{
  "id": 99494,
  "title": "Why control wells were not added to the TFRecords?",
  "url": "/competitions/recursion-cellular-image-classification/discussion/99494",
  "author_name": "",
  "post_date": "2019-07-11T18:11:39.349311800Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I'm looking through the TFRecords and it seems like the control wells are not there.\nLooking at the code that parses the images, I have noticed the following line:</p>\n\n<p>```</p>\n\n<h1>only pack images for the treatment wells, not the controls!</h1>\n\n<pre><code>metadata_df = metadata_df[metadata_df.well_type == \"treatment\"]\n</code></pre>\n\n<p>```</p>\n\n<p>Was there a reason not to include the control wells in the TFRecord dataset?</p>",
  "messages": [
    {
      "id": "573050",
      "postDate": "07/11/2019 18:11:39",
      "content": "<p>I'm looking through the TFRecords and it seems like the control wells are not there.\nLooking at the code that parses the images, I have noticed the following line:</p>\n\n<p>```</p>\n\n<h1>only pack images for the treatment wells, not the controls!</h1>\n\n<pre><code>metadata_df = metadata_df[metadata_df.well_type == \"treatment\"]\n</code></pre>\n\n<p>```</p>\n\n<p>Was there a reason not to include the control wells in the TFRecord dataset?</p>",
      "rawMarkdown": "I'm looking through the TFRecords and it seems like the control wells are not there.\nLooking at the code that parses the images, I have noticed the following line:\n\n```\n# only pack images for the treatment wells, not the controls!\n    metadata_df = metadata_df[metadata_df.well_type == \"treatment\"]\n```\n\nWas there a reason not to include the control wells in the TFRecord dataset?",
      "votes": null
    },
    {
      "id": "573268",
      "postDate": "07/12/2019 03:53:18",
      "content": "<p>If they were included in the TFRecords you would have to filter them out when training a classifier on the treatment wells. Likewise, if you wanted just the control wells you would need to scan through all of the TFRecords just to pull those out. That said, the one reason why we released the TFRecord generation code was that it would be easy for people to pack other versions of the dataset if you want to. </p>",
      "rawMarkdown": "If they were included in the TFRecords you would have to filter them out when training a classifier on the treatment wells. Likewise, if you wanted just the control wells you would need to scan through all of the TFRecords just to pull those out. That said, the one reason why we released the TFRecord generation code was that it would be easy for people to pack other versions of the dataset if you want to.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 573268,
      "author_name": "bmabey",
      "author_url": "",
      "post_date": "07/12/2019 03:53:18",
      "content": "<p>If they were included in the TFRecords you would have to filter them out when training a classifier on the treatment wells. Likewise, if you wanted just the control wells you would need to scan through all of the TFRecords just to pull those out. That said, the one reason why we released the TFRecord generation code was that it would be easy for people to pack other versions of the dataset if you want to. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "573050": "I'm looking through the TFRecords and it seems like the control wells are not there.\nLooking at the code that parses the images, I have noticed the following line:\n\n```\n# only pack images for the treatment wells, not the controls!\n    metadata_df = metadata_df[metadata_df.well_type == \"treatment\"]\n```\n\nWas there a reason not to include the control wells in the TFRecord dataset?",
    "573268": "If they were included in the TFRecords you would have to filter them out when training a classifier on the treatment wells. Likewise, if you wanted just the control wells you would need to scan through all of the TFRecords just to pull those out. That said, the one reason why we released the TFRecord generation code was that it would be easy for people to pack other versions of the dataset if you want to."
  },
  "source": "meta"
}