{
  "id": 236166,
  "title": "Problem with predicted InChI's /c layer text strings",
  "url": "/competitions/bms-molecular-translation/discussion/236166",
  "author_name": "",
  "post_date": "2021-05-03T06:28:15.495292300Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hello everyone,<br>\nI am simply using the ResNet18 encoder with the LSTM decoder without beam search (for now). The results I'm getting look like these:</p>\n<p>True = InChI=1S/C7H7ClN2S/c8-7-9-4-6-5(10-7)2-1-3-11-6/h4H,1-3H2<br>\nPred = InChI=1S/C7H7ClN2S/c8-7-6(10)4-5(2-1-3-7)9-11-8/h5H,1-4H2</p>\n<p>True = InChI=1S/C22H18O2/c1-24-22(23)19-10-12-21-18(14-19)8-7-17-13-16(9-11-20(17)21)15-5-3-2-4-6-15/h2-6,9-14H,7-8H2,1H3<br>\nPred = InChI=1S/C22H17O2/c1-24-21-13-11-18(12-14-21)22-15-17-8-10-19(23-20(17)16-6-2-3-7-16)9-4-5-11-22/h2-5,7-8,10-13H,6,9H2,1H3</p>\n<p>True = InChI=1S/C19H26N2O3/c1-13-3-6-15(7-4-13)21(2)19(23)12-24-16-8-9-17-14(11-16)5-10-18(22)20-17/h8-9,11,13,15H,3-7,10,12H2,1-2H3,(H,20,22)<br>\nPred = InChI=1S/C18H25N3O3/c1-12-3-5-14(6-4-12)18(23)21-9-7-15(8-10-21)22-11-13(2)17-16(17)19-20-24-22/h8-9,12,15H,3-7,10-11H2,1-2H3,(H,20,22)</p>\n<p>True = InChI=1S/C16H26FN/c1-4-11-18-16(10-5-13(2)3)12-14-6-8-15(17)9-7-14/h6-9,13,16,18H,4-5,10-12H2,1-3H3<br>\nPred = InChI=1S/C16H25FN2/c1-4-12(2)18-16(13(3)19)11-14-5-7-15(8-6-14)9-10-17/h5-8,12,15,18H,4,9-11H2,1-3H3</p>\n<p>True = InChI=1S/C12H25NO/c1-9(2)10(3)13-12-6-4-11(8-14)5-7-12/h9-14H,4-8H2,1-3H3<br>\nPred = InChI=1S/C12H25NO/c1-9(2)7-10-5-6-11(3)12(8-10)13-4/h9-11,13-14H,5-8H2,1-4H3</p>\n<p>The Main layer is fine but the /c layer is having trouble with the order of numbers. Can you please tell me what should I use to improve the order? I'm new to NLP tasks and cannot understand what to do next.<br>\nAny help or suggestion would be appreciated :)<br>\nThanks</p>",
  "messages": [
    {
      "id": "1291578",
      "postDate": "05/03/2021 06:28:15",
      "content": "<p>Hello everyone,<br>\nI am simply using the ResNet18 encoder with the LSTM decoder without beam search (for now). The results I'm getting look like these:</p>\n<p>True = InChI=1S/C7H7ClN2S/c8-7-9-4-6-5(10-7)2-1-3-11-6/h4H,1-3H2<br>\nPred = InChI=1S/C7H7ClN2S/c8-7-6(10)4-5(2-1-3-7)9-11-8/h5H,1-4H2</p>\n<p>True = InChI=1S/C22H18O2/c1-24-22(23)19-10-12-21-18(14-19)8-7-17-13-16(9-11-20(17)21)15-5-3-2-4-6-15/h2-6,9-14H,7-8H2,1H3<br>\nPred = InChI=1S/C22H17O2/c1-24-21-13-11-18(12-14-21)22-15-17-8-10-19(23-20(17)16-6-2-3-7-16)9-4-5-11-22/h2-5,7-8,10-13H,6,9H2,1H3</p>\n<p>True = InChI=1S/C19H26N2O3/c1-13-3-6-15(7-4-13)21(2)19(23)12-24-16-8-9-17-14(11-16)5-10-18(22)20-17/h8-9,11,13,15H,3-7,10,12H2,1-2H3,(H,20,22)<br>\nPred = InChI=1S/C18H25N3O3/c1-12-3-5-14(6-4-12)18(23)21-9-7-15(8-10-21)22-11-13(2)17-16(17)19-20-24-22/h8-9,12,15H,3-7,10-11H2,1-2H3,(H,20,22)</p>\n<p>True = InChI=1S/C16H26FN/c1-4-11-18-16(10-5-13(2)3)12-14-6-8-15(17)9-7-14/h6-9,13,16,18H,4-5,10-12H2,1-3H3<br>\nPred = InChI=1S/C16H25FN2/c1-4-12(2)18-16(13(3)19)11-14-5-7-15(8-6-14)9-10-17/h5-8,12,15,18H,4,9-11H2,1-3H3</p>\n<p>True = InChI=1S/C12H25NO/c1-9(2)10(3)13-12-6-4-11(8-14)5-7-12/h9-14H,4-8H2,1-3H3<br>\nPred = InChI=1S/C12H25NO/c1-9(2)7-10-5-6-11(3)12(8-10)13-4/h9-11,13-14H,5-8H2,1-4H3</p>\n<p>The Main layer is fine but the /c layer is having trouble with the order of numbers. Can you please tell me what should I use to improve the order? I'm new to NLP tasks and cannot understand what to do next.<br>\nAny help or suggestion would be appreciated :)<br>\nThanks</p>",
