{
  "id": 106905,
  "title": "19th place solution",
  "url": "/competitions/champs-scalar-coupling/writeups/toshi-k-19th-place-solution",
  "author_name": "",
  "post_date": "2019-08-31T17:38:04.494477300Z",
  "votes": 30,
  "comment_count": 10,
  "views": 0,
  "content": "<p>Hi Everyone,</p>\n\n<p>Ending final validation, I share my solution.\nMy solution is what I call <em>Triplet Update</em>.\nMy model has hidden layers for combinations of triple atoms.\nThis model is generalization of <em>Edge Update</em>, and similar to Higher-order Graph Neural Networks in a sense.</p>\n\n<p>Code: <a href=\"https://github.com/toshi-k/kaggle-champs-scalar-coupling\">https://github.com/toshi-k/kaggle-champs-scalar-coupling</a></p>\n\n<p><img src=\"https://raw.githubusercontent.com/toshi-k/kaggle-champs-scalar-coupling/master/img/concept.png\" alt=\"conceptual diagram\"></p>",
  "messages": [
    {
      "id": "614571",
      "postDate": "08/31/2019 17:38:04",
      "content": "<p>Hi Everyone,</p>\n\n<p>Ending final validation, I share my solution.\nMy solution is what I call <em>Triplet Update</em>.\nMy model has hidden layers for combinations of triple atoms.\nThis model is generalization of <em>Edge Update</em>, and similar to Higher-order Graph Neural Networks in a sense.</p>\n\n<p>Code: <a href=\"https://github.com/toshi-k/kaggle-champs-scalar-coupling\">https://github.com/toshi-k/kaggle-champs-scalar-coupling</a></p>\n\n<p><img src=\"https://raw.githubusercontent.com/toshi-k/kaggle-champs-scalar-coupling/master/img/concept.png\" alt=\"conceptual diagram\"></p>",
      "rawMarkdown": "Hi Everyone,\n\nEnding final validation, I share my solution.\nMy solution is what I call *Triplet Update*.\nMy model has hidden layers for combinations of triple atoms.\nThis model is generalization of *Edge Update*, and similar to Higher-order Graph Neural Networks in a sense.\n\nCode: https://github.com/toshi-k/kaggle-champs-scalar-coupling\n\n![conceptual diagram][1]\n\n\n  [1]: https://raw.githubusercontent.com/toshi-k/kaggle-champs-scalar-coupling/master/img/concept.png",
      "votes": null
    },
    {
      "id": "614588",
      "postDate": "08/31/2019 18:04:36",
      "content": "<p>Congrats...!!\nGreat Work...!!\nThanks for Sharing...!! <a href=\"/toshik\">@toshik</a> </p>",
      "rawMarkdown": "Congrats...!!\nGreat Work...!!\nThanks for Sharing...!! @toshik",
      "votes": null
    },
    {
      "id": "614590",
      "postDate": "08/31/2019 18:08:14",
      "content": "<p><a href=\"/toshik\">@toshik</a> Thank you for your kernel! It was my first competition and I am sure, that without your approach I wouldn't get such a high score (almost silver)! </p>",
      "rawMarkdown": "toshik Thank you for your kernel! It was my first competition and I am sure, that without your approach I wouldn't get such a high score (almost silver)!",
      "votes": null
    },
    {
      "id": "614683",
      "postDate": "08/31/2019 22:13:38",
      "content": "<p>Thanks for sharing with the github repo. There was a weekend during this competition that I spent getting to know your \"SchNet Starter Kit\" notebook. I really like your coding style- it was my first experience with chainer. We didn't end up using it in our final solution but I learned a lot because of it!</p>\n\n<p>Do you typically use chainer - or was it mainly because of the chainer-chemistry package?</p>\n\n<p>Excited to see you reach GM very soon.</p>",
      "rawMarkdown": "Thanks for sharing with the github repo. There was a weekend during this competition that I spent getting to know your \"SchNet Starter Kit\" notebook. I really like your coding style- it was my first experience with chainer. We didn't end up using it in our final solution but I learned a lot because of it!\n\nDo you typically use chainer - or was it mainly because of the chainer-chemistry package?\n\nExcited to see you reach GM very soon.",
      "votes": null
    },
    {
      "id": "614777",
      "postDate": "09/01/2019 03:49:55",
      "content": "<p>Congrats on your result! Extremely close to solo Gold! \nAlso, thank you for sharing your SchNet Starter Kit. This is what got me Silver.</p>",
