In this competition, I thought of treating the problem as a regression task on 2D images. My feeling was that such images would contain enough information concerning the geometry of the molecule so that a CNN would be able to perform better that a tree based model, since in the later, we have an a priory on the features that have to be fed to the model. However, I lack of experience to create efficient CNN architectures. I thus feel I could not get the full juice out of the method and thus want to make it public, to see if others manage to obtain better results.
\n\nIn the kernel creating bond centered 2D images, I show how I create the 2D images for every molecular bond. The code is not efficient and it's thus quite time consuming to create a whole dataset with such images. Such datasets are made available for the 1JHN symetry and part of the 1JHC molecular bonds. A tentative CNN architecture is also given here.
\n\nPS: with the 2D images constructed that way, the coordinates of the pixels are physically meaningful. Such information is lost in the CNN architecture I built. Which kind of architecture would be suitable for that ?? I tried to remove the maxpooling layer or to add filters with fixed coordinates, but this did not seem to improve the results ...
", "messages": [ { "id": "601575", "postDate": "08/17/2019 20:33:47", "content": "In this competition, I thought of treating the problem as a regression task on 2D images. My feeling was that such images would contain enough information concerning the geometry of the molecule so that a CNN would be able to perform better that a tree based model, since in the later, we have an a priory on the features that have to be fed to the model. However, I lack of experience to create efficient CNN architectures. I thus feel I could not get the full juice out of the method and thus want to make it public, to see if others manage to obtain better results.
\n\nIn the kernel creating bond centered 2D images, I show how I create the 2D images for every molecular bond. The code is not efficient and it's thus quite time consuming to create a whole dataset with such images. Such datasets are made available for the 1JHN symetry and part of the 1JHC molecular bonds. A tentative CNN architecture is also given here.
\n\nPS: with the 2D images constructed that way, the coordinates of the pixels are physically meaningful. Such information is lost in the CNN architecture I built. Which kind of architecture would be suitable for that ?? I tried to remove the maxpooling layer or to add filters with fixed coordinates, but this did not seem to improve the results ...
", "rawMarkdown": "In this competition, I thought of treating the problem as a regression task on 2D images. My feeling was that such images would contain enough information concerning the geometry of the molecule so that a CNN would be able to perform better that a tree based model, since in the later, we have an a priory on the features that have to be fed to the model. However, I lack of experience to create efficient CNN architectures. I thus feel I could not get the full juice out of the method and thus want to make it public, to see if others manage to obtain better results.\n\nIn the kernel [creating bond centered 2D images](https://www.kaggle.com/fabiendaniel/creating-bond-centered-2d-images/), I show how I create the 2D images for every molecular bond. The code is not efficient and it's thus quite time consuming to create a whole dataset with such images. Such datasets are made available for the [1JHN symetry](https://www.kaggle.com/fabiendaniel/full-images-1jhn/) and part of the [1JHC molecular bonds](https://www.kaggle.com/fabiendaniel/image-1jhc-part1). A tentative CNN architecture is also given [here](https://www.kaggle.com/fabiendaniel/cnn-on-1jhn-2d-images).\n\nPS: with the 2D images constructed that way, the coordinates of the pixels are physically meaningful. Such information is lost in the CNN architecture I built. Which kind of architecture would be suitable for that ?? I tried to remove the maxpooling layer or to add filters with fixed coordinates, but this did not seem to improve the results ...", "votes": null }, { "id": "601702", "postDate": "08/18/2019 04:26:36", "content": "Hi FabienDaniel. Good idea. Could you please check the link for CNN architecture. I am trying to open the architecture and its not opening up.
", "rawMarkdown": "Hi FabienDaniel. Good idea. Could you please check the link for CNN architecture. I am trying to open the architecture and its not opening up.", "votes": null }, { "id": "601823", "postDate": "08/18/2019 07:38:04", "content": "Thanks, I forgot to make the kernel public. That's fixed !
", "rawMarkdown": "Thanks, I forgot to make the kernel public. That's fixed !", "votes": null }, { "id": "603308", "postDate": "08/20/2019 06:03:00", "content": "I made public a few more datasets on other symetries:\n- 2JHN\n- 3JHN part1 and 3JHN part2\n- 2JHH part1 and 2JHH part2
", "rawMarkdown": "I made public a few more datasets on other symetries:\n- [2JHN](https://www.kaggle.com/fabiendaniel/full-images-2jhn)\n- [3JHN part1](https://www.kaggle.com/fabiendaniel/image-3jhn-chunk1) and [3JHN part2](https://www.kaggle.com/fabiendaniel/image-3jhn-chunk2)\n- [2JHH part1](https://www.kaggle.com/fabiendaniel/image-2jhh-part-1) and [2JHH part2](https://www.kaggle.com/fabiendaniel/image-2jhh-part-2)", "votes": null }, { "id": "603904", "postDate": "08/20/2019 19:42:23", "content": "I wanted to do this as well. If I had a good enough gpu I might've run it too. Good job.