      "rawMarkdown": "Hello everyone,\nI am simply using the ResNet18 encoder with the LSTM decoder without beam search (for now). The results I'm getting look like these:\n\nTrue = InChI=1S/C7H7ClN2S/c8-7-9-4-6-5(10-7)2-1-3-11-6/h4H,1-3H2\nPred = InChI=1S/C7H7ClN2S/c8-7-6(10)4-5(2-1-3-7)9-11-8/h5H,1-4H2\n\nTrue = InChI=1S/C22H18O2/c1-24-22(23)19-10-12-21-18(14-19)8-7-17-13-16(9-11-20(17)21)15-5-3-2-4-6-15/h2-6,9-14H,7-8H2,1H3\nPred = InChI=1S/C22H17O2/c1-24-21-13-11-18(12-14-21)22-15-17-8-10-19(23-20(17)16-6-2-3-7-16)9-4-5-11-22/h2-5,7-8,10-13H,6,9H2,1H3\n\nTrue = InChI=1S/C19H26N2O3/c1-13-3-6-15(7-4-13)21(2)19(23)12-24-16-8-9-17-14(11-16)5-10-18(22)20-17/h8-9,11,13,15H,3-7,10,12H2,1-2H3,(H,20,22)\nPred = InChI=1S/C18H25N3O3/c1-12-3-5-14(6-4-12)18(23)21-9-7-15(8-10-21)22-11-13(2)17-16(17)19-20-24-22/h8-9,12,15H,3-7,10-11H2,1-2H3,(H,20,22)\n\nTrue = InChI=1S/C16H26FN/c1-4-11-18-16(10-5-13(2)3)12-14-6-8-15(17)9-7-14/h6-9,13,16,18H,4-5,10-12H2,1-3H3\nPred = InChI=1S/C16H25FN2/c1-4-12(2)18-16(13(3)19)11-14-5-7-15(8-6-14)9-10-17/h5-8,12,15,18H,4,9-11H2,1-3H3\n\nTrue = InChI=1S/C12H25NO/c1-9(2)10(3)13-12-6-4-11(8-14)5-7-12/h9-14H,4-8H2,1-3H3\nPred = InChI=1S/C12H25NO/c1-9(2)7-10-5-6-11(3)12(8-10)13-4/h9-11,13-14H,5-8H2,1-4H3\n\nThe Main layer is fine but the /c layer is having trouble with the order of numbers. Can you please tell me what should I use to improve the order? I'm new to NLP tasks and cannot understand what to do next.\nAny help or suggestion would be appreciated :)\nThanks",
      "votes": null
    },
    {
      "id": "1291643",
      "postDate": "05/03/2021 08:08:02",
      "content": "<p>Try this one:<br>\n<a href=\"https://www.kaggle.com/nofreewill/normalize-your-predictions\" target=\"_blank\">https://www.kaggle.com/nofreewill/normalize-your-predictions</a></p>",
      "rawMarkdown": "Try this one:\nhttps://www.kaggle.com/nofreewill/normalize-your-predictions",
      "votes": null
    },
    {
      "id": "1291645",
      "postDate": "05/03/2021 08:11:44",
      "content": "<p>You can train for more epochs and see if the predictions are improving or not !. If you see a little/no improvement I would suggest to </p>\n<ol>\n<li>Change encoder architecture </li>\n<li>Implement multi-layer LSTM decoder with attention</li>\n<li>Beam Search</li>\n<li>Top K Sampling</li>\n</ol>\n<p>I am also a newbie exploring NLP Tasks.</p>",
      "rawMarkdown": "You can train for more epochs and see if the predictions are improving or not !. If you see a little/no improvement I would suggest to \n1. Change encoder architecture \n2. Implement multi-layer LSTM decoder with attention\n3. Beam Search\n4. Top K Sampling\n\nI am also a newbie exploring NLP Tasks.",
      "votes": null
    },
    {
      "id": "1292033",
      "postDate": "05/03/2021 15:18:11",
      "content": "<p>Thanks! Will try it.</p>",