      "rawMarkdown": "Congrats on your result! Extremely close to solo Gold! \nAlso, thank you for sharing your SchNet Starter Kit. This is what got me Silver.",
      "votes": null
    },
    {
      "id": "614809",
      "postDate": "09/01/2019 05:30:01",
      "content": "<blockquote>\n  <p>Do you typically use chainer - or was it mainly because of the chainer-chemistry package?</p>\n</blockquote>\n\n<p>I usually use other DL frameworks too (TensorFlow, PyTorch, and so on). The reason I used Chainer in this competition was that chainer-chemistry was a good entry point.</p>",
      "rawMarkdown": "&gt; Do you typically use chainer - or was it mainly because of the chainer-chemistry package?\n\nI usually use other DL frameworks too (TensorFlow, PyTorch, and so on). The reason I used Chainer in this competition was that chainer-chemistry was a good entry point.",
      "votes": null
    },
    {
      "id": "614844",
      "postDate": "09/01/2019 06:23:59",
      "content": "<p>Congratulations <a href=\"/toshik\">@toshik</a> ! Great work and solution. <br>\nThanks a lot for sharing the SchNet starter kit. Its a really nice library and I wouldn't have explored it without your notebook.</p>",
      "rawMarkdown": "Congratulations @toshik ! Great work and solution.   \nThanks a lot for sharing the SchNet starter kit. Its a really nice library and I wouldn't have explored it without your notebook.",
      "votes": null
    },
    {
      "id": "614871",
      "postDate": "09/01/2019 07:11:40",
      "content": "<p>Congrats <a href=\"/toshik\">@toshik</a> and thanks for sharing.</p>",
      "rawMarkdown": "Congrats @toshik and thanks for sharing.",
      "votes": null
    },
    {
      "id": "615031",
      "postDate": "09/01/2019 11:29:56",
      "content": "<p>Thank you for clearing that up. Did you find chainer faster comparing to pytorch?</p>",
      "rawMarkdown": "Thank you for clearing that up. Did you find chainer faster comparing to pytorch?",
      "votes": null
    },
    {
      "id": "615038",
      "postDate": "09/01/2019 11:41:41",
      "content": "<p>I suppose running time would almost the same when we implement well.</p>",
      "rawMarkdown": "I suppose running time would almost the same when we implement well.",
      "votes": null
    },
    {
      "id": "615483",
      "postDate": "09/02/2019 02:27:20",
      "content": "<p>Congratulations <a href=\"/toshik\">@toshik</a> - especially on getting this far alone!</p>",
      "rawMarkdown": "Congratulations @toshik - especially on getting this far alone!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 614588,
      "author_name": "veeralakrishna",
      "author_url": "",
      "post_date": "08/31/2019 18:04:36",
      "content": "<p>Congrats...!!\nGreat Work...!!\nThanks for Sharing...!! <a href=\"/toshik\">@toshik</a> </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 614590,
      "author_name": "betterthanbutter",
      "author_url": "",
      "post_date": "08/31/2019 18:08:14",
      "content": "<p><a href=\"/toshik\">@toshik</a> Thank you for your kernel! It was my first competition and I am sure, that without your approach I wouldn't get such a high score (almost silver)! </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 614683,
      "author_name": "robikscube",
      "author_url": "",
      "post_date": "08/31/2019 22:13:38",
      "content": "<p>Thanks for sharing with the github repo. There was a weekend during this competition that I spent getting to know your \"SchNet Starter Kit\" notebook. I really like your coding style- it was my first experience with chainer. We didn't end up using it in our final solution but I learned a lot because of it!</p>\n\n<p>Do you typically use chainer - or was it mainly because of the chainer-chemistry package?</p>\n\n<p>Excited to see you reach GM very soon.</p>",
      "votes": null,
      "replies": [
        {
          "id": 614809,
          "author_name": "toshik",
          "author_url": "",
          "post_date": "09/01/2019 05:30:01",
