{
  "id": 19996,
  "title": "can we train test set?",
  "url": "/competitions/state-farm-distracted-driver-detection/discussion/19996",
  "author_name": "",
  "post_date": "2016-04-07T14:47:12.580Z",
  "votes": null,
  "comment_count": 2,
  "views": 1009,
  "content": "<p>Is it allowed to train test set with pseudo labels (not hand labeled) to predict itself again?\nThank you very much!</p>",
  "messages": [
    {
      "id": "114093",
      "postDate": "04/07/2016 14:47:12",
      "content": "<p>Is it allowed to train test set with pseudo labels (not hand labeled) to predict itself again?\nThank you very much!</p>",
      "rawMarkdown": "Is it allowed to train test set with pseudo labels (not hand labeled) to predict itself again?\r\nThank you very much!",
      "votes": null
    },
    {
      "id": "114114",
      "postDate": "04/07/2016 15:39:59",
      "content": "<p>This is usually allowed, along with other semi-supervised methods applied to train+test (such as autoencoders or pre-training layers with RBM). As long as you don't use manual approach to select e.g. which test images to use. </p>\n\n<p>You should note that there are many dummy/unscored images in the test set that <em>might</em> reduce effectiveness of semi-supervised approaches, and you won't be able to manually remove them.</p>",
      "rawMarkdown": "This is usually allowed, along with other semi-supervised methods applied to train+test (such as autoencoders or pre-training layers with RBM). As long as you don't use manual approach to select e.g. which test images to use. \r\n\r\nYou should note that there are many dummy/unscored images in the test set that *might* reduce effectiveness of semi-supervised approaches, and you won't be able to manually remove them.",
      "votes": null
    },
    {
      "id": "114116",
      "postDate": "04/07/2016 16:01:40",
      "content": "<p>The general guideline is you should be able to feed your pipeline new images (i.e., ones it has never seen before), and your pipeline should be able to predict that image without any manual intervention.</p>",
      "rawMarkdown": "The general guideline is you should be able to feed your pipeline new images (i.e., ones it has never seen before), and your pipeline should be able to predict that image without any manual intervention.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 114114,
      "author_name": "slobo777",
      "author_url": "",
      "post_date": "04/07/2016 15:39:59",
      "content": "<p>This is usually allowed, along with other semi-supervised methods applied to train+test (such as autoencoders or pre-training layers with RBM). As long as you don't use manual approach to select e.g. which test images to use. </p>\n\n<p>You should note that there are many dummy/unscored images in the test set that <em>might</em> reduce effectiveness of semi-supervised approaches, and you won't be able to manually remove them.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 114116,
      "author_name": "inversion",
      "author_url": "",
      "post_date": "04/07/2016 16:01:40",
      "content": "<p>The general guideline is you should be able to feed your pipeline new images (i.e., ones it has never seen before), and your pipeline should be able to predict that image without any manual intervention.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "114093": "Is it allowed to train test set with pseudo labels (not hand labeled) to predict itself again?\r\nThank you very much!",
    "114114": "This is usually allowed, along with other semi-supervised methods applied to train+test (such as autoencoders or pre-training layers with RBM). As long as you don't use manual approach to select e.g. which test images to use. \r\n\r\nYou should note that there are many dummy/unscored images in the test set that *might* reduce effectiveness of semi-supervised approaches, and you won't be able to manually remove them.",
    "114116": "The general guideline is you should be able to feed your pipeline new images (i.e., ones it has never seen before), and your pipeline should be able to predict that image without any manual intervention."
  },
  "source": "meta"
}