{
  "id": 234462,
  "title": "How to train model",
  "url": "/competitions/bms-molecular-translation/discussion/234462",
  "author_name": "",
  "post_date": "2021-04-24T12:38:53.459783100Z",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Can we take it as a image classification problem? But, all InChI labels in train file are unique, i.e. we have just 1 example per class and totally about 2.42m classes. Will it be okay to train a classification model over the train files?<br>\nAlso, as InChI are unique i.e. one for one structure, so are test images are subset of train images?<br>\nPlease answer..</p>",
  "messages": [
    {
      "id": "1282958",
      "postDate": "04/24/2021 12:38:53",
      "content": "<p>Can we take it as a image classification problem? But, all InChI labels in train file are unique, i.e. we have just 1 example per class and totally about 2.42m classes. Will it be okay to train a classification model over the train files?<br>\nAlso, as InChI are unique i.e. one for one structure, so are test images are subset of train images?<br>\nPlease answer..</p>",
      "rawMarkdown": "Can we take it as a image classification problem? But, all InChI labels in train file are unique, i.e. we have just 1 example per class and totally about 2.42m classes. Will it be okay to train a classification model over the train files?\nAlso, as InChI are unique i.e. one for one structure, so are test images are subset of train images?\nPlease answer..",
      "votes": null
    },
    {
      "id": "1282994",
      "postDate": "04/24/2021 13:33:57",
      "content": "<blockquote>\n  <p>are test images are subset of train images?</p>\n</blockquote>\n<p>No, test molecules are also unique.</p>",
      "rawMarkdown": "> are test images are subset of train images?\n\nNo, test molecules are also unique.",
      "votes": null
    },
    {
      "id": "1283003",
      "postDate": "04/24/2021 13:42:12",
      "content": "<p>so, their InchI will also be different from what are in train, so how our model will identify these new class names?</p>",
      "rawMarkdown": "so, their InchI will also be different from what are in train, so how our model will identify these new class names?",
      "votes": null
    },
    {
      "id": "1283005",
      "postDate": "04/24/2021 13:44:33",
      "content": "<p>can you please give an idea how your are approaching this problem?</p>",
      "rawMarkdown": "can you please give an idea how your are approaching this problem?",
      "votes": null
    },
    {
      "id": "1283009",
      "postDate": "04/24/2021 13:49:05",
      "content": "<p>I recommend you to go to the Code section and sort by votes; pick the 1st one, for example, there you can get an idea. : )</p>",
      "rawMarkdown": "I recommend you to go to the Code section and sort by votes; pick the 1st one, for example, there you can get an idea. : )",
      "votes": null
    },
    {
      "id": "1283024",
      "postDate": "04/24/2021 14:02:39",
      "content": "<p>not all problems can be treated as classification.</p>\n<p>in this case, your test class has truth label: abc,def,acd,eff<br>\nyour training examples are aab,fef,ccd,ace</p>\n<p>none of the labels are the same.<br>\n but we note that the labels are made up of combinations of a,b,c,d,e,f</p>\n<p>hence this is a seq learning problem. We are solving two problems:</p>\n<ul>\n<li>how to identify each of the individual \"parts\" a,b,c,d …</li>\n<li>how to identify the sequence (target label), e.g. abc? def? …aaa, abc?</li>\n</ul>",
      "rawMarkdown": "not all problems can be treated as classification.\n\nin this case, your test class has truth label: abc,def,acd,eff\nyour training examples are aab,fef,ccd,ace\n\nnone of the labels are the same.\n but we note that the labels are made up of combinations of a,b,c,d,e,f\n\nhence this is a seq learning problem. We are solving two problems:\n- how to identify each of the individual \"parts\" a,b,c,d ...\n- how to identify the sequence (target label), e.g. abc? def? ...aaa, abc?",
      "votes": null
    },
    {
      "id": "1283405",
      "postDate": "04/24/2021 22:25:34",
      "content": "<p>There are multiple approaches, but a state of the art method is to encode the images with a CNN and to decode it with an RNN (mostly LSTMs) to predict the string sequence based on what the model has learned.</p>\n<p>I recommend looking at one of the starter notebooks e.g: <a href=\"https://www.kaggle.com/yasufuminakama/inchi-resnet-lstm-with-attention-starter\" target=\"_blank\">https://www.kaggle.com/yasufuminakama/inchi-resnet-lstm-with-attention-starter</a></p>",
