{
  "id": 105917,
  "title": "What's the key point to be better?",
  "url": "/competitions/champs-scalar-coupling/discussion/105917",
  "author_name": "",
  "post_date": "2019-08-27T07:11:08.908975700Z",
  "votes": 2,
  "comment_count": 8,
  "views": 0,
  "content": "<p>We have tried many approaches to get a better score, but it seems very hard to be higher.\nThe key point is the feature, model or stacking? I'm not sure yet, but I hope I can learn more good solutions from the high score teams after the end of this competition.\n💪 </p>",
  "messages": [
    {
      "id": "608795",
      "postDate": "08/27/2019 07:11:08",
      "content": "<p>We have tried many approaches to get a better score, but it seems very hard to be higher.\nThe key point is the feature, model or stacking? I'm not sure yet, but I hope I can learn more good solutions from the high score teams after the end of this competition.\n💪 </p>",
      "rawMarkdown": "We have tried many approaches to get a better score, but it seems very hard to be higher.\nThe key point is the feature, model or stacking? I'm not sure yet, but I hope I can learn more good solutions from the high score teams after the end of this competition.\n💪",
      "votes": null
    },
    {
      "id": "608799",
      "postDate": "08/27/2019 07:15:06",
      "content": "<p>I think that the key is the architecture of Neural Net and hardware. But can't share anything until the end of the competition. :) </p>",
      "rawMarkdown": "I think that the key is the architecture of Neural Net and hardware. But can't share anything until the end of the competition. :)",
      "votes": null
    },
    {
      "id": "608808",
      "postDate": "08/27/2019 07:19:47",
      "content": "<p>What hardware do you have?  On our team we have 4 1080 Ti and 2 K40.</p>",
      "rawMarkdown": "What hardware do you have?  On our team we have 4 1080 Ti and 2 K40.",
      "votes": null
    },
    {
      "id": "608823",
      "postDate": "08/27/2019 07:32:25",
      "content": "<p>Thanks for your advice. I do agree. For deep learning, it's true that the proper architecture is able to fit the data better, however sometimes the 'proper' means more complex NN, and it would take more computing power to train. At the same time, to design the proper architecture, it need more analysis to the data and the knowledge of the NN. \nI hope I can learn much from your teams' solution after the competition.</p>",
      "rawMarkdown": "Thanks for your advice. I do agree. For deep learning, it's true that the proper architecture is able to fit the data better, however sometimes the 'proper' means more complex NN, and it would take more computing power to train. At the same time, to design the proper architecture, it need more analysis to the data and the knowledge of the NN. \nI hope I can learn much from your teams' solution after the competition.",
      "votes": null
    },
    {
      "id": "608874",
      "postDate": "08/27/2019 08:33:04",
      "content": "<p>I use my 1 1080ti, 1 P100 on google cloud instance and actively used 4 kernels with P100 (until the limit became 1). My teammates have more hardware, especially <a href=\"/borisdee\">@borisdee</a> </p>",
      "rawMarkdown": "I use my 1 1080ti, 1 P100 on google cloud instance and actively used 4 kernels with P100 (until the limit became 1). My teammates have more hardware, especially @borisdee",
      "votes": null
    },
    {
      "id": "609165",
      "postDate": "08/27/2019 13:04:28",
      "content": "<p>Looking forward to see your solution tomorrow. I was tried with different architecture but was get any good results. It seems like your team tried a lot of architectures. Is your performance improve constantly in training? Because my training also seems to go to a plateau no matter which architecture I tried.</p>",
      "rawMarkdown": "Looking forward to see your solution tomorrow. I was tried with different architecture but was get any good results. It seems like your team tried a lot of architectures. Is your performance improve constantly in training? Because my training also seems to go to a plateau no matter which architecture I tried.",
      "votes": null
    },
    {
      "id": "609178",
      "postDate": "08/27/2019 13:18:35",
      "content": "<blockquote>\n  <p>my training also seems to go to a plateau\n  As far as I see, you didn't make any submissions...</p>\n</blockquote>",
      "rawMarkdown": "&gt; my training also seems to go to a plateau\nAs far as I see, you didn't make any submissions...",
      "votes": null
    },
    {
      "id": "609180",
      "postDate": "08/27/2019 13:24:27",
      "content": "<p>I submitted on my previous (deleted) account. Didn't noticed about the deadline for the first submission ==&gt; out of competition. :(</p>",
      "rawMarkdown": "I submitted on my previous (deleted) account. Didn't noticed about the deadline for the first submission ==&gt; out of competition. :(",
      "votes": null
    },
    {
      "id": "610240",
      "postDate": "08/28/2019 15:12:59",
      "content": "<p>There is no substitute for failing and failing often - then spend a week AFTER the competition ends to look at all ideas that worked - this week is much more important to improve than those that preceded it.  The failing part certainly helps the memory!  I would be very surprised it one of the improvements was to detect similar hydrogens on the same carbon - whether directly or indirectly through graphs.</p>",
      "rawMarkdown": "There is no substitute for failing and failing often - then spend a week AFTER the competition ends to look at all ideas that worked - this week is much more important to improve than those that preceded it.  The failing part certainly helps the memory!  I would be very surprised it one of the improvements was to detect similar hydrogens on the same carbon - whether directly or indirectly through graphs.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 608799,