", "rawMarkdown": "I wanted to do this as well. If I had a good enough gpu I might've run it too. Good job.", "votes": null }, { "id": "603906", "postDate": "08/20/2019 19:46:34", "content": "Thanks ! I have done everything on Kaggle's kernel ;)
", "rawMarkdown": "Thanks ! I have done everything on Kaggle's kernel ;)", "votes": null }, { "id": "606155", "postDate": "08/23/2019 09:13:33", "content": "Génial! Merci pour ton partage.
", "rawMarkdown": "Génial! Merci pour ton partage.", "votes": null }, { "id": "606160", "postDate": "08/23/2019 09:26:52", "content": "avec plaisir :)
", "rawMarkdown": "avec plaisir :)", "votes": null }, { "id": "606230", "postDate": "08/23/2019 11:03:05", "content": "Thank you for sharing. 3JHN part2 is not working. Could you reload the data?
", "rawMarkdown": "Thank you for sharing. 3JHN part2 is not working. Could you reload the data?", "votes": null }, { "id": "606241", "postDate": "08/23/2019 11:09:54", "content": "The 3JHN dataset should be OK: in this kernel, I used it, so that you may check the way I load the data.
", "rawMarkdown": "The 3JHN dataset should be OK: in [this kernel](https://www.kaggle.com/fabiendaniel/3jhn-model-1-fold0-mse/), I used it, so that you may check the way I load the data.", "votes": null }, { "id": "606248", "postDate": "08/23/2019 11:16:23", "content": "I meant the link above is not working. When I clicked it, it says 404.
", "rawMarkdown": "I meant the link above is not working. When I clicked it, it says 404.", "votes": null }, { "id": "606249", "postDate": "08/23/2019 11:17:54", "content": "Chunk1 is ok but chunk2 for 3JHN is not
", "rawMarkdown": "Chunk1 is ok but chunk2 for 3JHN is not", "votes": null }, { "id": "606250", "postDate": "08/23/2019 11:19:26", "content": "ups, the dataset was still private ... fixed !
", "rawMarkdown": "ups, the dataset was still private ... fixed !", "votes": null }, { "id": "606292", "postDate": "08/23/2019 11:50:09", "content": "Thank you. I really like the way you represent edge with (r,R), which satisfies translational and rotational invariant. The only downside is that it takes forever to run :D
", "rawMarkdown": "Thank you. I really like the way you represent edge with (r,R), which satisfies translational and rotational invariant. The only downside is that it takes forever to run :D", "votes": null } ], "comments": [ { "id": 601702, "author_name": "manojprabhaakr", "author_url": "", "post_date": "08/18/2019 04:26:36", "content": "Hi FabienDaniel. Good idea. Could you please check the link for CNN architecture. I am trying to open the architecture and its not opening up.
", "votes": null, "replies": [ { "id": 601823, "author_name": "fabiendaniel", "author_url": "", "post_date": "08/18/2019 07:38:04", "content": "Thanks, I forgot to make the kernel public. That's fixed !
", "votes": null, "replies": [] } ] }, { "id": 603308, "author_name": "fabiendaniel", "author_url": "", "post_date": "08/20/2019 06:03:00", "content": "I made public a few more datasets on other symetries:\n- 2JHN\n- 3JHN part1 and 3JHN part2\n- 2JHH part1 and 2JHH part2
", "votes": null, "replies": [] }, { "id": 603904, "author_name": "roushrsh", "author_url": "", "post_date": "08/20/2019 19:42:23", "content": "I wanted to do this as well. If I had a good enough gpu I might've run it too. Good job.
", "votes": null, "replies": [ { "id": 603906, "author_name": "fabiendaniel", "author_url": "", "post_date": "08/20/2019 19:46:34", "content": "Thanks ! I have done everything on Kaggle's kernel ;)
", "votes": null, "replies": [] } ] }, { "id": 606155, "author_name": "alexpartisan", "author_url": "", "post_date": "08/23/2019 09:13:33", "content": "Génial! Merci pour ton partage.