      "rawMarkdown": "There are multiple approaches, but a state of the art method is to encode the images with a CNN and to decode it with an RNN (mostly LSTMs) to predict the string sequence based on what the model has learned.\n\nI recommend looking at one of the starter notebooks e.g: https://www.kaggle.com/yasufuminakama/inchi-resnet-lstm-with-attention-starter",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1282994,
      "author_name": "nofreewill",
      "author_url": "",
      "post_date": "04/24/2021 13:33:57",
      "content": "<blockquote>\n  <p>are test images are subset of train images?</p>\n</blockquote>\n<p>No, test molecules are also unique.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1283003,
          "author_name": "karan23258",
          "author_url": "",
          "post_date": "04/24/2021 13:42:12",
          "content": "<p>so, their InchI will also be different from what are in train, so how our model will identify these new class names?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1283005,
          "author_name": "karan23258",
          "author_url": "",
          "post_date": "04/24/2021 13:44:33",
          "content": "<p>can you please give an idea how your are approaching this problem?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1283009,
          "author_name": "nofreewill",
          "author_url": "",
          "post_date": "04/24/2021 13:49:05",
          "content": "<p>I recommend you to go to the Code section and sort by votes; pick the 1st one, for example, there you can get an idea. : )</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1283405,
          "author_name": "aliabdin1",
          "author_url": "",
          "post_date": "04/24/2021 22:25:34",
          "content": "<p>There are multiple approaches, but a state of the art method is to encode the images with a CNN and to decode it with an RNN (mostly LSTMs) to predict the string sequence based on what the model has learned.</p>\n<p>I recommend looking at one of the starter notebooks e.g: <a href=\"https://www.kaggle.com/yasufuminakama/inchi-resnet-lstm-with-attention-starter\" target=\"_blank\">https://www.kaggle.com/yasufuminakama/inchi-resnet-lstm-with-attention-starter</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1283024,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "04/24/2021 14:02:39",
      "content": "<p>not all problems can be treated as classification.</p>\n<p>in this case, your test class has truth label: abc,def,acd,eff<br>\nyour training examples are aab,fef,ccd,ace</p>\n<p>none of the labels are the same.<br>\n but we note that the labels are made up of combinations of a,b,c,d,e,f</p>\n<p>hence this is a seq learning problem. We are solving two problems:</p>\n<ul>\n<li>how to identify each of the individual \"parts\" a,b,c,d …</li>\n<li>how to identify the sequence (target label), e.g. abc? def? …aaa, abc?</li>\n</ul>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1282958": "Can we take it as a image classification problem? But, all InChI labels in train file are unique, i.e. we have just 1 example per class and totally about 2.42m classes. Will it be okay to train a classification model over the train files?\nAlso, as InChI are unique i.e. one for one structure, so are test images are subset of train images?\nPlease answer..",
    "1282994": "> are test images are subset of train images?\n\nNo, test molecules are also unique.",
    "1283003": "so, their InchI will also be different from what are in train, so how our model will identify these new class names?",
    "1283005": "can you please give an idea how your are approaching this problem?",
    "1283009": "I recommend you to go to the Code section and sort by votes; pick the 1st one, for example, there you can get an idea. : )",
    "1283024": "not all problems can be treated as classification.\n\nin this case, your test class has truth label: abc,def,acd,eff\nyour training examples are aab,fef,ccd,ace\n\nnone of the labels are the same.\n but we note that the labels are made up of combinations of a,b,c,d,e,f\n\nhence this is a seq learning problem. We are solving two problems:\n- how to identify each of the individual \"parts\" a,b,c,d ...\n- how to identify the sequence (target label), e.g. abc? def? ...aaa, abc?",
    "1283405": "There are multiple approaches, but a state of the art method is to encode the images with a CNN and to decode it with an RNN (mostly LSTMs) to predict the string sequence based on what the model has learned.\n\nI recommend looking at one of the starter notebooks e.g: https://www.kaggle.com/yasufuminakama/inchi-resnet-lstm-with-attention-starter"
  },
  "source": "meta"
}