{
  "id": 27751,
  "title": "How do you validate?",
  "url": "/competitions/dstl-satellite-imagery-feature-detection/discussion/27751",
  "author_name": "",
  "post_date": "2017-01-15T03:54:40.517Z",
  "votes": 1,
  "comment_count": 2,
  "views": 390,
  "content": "<p>How do you guys validate your models? </p>\n\n<p>I am training my models picking random crops of the size 224x224 and the doing similar for validation.</p>\n\n<p>So it definitely happens that there may be an overlap between train and validation images which makes evaluation process less reliable.</p>\n\n<p>Another way would be to partition train / validation based on the images.</p>\n\n<p>But because some classes are not represented well we may end with the situation when we do not have enough training data, say waterways.</p>\n\n<p>So, what is your reliable approach?</p>",
  "messages": [
    {
      "id": "156238",
      "postDate": "01/15/2017 03:54:40",
      "content": "<p>How do you guys validate your models? </p>\n\n<p>I am training my models picking random crops of the size 224x224 and the doing similar for validation.</p>\n\n<p>So it definitely happens that there may be an overlap between train and validation images which makes evaluation process less reliable.</p>\n\n<p>Another way would be to partition train / validation based on the images.</p>\n\n<p>But because some classes are not represented well we may end with the situation when we do not have enough training data, say waterways.</p>\n\n<p>So, what is your reliable approach?</p>",
      "rawMarkdown": "How do you guys validate your models? \r\n\r\nI am training my models picking random crops of the size 224x224 and the doing similar for validation.\r\n\r\nSo it definitely happens that there may be an overlap between train and validation images which makes evaluation process less reliable.\r\n\r\nAnother way would be to partition train / validation based on the images.\r\n\r\nBut because some classes are not represented well we may end with the situation when we do not have enough training data, say waterways.\r\n\r\nSo, what is your reliable approach?",
      "votes": null
    },
    {
      "id": "156599",
      "postDate": "01/17/2017 07:21:31",
      "content": "<p>I've been using a couple of images to validate and training with patches from all others.  So nothing fancy so far...still trying to get some models tuned properly. Admittedly slow progress!</p>",
      "rawMarkdown": "I've been using a couple of images to validate and training with patches from all others.  So nothing fancy so far...still trying to get some models tuned properly. Admittedly slow progress!",
      "votes": null
    },
    {
      "id": "156832",
      "postDate": "01/18/2017 02:52:51",
      "content": "<p>Is there any possibility to find similarity/clustering of images into 2 parts and so based on which we can make sure that images which are similar will form part of the same set?</p>",
      "rawMarkdown": "Is there any possibility to find similarity/clustering of images into 2 parts and so based on which we can make sure that images which are similar will form part of the same set?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 156599,
      "author_name": "zerozero",
      "author_url": "",
      "post_date": "01/17/2017 07:21:31",
      "content": "<p>I've been using a couple of images to validate and training with patches from all others.  So nothing fancy so far...still trying to get some models tuned properly. Admittedly slow progress!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 156832,
      "author_name": "puneetjindal",
      "author_url": "",
      "post_date": "01/18/2017 02:52:51",
      "content": "<p>Is there any possibility to find similarity/clustering of images into 2 parts and so based on which we can make sure that images which are similar will form part of the same set?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "156238": "How do you guys validate your models? \r\n\r\nI am training my models picking random crops of the size 224x224 and the doing similar for validation.\r\n\r\nSo it definitely happens that there may be an overlap between train and validation images which makes evaluation process less reliable.\r\n\r\nAnother way would be to partition train / validation based on the images.\r\n\r\nBut because some classes are not represented well we may end with the situation when we do not have enough training data, say waterways.\r\n\r\nSo, what is your reliable approach?",
    "156599": "I've been using a couple of images to validate and training with patches from all others.  So nothing fancy so far...still trying to get some models tuned properly. Admittedly slow progress!",
    "156832": "Is there any possibility to find similarity/clustering of images into 2 parts and so based on which we can make sure that images which are similar will form part of the same set?"
  },
  "source": "meta"
}