{ "id": 229515, "title": "Test Images manipulation?", "url": "/competitions/bms-molecular-translation/discussion/229515", "author_name": "", "post_date": "2021-03-30T14:10:17.726146400Z", "votes": 2, "comment_count": 9, "views": 0, "content": "

Hi everyone,
\nCan we convert the whole test set beforehand using rdkit and make the final predictions on them?
\nIf we cannot do so can we convert them on the fly or is this also not allowed??

\n

I am very confused any information regarding this will be really helpful.
\nThank you :)

", "messages": [ { "id": "1257088", "postDate": "03/30/2021 14:10:17", "content": "

Hi everyone,
\nCan we convert the whole test set beforehand using rdkit and make the final predictions on them?
\nIf we cannot do so can we convert them on the fly or is this also not allowed??

\n

I am very confused any information regarding this will be really helpful.
\nThank you :)

", "rawMarkdown": "Hi everyone, \nCan we convert the whole test set beforehand using rdkit and make the final predictions on them?\nIf we cannot do so can we convert them on the fly or is this also not allowed??\n\nI am very confused any information regarding this will be really helpful.\nThank you :)", "votes": null }, { "id": "1257180", "postDate": "03/30/2021 15:36:26", "content": "

As far as I am aware the 1.6 available test images contain both the public leaderboard images, as well as the private leaderboard images. You submit all of them and 25% is used for the public leaderboard and the other 75% will be used for the private leaderboard. You are thus currently making the final predictions for both the public and private leaderboard, but you can currently only see your public score.

\n
This leaderboard is calculated with approximately 25% of the test data.\nThe final results will be based on the other 75%, so the final standings may be different.\n
", "rawMarkdown": "As far as I am aware the 1.6 available test images contain both the public leaderboard images, as well as the private leaderboard images. You submit all of them and 25% is used for the public leaderboard and the other 75% will be used for the private leaderboard. You are thus currently making the final predictions for both the public and private leaderboard, but you can currently only see your public score.\n\n```\nThis leaderboard is calculated with approximately 25% of the test data.\nThe final results will be based on the other 75%, so the final standings may be different.\n```", "votes": null }, { "id": "1257187", "postDate": "03/30/2021 15:43:51", "content": "

So can I convert them using rdkit or not?

", "rawMarkdown": "So can I convert them using rdkit or not?", "votes": null }, { "id": "1257260", "postDate": "03/30/2021 16:53:27", "content": "

yes you can.

\n

Edit: You will be not breaking the rule if you mention RDKit in external data.

", "rawMarkdown": "yes you can. \n\nEdit: You will be not breaking the rule if you mention `RDKit` in external data.", "votes": null }, { "id": "1257283", "postDate": "03/30/2021 17:20:12", "content": "

Great thank you @drhabib

", "rawMarkdown": "Great thank you @drhabib", "votes": null }, { "id": "1257363", "postDate": "03/30/2021 18:52:39", "content": "

I understand converting the training set using rdkit by feeding it the InChI strings, but how do you convert the test set images using rdkit? I feel like I'm missing something. Can you elaborate?

", "rawMarkdown": "I understand converting the training set using rdkit by feeding it the InChI strings, but how do you convert the test set images using rdkit? I feel like I'm missing something. Can you elaborate?", "votes": null }, { "id": "1257676", "postDate": "03/31/2021 03:08:54", "content": "

I thought about it too @jjinho…
\nI was assuming rdkit has a function where we can pass in image to get back another image with rdkit representation. But I don't think such a function exists. :(

", "rawMarkdown": "I thought about it too @jjinho...\nI was assuming rdkit has a function where we can pass in image to get back another image with rdkit representation. But I don't think such a function exists. :(", "votes": null }, { "id": "1257693", "postDate": "03/31/2021 03:25:01", "content": "

@nitindatta kind of disappointed! Thought you had found a way to do that, because rdkit would basically solve this problem (and maybe that would explain the amazing scores at the top of the leaderboard…) =) Although when @drhabib said \"yes you can\" I thought that was implying this is what he might be doing

", "rawMarkdown": "nitindatta kind of disappointed! Thought you had found a way to do that, because rdkit would basically solve this problem (and maybe that would explain the amazing scores at the top of the leaderboard...) =) Although when @drhabib said \"yes you can\" I thought that was implying this is what he might be doing", "votes": null }, { "id": "1257713", "postDate": "03/31/2021 03:52:06", "content": "

@jjinho , If I find a way I will surely share it here :)

", "rawMarkdown": "jjinho , If I find a way I will surely share it here :)", "votes": null }, { "id": "1257960", "postDate": "03/31/2021 08:36:24", "content": "

But in order for rdkit to create another image of a molecule, it has to recognize it first.

\n\n

So your's cannot be better then rdkit's prediction. (Or any library's that can make this image conversion.)

\n

If you find a pretty good library, then it may be interesting to generate image for ensemble, though, I guess.

\n

Am I missing something?

