and use them for ensemble?\ne.g. train another model on canoical SMILES", "votes": null }, { "id": "1258202", "postDate": "03/31/2021 12:45:28", "content": "

Interesting idea.
\nI'm not sure how you could translate them to a common language, though, where you can ensemble them.
\nBut I haven't looked into what SMILES really is, yet.

", "rawMarkdown": "Interesting idea.\nI'm not sure how you could translate them to a common language, though, where you can ensemble them.\nBut I haven't looked into what SMILES really is, yet.", "votes": null }, { "id": "1258260", "postDate": "03/31/2021 13:42:18", "content": "

Thanks @nofreewill
\nI rewrote your notebook, and made it can run on kaggle. Unfortunately, the LB score got worse. ~-0.02
\nhttps://www.kaggle.com/wuliaokaola/bmsmt-0331-normalize-your-predictions

", "rawMarkdown": "Thanks @nofreewill \nI rewrote your notebook, and made it can run on kaggle. Unfortunately, the LB score got worse. ~-0.02\n[https://www.kaggle.com/wuliaokaola/bmsmt-0331-normalize-your-predictions](https://www.kaggle.com/wuliaokaola/bmsmt-0331-normalize-your-predictions)", "votes": null }, { "id": "1258284", "postDate": "03/31/2021 14:04:53", "content": "

So it is possible that the LB gets worse, then. Thanks for pointing out, in this case I may try my original submission, too.

", "rawMarkdown": "So it is possible that the LB gets worse, then. Thanks for pointing out, in this case I may try my original submission, too.", "votes": null } ], "comments": [ { "id": 1257696, "author_name": "jjinho", "author_url": "", "post_date": "03/31/2021 03:28:27", "content": "

Thanks @nofreewill. Can confirm that whatever it says is the difference at the end of the code is approximately the improvement you'll see on LB.

", "votes": null, "replies": [ { "id": 1257806, "author_name": "nofreewill", "author_url": "", "post_date": "03/31/2021 05:37:50", "content": "

TLDR - START;
\nThe same molecule can be described with multiple \"valid\" InChIs, but there is only one standard one, that's most probably the label.
\nIf you have predicted a correct valid inchi other than the standard one, then you can normalize it back.
\nTLDR - END;

\n

(Well standardize, I guess. Should I change the title?)

\n

So the improvement cannot be more than this number. And I don't know if it is possible that this makes the score worse.

", "votes": null, "replies": [] } ] }, { "id": 1258194, "author_name": "hengck23", "author_url": "", "post_date": "03/31/2021 12:37:21", "content": "

i wonder is it a good idea to train different models using different representations and use them for ensemble?
\ne.g. train another model on canoical SMILES

", "votes": null, "replies": [ { "id": 1258202, "author_name": "nofreewill", "author_url": "", "post_date": "03/31/2021 12:45:28", "content": "

Interesting idea.
\nI'm not sure how you could translate them to a common language, though, where you can ensemble them.
\nBut I haven't looked into what SMILES really is, yet.

", "votes": null, "replies": [] } ] }, { "id": 1258260, "author_name": "wuliaokaola", "author_url": "", "post_date": "03/31/2021 13:42:18", "content": "

Thanks @nofreewill
\nI rewrote your notebook, and made it can run on kaggle. Unfortunately, the LB score got worse. ~-0.02
\nhttps://www.kaggle.com/wuliaokaola/bmsmt-0331-normalize-your-predictions

", "votes": null, "replies": [ { "id": 1258284, "author_name": "nofreewill", "author_url": "", "post_date": "03/31/2021 14:04:53", "content": "

So it is possible that the LB gets worse, then. Thanks for pointing out, in this case I may try my original submission, too.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1257465": "I have no idea how to share the code in the \"Code\" section, so here it is:\nhttps://www.kaggle.com/nofreewill/normalize-your-predictions", "1257696": "Thanks @nofreewill. Can confirm that whatever it says is the difference at the end of the code is approximately the improvement you'll see on LB.", "1257806": "TLDR - START;\nThe same molecule can be described with multiple \"valid\" InChIs, but there is only one standard one, that's most probably the label.\nIf you have predicted a correct valid inchi other than the standard one, then you can normalize it back.\nTLDR - END;\n\n(Well standardize, I guess. Should I change the title?)\n\nSo the improvement cannot be more than this number. And I don't know if it is possible that this makes the score worse.", "1258194": "i wonder is it a good idea to train different models using different representations and use them for ensemble?\ne.g. train another model on canoical SMILES", "1258202": "Interesting idea.\nI'm not sure how you could translate them to a common language, though, where you can ensemble them.\nBut I haven't looked into what SMILES really is, yet.", "1258260": "Thanks @nofreewill \nI rewrote your notebook, and made it can run on kaggle. Unfortunately, the LB score got worse. ~-0.02\n[https://www.kaggle.com/wuliaokaola/bmsmt-0331-normalize-your-predictions](https://www.kaggle.com/wuliaokaola/bmsmt-0331-normalize-your-predictions)", "1258284": "So it is possible that the LB gets worse, then. Thanks for pointing out, in this case I may try my original submission, too." }, "source": "meta" }