{
  "id": 225358,
  "title": "DECIMER from Images to SMILES",
  "url": "/competitions/bms-molecular-translation/discussion/225358",
  "author_name": "Alexander Scarlat MD",
  "post_date": "2021-03-11T21:52:08.271000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>As many have mentioned this tool - <a href=\"https://github.com/Kohulan/DECIMER-Image-to-SMILES\" target=\"_blank\">https://github.com/Kohulan/DECIMER-Image-to-SMILES</a> I've decided to see if it can be used for this competition. <br>\nThe idea is to get from images to SMILES and then with rdkit to Mol and finally to InChI.</p>\n<p>I've cloned the original github repo and it works by identifying an image into SMILES.<br>\nHowever, when tried with the competition images and then translating with rdkit from SMILES into Mol and then into InCHI - the Levenshtein distance is very high.</p>\n<p>The author mentioned the DECIMER model (which was trained on 15 million molecules for many days of GPU) - fails with Kaggle competition images, but didn't mention any reason: </p>\n<p><a href=\"https://github.com/Kohulan/DECIMER-Image-to-SMILES/issues/6\" target=\"_blank\">https://github.com/Kohulan/DECIMER-Image-to-SMILES/issues/6</a></p>\n<p>Did anybody succeed in implementing DECIMER for the BMS competition ?</p>\n<p>Thanks</p>",
  "messages": [
    {
      "id": 1235152,
      "postDate": "2021-03-11T21:52:08.270Z",
      "content": "<p>As many have mentioned this tool - <a href=\"https://github.com/Kohulan/DECIMER-Image-to-SMILES\" target=\"_blank\">https://github.com/Kohulan/DECIMER-Image-to-SMILES</a> I've decided to see if it can be used for this competition. <br>\nThe idea is to get from images to SMILES and then with rdkit to Mol and finally to InChI.</p>\n<p>I've cloned the original github repo and it works by identifying an image into SMILES.<br>\nHowever, when tried with the competition images and then translating with rdkit from SMILES into Mol and then into InCHI - the Levenshtein distance is very high.</p>\n<p>The author mentioned the DECIMER model (which was trained on 15 million molecules for many days of GPU) - fails with Kaggle competition images, but didn't mention any reason: </p>\n<p><a href=\"https://github.com/Kohulan/DECIMER-Image-to-SMILES/issues/6\" target=\"_blank\">https://github.com/Kohulan/DECIMER-Image-to-SMILES/issues/6</a></p>\n<p>Did anybody succeed in implementing DECIMER for the BMS competition ?</p>\n<p>Thanks</p>",
      "rawMarkdown": "As many have mentioned this tool - https://github.com/Kohulan/DECIMER-Image-to-SMILES I've decided to see if it can be used for this competition. \nThe idea is to get from images to SMILES and then with rdkit to Mol and finally to InChI.\n\nI've cloned the original github repo and it works by identifying an image into SMILES.\nHowever, when tried with the competition images and then translating with rdkit from SMILES into Mol and then into InCHI - the Levenshtein distance is very high.\n\nThe author mentioned the DECIMER model (which was trained on 15 million molecules for many days of GPU) - fails with Kaggle competition images, but didn't mention any reason: \n\nhttps://github.com/Kohulan/DECIMER-Image-to-SMILES/issues/6\n\nDid anybody succeed in implementing DECIMER for the BMS competition ?\n\nThanks",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1235152": "As many have mentioned this tool - https://github.com/Kohulan/DECIMER-Image-to-SMILES I've decided to see if it can be used for this competition. \nThe idea is to get from images to SMILES and then with rdkit to Mol and finally to InChI.\n\nI've cloned the original github repo and it works by identifying an image into SMILES.\nHowever, when tried with the competition images and then translating with rdkit from SMILES into Mol and then into InCHI - the Levenshtein distance is very high.\n\nThe author mentioned the DECIMER model (which was trained on 15 million molecules for many days of GPU) - fails with Kaggle competition images, but didn't mention any reason: \n\nhttps://github.com/Kohulan/DECIMER-Image-to-SMILES/issues/6\n\nDid anybody succeed in implementing DECIMER for the BMS competition ?\n\nThanks"
  }
}