      "rawMarkdown": "Thanks! Will try it.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1291643,
      "author_name": "nofreewill",
      "author_url": "",
      "post_date": "05/03/2021 08:08:02",
      "content": "<p>Try this one:<br>\n<a href=\"https://www.kaggle.com/nofreewill/normalize-your-predictions\" target=\"_blank\">https://www.kaggle.com/nofreewill/normalize-your-predictions</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 1292033,
          "author_name": "virksaab",
          "author_url": "",
          "post_date": "05/03/2021 15:18:11",
          "content": "<p>Thanks! Will try it.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1291645,
      "author_name": "tawheedrony",
      "author_url": "",
      "post_date": "05/03/2021 08:11:44",
      "content": "<p>You can train for more epochs and see if the predictions are improving or not !. If you see a little/no improvement I would suggest to </p>\n<ol>\n<li>Change encoder architecture </li>\n<li>Implement multi-layer LSTM decoder with attention</li>\n<li>Beam Search</li>\n<li>Top K Sampling</li>\n</ol>\n<p>I am also a newbie exploring NLP Tasks.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1291578": "Hello everyone,\nI am simply using the ResNet18 encoder with the LSTM decoder without beam search (for now). The results I'm getting look like these:\n\nTrue = InChI=1S/C7H7ClN2S/c8-7-9-4-6-5(10-7)2-1-3-11-6/h4H,1-3H2\nPred = InChI=1S/C7H7ClN2S/c8-7-6(10)4-5(2-1-3-7)9-11-8/h5H,1-4H2\n\nTrue = InChI=1S/C22H18O2/c1-24-22(23)19-10-12-21-18(14-19)8-7-17-13-16(9-11-20(17)21)15-5-3-2-4-6-15/h2-6,9-14H,7-8H2,1H3\nPred = InChI=1S/C22H17O2/c1-24-21-13-11-18(12-14-21)22-15-17-8-10-19(23-20(17)16-6-2-3-7-16)9-4-5-11-22/h2-5,7-8,10-13H,6,9H2,1H3\n\nTrue = InChI=1S/C19H26N2O3/c1-13-3-6-15(7-4-13)21(2)19(23)12-24-16-8-9-17-14(11-16)5-10-18(22)20-17/h8-9,11,13,15H,3-7,10,12H2,1-2H3,(H,20,22)\nPred = InChI=1S/C18H25N3O3/c1-12-3-5-14(6-4-12)18(23)21-9-7-15(8-10-21)22-11-13(2)17-16(17)19-20-24-22/h8-9,12,15H,3-7,10-11H2,1-2H3,(H,20,22)\n\nTrue = InChI=1S/C16H26FN/c1-4-11-18-16(10-5-13(2)3)12-14-6-8-15(17)9-7-14/h6-9,13,16,18H,4-5,10-12H2,1-3H3\nPred = InChI=1S/C16H25FN2/c1-4-12(2)18-16(13(3)19)11-14-5-7-15(8-6-14)9-10-17/h5-8,12,15,18H,4,9-11H2,1-3H3\n\nTrue = InChI=1S/C12H25NO/c1-9(2)10(3)13-12-6-4-11(8-14)5-7-12/h9-14H,4-8H2,1-3H3\nPred = InChI=1S/C12H25NO/c1-9(2)7-10-5-6-11(3)12(8-10)13-4/h9-11,13-14H,5-8H2,1-4H3\n\nThe Main layer is fine but the /c layer is having trouble with the order of numbers. Can you please tell me what should I use to improve the order? I'm new to NLP tasks and cannot understand what to do next.\nAny help or suggestion would be appreciated :)\nThanks",
    "1291643": "Try this one:\nhttps://www.kaggle.com/nofreewill/normalize-your-predictions",
    "1291645": "You can train for more epochs and see if the predictions are improving or not !. If you see a little/no improvement I would suggest to \n1. Change encoder architecture \n2. Implement multi-layer LSTM decoder with attention\n3. Beam Search\n4. Top K Sampling\n\nI am also a newbie exploring NLP Tasks.",
    "1292033": "Thanks! Will try it."
  },
  "source": "meta"
}