{
  "id": 37737,
  "title": "Unsupervised pre-training, does it make sense here?",
  "url": "/competitions/carvana-image-masking-challenge/discussion/37737",
  "author_name": "",
  "post_date": "2017-08-08T11:22:44.753360900Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Considering we have 100K images in test vs. and ~5K in training/validation it could potentially make sense to use unsupervised pre-training. </p>\n\n<p>Has anybody tried it?</p>\n\n<p>See e.g. <a href=\"https://arxiv.org/pdf/1611.09119.pdf\">https://arxiv.org/pdf/1611.09119.pdf</a> (Mar 2017)</p>",
  "messages": [
    {
      "id": "211230",
      "postDate": "08/08/2017 11:22:44",
      "content": "<p>Considering we have 100K images in test vs. and ~5K in training/validation it could potentially make sense to use unsupervised pre-training. </p>\n\n<p>Has anybody tried it?</p>\n\n<p>See e.g. <a href=\"https://arxiv.org/pdf/1611.09119.pdf\">https://arxiv.org/pdf/1611.09119.pdf</a> (Mar 2017)</p>",
      "rawMarkdown": "Considering we have 100K images in test vs. and ~5K in training/validation it could potentially make sense to use unsupervised pre-training. \n\nHas anybody tried it?\n\nSee e.g. https://arxiv.org/pdf/1611.09119.pdf (Mar 2017)",
      "votes": null
    },
    {
      "id": "211499",
      "postDate": "08/09/2017 08:04:45",
      "content": "<p>Well\nIt works good for weight initializing but in this competition the score is already 99.7 to achieve a better one - we need co solve the problem with the shadows etc\nCRF can probably help here - but I do not think unsupervised will solve this problem</p>",
      "rawMarkdown": "Well\nIt works good for weight initializing but in this competition the score is already 99.7 to achieve a better one - we need co solve the problem with the shadows etc\nCRF can probably help here - but I do not think unsupervised will solve this problem",
      "votes": null
    },
    {
      "id": "211702",
      "postDate": "08/09/2017 18:55:33",
      "content": "<p>Thanks!</p>",
      "rawMarkdown": "Thanks!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 211499,
      "author_name": "venheads",
      "author_url": "",
      "post_date": "08/09/2017 08:04:45",
      "content": "<p>Well\nIt works good for weight initializing but in this competition the score is already 99.7 to achieve a better one - we need co solve the problem with the shadows etc\nCRF can probably help here - but I do not think unsupervised will solve this problem</p>",
      "votes": null,
      "replies": [
        {
          "id": 211702,
          "author_name": "antorsae",
          "author_url": "",
          "post_date": "08/09/2017 18:55:33",
          "content": "<p>Thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "211230": "Considering we have 100K images in test vs. and ~5K in training/validation it could potentially make sense to use unsupervised pre-training. \n\nHas anybody tried it?\n\nSee e.g. https://arxiv.org/pdf/1611.09119.pdf (Mar 2017)",
    "211499": "Well\nIt works good for weight initializing but in this competition the score is already 99.7 to achieve a better one - we need co solve the problem with the shadows etc\nCRF can probably help here - but I do not think unsupervised will solve this problem",
    "211702": "Thanks!"
  },
  "source": "meta"
}