          "content": "<blockquote>\n  <p>Do you typically use chainer - or was it mainly because of the chainer-chemistry package?</p>\n</blockquote>\n\n<p>I usually use other DL frameworks too (TensorFlow, PyTorch, and so on). The reason I used Chainer in this competition was that chainer-chemistry was a good entry point.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 615031,
          "author_name": "jiweiliu",
          "author_url": "",
          "post_date": "09/01/2019 11:29:56",
          "content": "<p>Thank you for clearing that up. Did you find chainer faster comparing to pytorch?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 615038,
          "author_name": "toshik",
          "author_url": "",
          "post_date": "09/01/2019 11:41:41",
          "content": "<p>I suppose running time would almost the same when we implement well.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 614777,
      "author_name": "stocks",
      "author_url": "",
      "post_date": "09/01/2019 03:49:55",
      "content": "<p>Congrats on your result! Extremely close to solo Gold! \nAlso, thank you for sharing your SchNet Starter Kit. This is what got me Silver.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 614844,
      "author_name": "rohanrao",
      "author_url": "",
      "post_date": "09/01/2019 06:23:59",
      "content": "<p>Congratulations <a href=\"/toshik\">@toshik</a> ! Great work and solution. <br>\nThanks a lot for sharing the SchNet starter kit. Its a really nice library and I wouldn't have explored it without your notebook.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 614871,
      "author_name": "joelhanson",
      "author_url": "",
      "post_date": "09/01/2019 07:11:40",
      "content": "<p>Congrats <a href=\"/toshik\">@toshik</a> and thanks for sharing.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 615483,
      "author_name": "michelml",
      "author_url": "",
      "post_date": "09/02/2019 02:27:20",
      "content": "<p>Congratulations <a href=\"/toshik\">@toshik</a> - especially on getting this far alone!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "614571": "Hi Everyone,\n\nEnding final validation, I share my solution.\nMy solution is what I call *Triplet Update*.\nMy model has hidden layers for combinations of triple atoms.\nThis model is generalization of *Edge Update*, and similar to Higher-order Graph Neural Networks in a sense.\n\nCode: https://github.com/toshi-k/kaggle-champs-scalar-coupling\n\n![conceptual diagram][1]\n\n\n  [1]: https://raw.githubusercontent.com/toshi-k/kaggle-champs-scalar-coupling/master/img/concept.png",
    "614588": "Congrats...!!\nGreat Work...!!\nThanks for Sharing...!! @toshik",
    "614590": "toshik Thank you for your kernel! It was my first competition and I am sure, that without your approach I wouldn't get such a high score (almost silver)!",
    "614683": "Thanks for sharing with the github repo. There was a weekend during this competition that I spent getting to know your \"SchNet Starter Kit\" notebook. I really like your coding style- it was my first experience with chainer. We didn't end up using it in our final solution but I learned a lot because of it!\n\nDo you typically use chainer - or was it mainly because of the chainer-chemistry package?\n\nExcited to see you reach GM very soon.",
    "614777": "Congrats on your result! Extremely close to solo Gold! \nAlso, thank you for sharing your SchNet Starter Kit. This is what got me Silver.",
    "614809": "&gt; Do you typically use chainer - or was it mainly because of the chainer-chemistry package?\n\nI usually use other DL frameworks too (TensorFlow, PyTorch, and so on). The reason I used Chainer in this competition was that chainer-chemistry was a good entry point.",
    "614844": "Congratulations @toshik ! Great work and solution.   \nThanks a lot for sharing the SchNet starter kit. Its a really nice library and I wouldn't have explored it without your notebook.",
    "614871": "Congrats @toshik and thanks for sharing.",
    "615031": "Thank you for clearing that up. Did you find chainer faster comparing to pytorch?",
    "615038": "I suppose running time would almost the same when we implement well.",
    "615483": "Congratulations @toshik - especially on getting this far alone!"
  },
  "source": "meta"
}