{
  "id": 107751,
  "title": "Forbidden Features Appear to Help",
  "url": "/competitions/champs-scalar-coupling/discussion/107751",
  "author_name": "",
  "post_date": "2019-09-06T12:02:33.169961100Z",
  "votes": 10,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi Gang,   </p>\n\n<p>A big congratulations to the top teams and organizers on some great results.   While I did not have the chance to do this competition, wanted to offer the following, being curious to see if any \"fast\" approximations to J couplings and shieldings would be useful as features in neural net or boosted tree models.   While such features would be considered illegal for this competition, they could potentially be helpful in practice.   I ran the public binary version of <a href=\"http://www.demon-software.com/public_html/index.html\">deMon</a> with default settings across the QM9 molecules and parsed out the results.   This took 2.5 days running in parallel on 28 nodes and results are available at <a href=\"https://www.kaggle.com/sasrdw/demon-qm9-features\">deMon QM9 Features</a>.</p>\n\n<p>My preliminary findings using these features are a fairly large improvement in boosted tree models (e.g. jumping from -1.5 to -2.5 in the competition metric) and smaller improvements with neural nets (-2.6 to -2.8).   I'm trying Robin's larger model now but having some trouble with a fastai/pytorch memory leak, so only able to run a few epochs at a time.</p>\n\n<p>If you are able to add these features to your model, it would be interesting to hear if they help.  I'm also wondering if this kind of approach has appeal to computational chemists on top of our newfound insights.  Would guess better features along these lines could be developed from various software packages with more appropriate choices for the functional and basis set that optimally balance the tradeoff between accuracy and speed.</p>",
  "messages": [
    {
      "id": "619645",
      "postDate": "09/06/2019 12:02:33",
      "content": "<p>Hi Gang,   </p>\n\n<p>A big congratulations to the top teams and organizers on some great results.   While I did not have the chance to do this competition, wanted to offer the following, being curious to see if any \"fast\" approximations to J couplings and shieldings would be useful as features in neural net or boosted tree models.   While such features would be considered illegal for this competition, they could potentially be helpful in practice.   I ran the public binary version of <a href=\"http://www.demon-software.com/public_html/index.html\">deMon</a> with default settings across the QM9 molecules and parsed out the results.   This took 2.5 days running in parallel on 28 nodes and results are available at <a href=\"https://www.kaggle.com/sasrdw/demon-qm9-features\">deMon QM9 Features</a>.</p>\n\n<p>My preliminary findings using these features are a fairly large improvement in boosted tree models (e.g. jumping from -1.5 to -2.5 in the competition metric) and smaller improvements with neural nets (-2.6 to -2.8).   I'm trying Robin's larger model now but having some trouble with a fastai/pytorch memory leak, so only able to run a few epochs at a time.</p>\n\n<p>If you are able to add these features to your model, it would be interesting to hear if they help.  I'm also wondering if this kind of approach has appeal to computational chemists on top of our newfound insights.  Would guess better features along these lines could be developed from various software packages with more appropriate choices for the functional and basis set that optimally balance the tradeoff between accuracy and speed.</p>",
      "rawMarkdown": "Hi Gang,   \n\nA big congratulations to the top teams and organizers on some great results.   While I did not have the chance to do this competition, wanted to offer the following, being curious to see if any \"fast\" approximations to J couplings and shieldings would be useful as features in neural net or boosted tree models.   While such features would be considered illegal for this competition, they could potentially be helpful in practice.   I ran the public binary version of [deMon](http://www.demon-software.com/public_html/index.html) with default settings across the QM9 molecules and parsed out the results.   This took 2.5 days running in parallel on 28 nodes and results are available at [deMon QM9 Features](https://www.kaggle.com/sasrdw/demon-qm9-features).\n\nMy preliminary findings using these features are a fairly large improvement in boosted tree models (e.g. jumping from -1.5 to -2.5 in the competition metric) and smaller improvements with neural nets (-2.6 to -2.8).   I'm trying Robin's larger model now but having some trouble with a fastai/pytorch memory leak, so only able to run a few epochs at a time.\n\nIf you are able to add these features to your model, it would be interesting to hear if they help.  I'm also wondering if this kind of approach has appeal to computational chemists on top of our newfound insights.  Would guess better features along these lines could be developed from various software packages with more appropriate choices for the functional and basis set that optimally balance the tradeoff between accuracy and speed.",
      "votes": null
    },
    {
      "id": "619959",
      "postDate": "09/06/2019 20:08:40",
      "content": "<p>nice job!</p>",
      "rawMarkdown": "nice job!",
      "votes": null
    },
    {
      "id": "621268",
      "postDate": "09/08/2019 11:44:27",
      "content": "<p>Well done!</p>",
      "rawMarkdown": "Well done!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 619959,
      "author_name": "chabir",
      "author_url": "",
      "post_date": "09/06/2019 20:08:40",
      "content": "<p>nice job!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 621268,
      "author_name": "babubhai",
      "author_url": "",
      "post_date": "09/08/2019 11:44:27",
      "content": "<p>Well done!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "619645": "Hi Gang,   \n\nA big congratulations to the top teams and organizers on some great results.   While I did not have the chance to do this competition, wanted to offer the following, being curious to see if any \"fast\" approximations to J couplings and shieldings would be useful as features in neural net or boosted tree models.   While such features would be considered illegal for this competition, they could potentially be helpful in practice.   I ran the public binary version of [deMon](http://www.demon-software.com/public_html/index.html) with default settings across the QM9 molecules and parsed out the results.   This took 2.5 days running in parallel on 28 nodes and results are available at [deMon QM9 Features](https://www.kaggle.com/sasrdw/demon-qm9-features).\n\nMy preliminary findings using these features are a fairly large improvement in boosted tree models (e.g. jumping from -1.5 to -2.5 in the competition metric) and smaller improvements with neural nets (-2.6 to -2.8).   I'm trying Robin's larger model now but having some trouble with a fastai/pytorch memory leak, so only able to run a few epochs at a time.\n\nIf you are able to add these features to your model, it would be interesting to hear if they help.  I'm also wondering if this kind of approach has appeal to computational chemists on top of our newfound insights.  Would guess better features along these lines could be developed from various software packages with more appropriate choices for the functional and basis set that optimally balance the tradeoff between accuracy and speed.",
    "619959": "nice job!",
    "621268": "Well done!"
  },
  "source": "meta"
}