I can see from public kernels, also my own kernels that all solutions are constant strings 😃
\nWas anyone successful in development solutions with non-constant outputs? (the score about 75+)
", "messages": [ { "id": 1227392, "postDate": "2021-03-05T13:40:57.357Z", "content": "I can see from public kernels, also my own kernels that all solutions are constant strings 😃
\nWas anyone successful in development solutions with non-constant outputs? (the score about 75+)
", "rawMarkdown": "I can see from public kernels, also my own kernels that all solutions are constant strings 😃\n\nWas anyone successful in development solutions with non-constant outputs? (the score about 75+)", "votes": 10 }, { "id": 1227437, "postDate": "2021-03-05T14:25:48.233Z", "content": "Yes, but only the chemical formula sublayer.
", "rawMarkdown": "Yes, but only the chemical formula sublayer.", "votes": 5 }, { "id": 1229485, "postDate": "2021-03-07T11:33:32.733Z", "content": "My notebook here gives a solution with non-constant strings: https://www.kaggle.com/pasewark/pytorch-resnet-lstm-with-attention
", "rawMarkdown": "My notebook here gives a solution with non-constant strings: https://www.kaggle.com/pasewark/pytorch-resnet-lstm-with-attention", "votes": 3 }, { "id": 1227857, "postDate": "2021-03-05T21:46:01.727Z", "content": "Working on it… the competition has only been going 3 days.
", "rawMarkdown": "Working on it... the competition has only been going 3 days.", "votes": 3 }, { "id": 1228793, "postDate": "2021-03-06T19:10:01.930Z", "content": "Not yet, stuff like my 68 and somewhat better are clearly also all constant strings. I found the constant string idea intriguing, so did one, but will finish RANZR CLiP first before deciding to have a proper try on the other X-ray one or this one. In sort of guessing that to get to 4x people are doing a stratified version of constant strings (e.g. stratified large versus small molecules), but would be interesting to hear whether that's it.
\nI kind of find this particular problem pretty fascinating (enough so, to invest some time to understand the data and do some EDA to get my mind thinking on how to model this…) due to the lack of prepackaged solutions that you can just apply out of the box. So, there'll be interesting opportunities to learn new stuff and be creative.
", "rawMarkdown": "Not yet, stuff like my 68 and somewhat better are clearly also all constant strings. I found the constant string idea intriguing, so did one, but will finish RANZR CLiP first before deciding to have a proper try on the other X-ray one or this one. In sort of guessing that to get to 4x people are doing a stratified version of constant strings (e.g. stratified large versus small molecules), but would be interesting to hear whether that's it.\n\nI kind of find this particular problem pretty fascinating (enough so, to invest some time to [understand the data and do some EDA](https://www.kaggle.com/bjoernholzhauer/bristol-myers-squibb-eda-and-preparing-data) to get my mind thinking on how to model this...) due to the lack of prepackaged solutions that you can just apply out of the box. So, there'll be interesting opportunities to learn new stuff and be creative." }, { "id": 1228036, "postDate": "2021-03-06T04:09:13.910Z", "content": "I believe you can shave a few points by binning the images by the number of black pixels (more black pixels correlates with more atoms in the molecule) and use a different constant string for each bin. I will try not to get distracted by pursuing this solution, but it may be useful as a baseline.
", "rawMarkdown": "I believe you can shave a few points by binning the images by the number of black pixels (more black pixels correlates with more atoms in the molecule) and use a different constant string for each bin. I will try not to get distracted by pursuing this solution, but it may be useful as a baseline." }, { "id": 1227719, "postDate": "2021-03-05T19:41:50.703Z", "content": "What do you mean by (non-)constant?
", "rawMarkdown": "What do you mean by (non-)constant?", "replies": [ { "id": 1227732, "postDate": "2021-03-05T19:56:51.067Z", "content": "@nofreewill ,
\nConstant - all records from test set have the same answer.
\nNon-constant - 1 prediction for 1 row in test set and all of them are different (maybe not all but the majority).
Oh I see. So does your question literally mean that \"has anyone tried to solve the problem yet\"? :D
\nBtw, thank you for the clarification!
Yes, exactly 😁
", "rawMarkdown": "Yes, exactly 😁", "votes": 1 } ] }, { "id": 1227431, "postDate": "2021-03-05T14:19:37.517Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1227437, "author_name": "Johnny Lee", "author_url": "", "post_date": "2021-03-05T14:25:48.233000", "content": "Yes, but only the chemical formula sublayer.