", "votes": null, "replies": [ { "id": 606160, "author_name": "fabiendaniel", "author_url": "", "post_date": "08/23/2019 09:26:52", "content": "avec plaisir :)
", "votes": null, "replies": [] } ] }, { "id": 606230, "author_name": "zhenlanwang", "author_url": "", "post_date": "08/23/2019 11:03:05", "content": "Thank you for sharing. 3JHN part2 is not working. Could you reload the data?
", "votes": null, "replies": [ { "id": 606241, "author_name": "fabiendaniel", "author_url": "", "post_date": "08/23/2019 11:09:54", "content": "The 3JHN dataset should be OK: in this kernel, I used it, so that you may check the way I load the data.
", "votes": null, "replies": [] }, { "id": 606248, "author_name": "zhenlanwang", "author_url": "", "post_date": "08/23/2019 11:16:23", "content": "I meant the link above is not working. When I clicked it, it says 404.
", "votes": null, "replies": [] }, { "id": 606249, "author_name": "zhenlanwang", "author_url": "", "post_date": "08/23/2019 11:17:54", "content": "Chunk1 is ok but chunk2 for 3JHN is not
", "votes": null, "replies": [] }, { "id": 606250, "author_name": "fabiendaniel", "author_url": "", "post_date": "08/23/2019 11:19:26", "content": "ups, the dataset was still private ... fixed !
", "votes": null, "replies": [] }, { "id": 606292, "author_name": "zhenlanwang", "author_url": "", "post_date": "08/23/2019 11:50:09", "content": "Thank you. I really like the way you represent edge with (r,R), which satisfies translational and rotational invariant. The only downside is that it takes forever to run :D
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "601575": "In this competition, I thought of treating the problem as a regression task on 2D images. My feeling was that such images would contain enough information concerning the geometry of the molecule so that a CNN would be able to perform better that a tree based model, since in the later, we have an a priory on the features that have to be fed to the model. However, I lack of experience to create efficient CNN architectures. I thus feel I could not get the full juice out of the method and thus want to make it public, to see if others manage to obtain better results.\n\nIn the kernel [creating bond centered 2D images](https://www.kaggle.com/fabiendaniel/creating-bond-centered-2d-images/), I show how I create the 2D images for every molecular bond. The code is not efficient and it's thus quite time consuming to create a whole dataset with such images. Such datasets are made available for the [1JHN symetry](https://www.kaggle.com/fabiendaniel/full-images-1jhn/) and part of the [1JHC molecular bonds](https://www.kaggle.com/fabiendaniel/image-1jhc-part1). A tentative CNN architecture is also given [here](https://www.kaggle.com/fabiendaniel/cnn-on-1jhn-2d-images).\n\nPS: with the 2D images constructed that way, the coordinates of the pixels are physically meaningful. Such information is lost in the CNN architecture I built. Which kind of architecture would be suitable for that ?? I tried to remove the maxpooling layer or to add filters with fixed coordinates, but this did not seem to improve the results ...", "601702": "Hi FabienDaniel. Good idea. Could you please check the link for CNN architecture. I am trying to open the architecture and its not opening up.", "601823": "Thanks, I forgot to make the kernel public. That's fixed !", "603308": "I made public a few more datasets on other symetries:\n- [2JHN](https://www.kaggle.com/fabiendaniel/full-images-2jhn)\n- [3JHN part1](https://www.kaggle.com/fabiendaniel/image-3jhn-chunk1) and [3JHN part2](https://www.kaggle.com/fabiendaniel/image-3jhn-chunk2)\n- [2JHH part1](https://www.kaggle.com/fabiendaniel/image-2jhh-part-1) and [2JHH part2](https://www.kaggle.com/fabiendaniel/image-2jhh-part-2)", "603904": "I wanted to do this as well. If I had a good enough gpu I might've run it too. Good job.", "603906": "Thanks ! I have done everything on Kaggle's kernel ;)", "606155": "Génial! Merci pour ton partage.", "606160": "avec plaisir :)", "606230": "Thank you for sharing. 3JHN part2 is not working. Could you reload the data?", "606241": "The 3JHN dataset should be OK: in [this kernel](https://www.kaggle.com/fabiendaniel/3jhn-model-1-fold0-mse/), I used it, so that you may check the way I load the data.", "606248": "I meant the link above is not working. When I clicked it, it says 404.", "606249": "Chunk1 is ok but chunk2 for 3JHN is not", "606250": "ups, the dataset was still private ... fixed !", "606292": "Thank you. I really like the way you represent edge with (r,R), which satisfies translational and rotational invariant. The only downside is that it takes forever to run :D" }, "source": "meta" }