      "author_name": "artgor",
      "author_url": "",
      "post_date": "08/27/2019 07:15:06",
      "content": "<p>I think that the key is the architecture of Neural Net and hardware. But can't share anything until the end of the competition. :) </p>",
      "votes": null,
      "replies": [
        {
          "id": 608808,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "08/27/2019 07:19:47",
          "content": "<p>What hardware do you have?  On our team we have 4 1080 Ti and 2 K40.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 608823,
          "author_name": "assare",
          "author_url": "",
          "post_date": "08/27/2019 07:32:25",
          "content": "<p>Thanks for your advice. I do agree. For deep learning, it's true that the proper architecture is able to fit the data better, however sometimes the 'proper' means more complex NN, and it would take more computing power to train. At the same time, to design the proper architecture, it need more analysis to the data and the knowledge of the NN. \nI hope I can learn much from your teams' solution after the competition.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 608874,
          "author_name": "artgor",
          "author_url": "",
          "post_date": "08/27/2019 08:33:04",
          "content": "<p>I use my 1 1080ti, 1 P100 on google cloud instance and actively used 4 kernels with P100 (until the limit became 1). My teammates have more hardware, especially <a href=\"/borisdee\">@borisdee</a> </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 609165,
          "author_name": "vithng",
          "author_url": "",
          "post_date": "08/27/2019 13:04:28",
          "content": "<p>Looking forward to see your solution tomorrow. I was tried with different architecture but was get any good results. It seems like your team tried a lot of architectures. Is your performance improve constantly in training? Because my training also seems to go to a plateau no matter which architecture I tried.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 609178,
          "author_name": "artgor",
          "author_url": "",
          "post_date": "08/27/2019 13:18:35",
          "content": "<blockquote>\n  <p>my training also seems to go to a plateau\n  As far as I see, you didn't make any submissions...</p>\n</blockquote>",
          "votes": null,
          "replies": []
        },
        {
          "id": 609180,
          "author_name": "vithng",
          "author_url": "",
          "post_date": "08/27/2019 13:24:27",
          "content": "<p>I submitted on my previous (deleted) account. Didn't noticed about the deadline for the first submission ==&gt; out of competition. :(</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 610240,
      "author_name": "scirpus",
      "author_url": "",
      "post_date": "08/28/2019 15:12:59",
      "content": "<p>There is no substitute for failing and failing often - then spend a week AFTER the competition ends to look at all ideas that worked - this week is much more important to improve than those that preceded it.  The failing part certainly helps the memory!  I would be very surprised it one of the improvements was to detect similar hydrogens on the same carbon - whether directly or indirectly through graphs.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "608795": "We have tried many approaches to get a better score, but it seems very hard to be higher.\nThe key point is the feature, model or stacking? I'm not sure yet, but I hope I can learn more good solutions from the high score teams after the end of this competition.\n💪",
    "608799": "I think that the key is the architecture of Neural Net and hardware. But can't share anything until the end of the competition. :)",
    "608808": "What hardware do you have?  On our team we have 4 1080 Ti and 2 K40.",
    "608823": "Thanks for your advice. I do agree. For deep learning, it's true that the proper architecture is able to fit the data better, however sometimes the 'proper' means more complex NN, and it would take more computing power to train. At the same time, to design the proper architecture, it need more analysis to the data and the knowledge of the NN. \nI hope I can learn much from your teams' solution after the competition.",
    "608874": "I use my 1 1080ti, 1 P100 on google cloud instance and actively used 4 kernels with P100 (until the limit became 1). My teammates have more hardware, especially @borisdee",
    "609165": "Looking forward to see your solution tomorrow. I was tried with different architecture but was get any good results. It seems like your team tried a lot of architectures. Is your performance improve constantly in training? Because my training also seems to go to a plateau no matter which architecture I tried.",
    "609178": "&gt; my training also seems to go to a plateau\nAs far as I see, you didn't make any submissions...",
    "609180": "I submitted on my previous (deleted) account. Didn't noticed about the deadline for the first submission ==&gt; out of competition. :(",
    "610240": "There is no substitute for failing and failing often - then spend a week AFTER the competition ends to look at all ideas that worked - this week is much more important to improve than those that preceded it.  The failing part certainly helps the memory!  I would be very surprised it one of the improvements was to detect similar hydrogens on the same carbon - whether directly or indirectly through graphs."
  },
  "source": "meta"
}