", "votes": 5, "replies": [] }, { "id": 1229485, "author_name": "Eric Pasewark", "author_url": "", "post_date": "2021-03-07T11:33:32.733000", "content": "My notebook here gives a solution with non-constant strings: https://www.kaggle.com/pasewark/pytorch-resnet-lstm-with-attention
", "votes": 3, "replies": [] }, { "id": 1227857, "author_name": "Andy Penrose", "author_url": "", "post_date": "2021-03-05T21:46:01.727000", "content": "Working on it… the competition has only been going 3 days.
", "votes": 3, "replies": [] }, { "id": 1228793, "author_name": "Björn", "author_url": "", "post_date": "2021-03-06T19:10:01.930000", "content": "Not yet, stuff like my 68 and somewhat better are clearly also all constant strings. I found the constant string idea intriguing, so did one, but will finish RANZR CLiP first before deciding to have a proper try on the other X-ray one or this one. In sort of guessing that to get to 4x people are doing a stratified version of constant strings (e.g. stratified large versus small molecules), but would be interesting to hear whether that's it.
\nI kind of find this particular problem pretty fascinating (enough so, to invest some time to understand the data and do some EDA to get my mind thinking on how to model this…) due to the lack of prepackaged solutions that you can just apply out of the box. So, there'll be interesting opportunities to learn new stuff and be creative.
", "votes": 0, "replies": [] }, { "id": 1228036, "author_name": "Paul Jurczak", "author_url": "", "post_date": "2021-03-06T04:09:13.910000", "content": "I believe you can shave a few points by binning the images by the number of black pixels (more black pixels correlates with more atoms in the molecule) and use a different constant string for each bin. I will try not to get distracted by pursuing this solution, but it may be useful as a baseline.
", "votes": 0, "replies": [] }, { "id": 1227719, "author_name": "nofreewill42", "author_url": "", "post_date": "2021-03-05T19:41:50.703000", "content": "What do you mean by (non-)constant?
", "votes": 0, "replies": [ { "id": 1227732, "author_name": "Kostiantyn Isaienkov", "author_url": "", "post_date": "2021-03-05T19:56:51.067000", "content": "@nofreewill ,
\nConstant - all records from test set have the same answer.
\nNon-constant - 1 prediction for 1 row in test set and all of them are different (maybe not all but the majority).
Oh I see. So does your question literally mean that \"has anyone tried to solve the problem yet\"? :D
\nBtw, thank you for the clarification!
Yes, exactly 😁
", "votes": 1, "replies": [] } ] }, { "id": 1227431, "author_name": "", "author_url": "", "post_date": "2021-03-05T14:19:37.517000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1227392": "I can see from public kernels, also my own kernels that all solutions are constant strings 😃\n\nWas anyone successful in development solutions with non-constant outputs? (the score about 75+)", "1227437": "Yes, but only the chemical formula sublayer.", "1229485": "My notebook here gives a solution with non-constant strings: https://www.kaggle.com/pasewark/pytorch-resnet-lstm-with-attention", "1227857": "Working on it... the competition has only been going 3 days.", "1228793": "Not yet, stuff like my 68 and somewhat better are clearly also all constant strings. I found the constant string idea intriguing, so did one, but will finish RANZR CLiP first before deciding to have a proper try on the other X-ray one or this one. In sort of guessing that to get to 4x people are doing a stratified version of constant strings (e.g. stratified large versus small molecules), but would be interesting to hear whether that's it.\n\nI kind of find this particular problem pretty fascinating (enough so, to invest some time to [understand the data and do some EDA](https://www.kaggle.com/bjoernholzhauer/bristol-myers-squibb-eda-and-preparing-data) to get my mind thinking on how to model this...) due to the lack of prepackaged solutions that you can just apply out of the box. So, there'll be interesting opportunities to learn new stuff and be creative.", "1228036": "I believe you can shave a few points by binning the images by the number of black pixels (more black pixels correlates with more atoms in the molecule) and use a different constant string for each bin. I will try not to get distracted by pursuing this solution, but it may be useful as a baseline.", "1227719": "What do you mean by (non-)constant?", "1227431": "" } }