", "rawMarkdown": "But in order for rdkit to create another image of a molecule, it has to recognize it first.\n - If this recognition is perfect, then just use the InChI that rdkit produces, don't make another image for your model to predict on.\n - If this recognition is not perfect, then the generated image will carry on this error and even if your model is perfect, you will predict the wrong molecule's InChI.\n\nSo your's cannot be better then rdkit's prediction. (Or any library's that can make this image conversion.)\n\nIf you find a pretty good library, then it may be interesting to generate image for ensemble, though, I guess.\n\nAm I missing something?", "votes": null } ], "comments": [ { "id": 1257180, "author_name": "markwijkhuizen", "author_url": "", "post_date": "03/30/2021 15:36:26", "content": "

As far as I am aware the 1.6 available test images contain both the public leaderboard images, as well as the private leaderboard images. You submit all of them and 25% is used for the public leaderboard and the other 75% will be used for the private leaderboard. You are thus currently making the final predictions for both the public and private leaderboard, but you can currently only see your public score.

\n
This leaderboard is calculated with approximately 25% of the test data.\nThe final results will be based on the other 75%, so the final standings may be different.\n
", "votes": null, "replies": [ { "id": 1257187, "author_name": "nitindatta", "author_url": "", "post_date": "03/30/2021 15:43:51", "content": "

So can I convert them using rdkit or not?

", "votes": null, "replies": [] }, { "id": 1257260, "author_name": "drhabib", "author_url": "", "post_date": "03/30/2021 16:53:27", "content": "

yes you can.

\n

Edit: You will be not breaking the rule if you mention RDKit in external data.

", "votes": null, "replies": [] }, { "id": 1257283, "author_name": "nitindatta", "author_url": "", "post_date": "03/30/2021 17:20:12", "content": "

Great thank you @drhabib

", "votes": null, "replies": [] }, { "id": 1257363, "author_name": "jjinho", "author_url": "", "post_date": "03/30/2021 18:52:39", "content": "

I understand converting the training set using rdkit by feeding it the InChI strings, but how do you convert the test set images using rdkit? I feel like I'm missing something. Can you elaborate?

", "votes": null, "replies": [] }, { "id": 1257676, "author_name": "nitindatta", "author_url": "", "post_date": "03/31/2021 03:08:54", "content": "

I thought about it too @jjinho…
\nI was assuming rdkit has a function where we can pass in image to get back another image with rdkit representation. But I don't think such a function exists. :(

", "votes": null, "replies": [] }, { "id": 1257693, "author_name": "jjinho", "author_url": "", "post_date": "03/31/2021 03:25:01", "content": "

@nitindatta kind of disappointed! Thought you had found a way to do that, because rdkit would basically solve this problem (and maybe that would explain the amazing scores at the top of the leaderboard…) =) Although when @drhabib said \"yes you can\" I thought that was implying this is what he might be doing

", "votes": null, "replies": [] }, { "id": 1257713, "author_name": "nitindatta", "author_url": "", "post_date": "03/31/2021 03:52:06", "content": "

@jjinho , If I find a way I will surely share it here :)

", "votes": null, "replies": [] }, { "id": 1257960, "author_name": "nofreewill", "author_url": "", "post_date": "03/31/2021 08:36:24", "content": "

But in order for rdkit to create another image of a molecule, it has to recognize it first.

\n\n

So your's cannot be better then rdkit's prediction. (Or any library's that can make this image conversion.)

\n

If you find a pretty good library, then it may be interesting to generate image for ensemble, though, I guess.

\n

Am I missing something?

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1257088": "Hi everyone, \nCan we convert the whole test set beforehand using rdkit and make the final predictions on them?\nIf we cannot do so can we convert them on the fly or is this also not allowed??\n\nI am very confused any information regarding this will be really helpful.\nThank you :)", "1257180": "As far as I am aware the 1.6 available test images contain both the public leaderboard images, as well as the private leaderboard images. You submit all of them and 25% is used for the public leaderboard and the other 75% will be used for the private leaderboard. You are thus currently making the final predictions for both the public and private leaderboard, but you can currently only see your public score.\n\n```\nThis leaderboard is calculated with approximately 25% of the test data.\nThe final results will be based on the other 75%, so the final standings may be different.\n```", "1257187": "So can I convert them using rdkit or not?", "1257260": "yes you can. \n\nEdit: You will be not breaking the rule if you mention `RDKit` in external data.", "1257283": "Great thank you @drhabib", "1257363": "I understand converting the training set using rdkit by feeding it the InChI strings, but how do you convert the test set images using rdkit? I feel like I'm missing something. Can you elaborate?", "1257676": "I thought about it too @jjinho...\nI was assuming rdkit has a function where we can pass in image to get back another image with rdkit representation. But I don't think such a function exists. :(", "1257693": "nitindatta kind of disappointed! Thought you had found a way to do that, because rdkit would basically solve this problem (and maybe that would explain the amazing scores at the top of the leaderboard...) =) Although when @drhabib said \"yes you can\" I thought that was implying this is what he might be doing", "1257713": "jjinho , If I find a way I will surely share it here :)", "1257960": "But in order for rdkit to create another image of a molecule, it has to recognize it first.\n - If this recognition is perfect, then just use the InChI that rdkit produces, don't make another image for your model to predict on.\n - If this recognition is not perfect, then the generated image will carry on this error and even if your model is perfect, you will predict the wrong molecule's InChI.\n\nSo your's cannot be better then rdkit's prediction. (Or any library's that can make this image conversion.)\n\nIf you find a pretty good library, then it may be interesting to generate image for ensemble, though, I guess.\n\nAm I missing something?" }, "source": "meta" }