{
  "id": 28042,
  "title": "Class distribution in Training set and Strategies",
  "url": "/competitions/dstl-satellite-imagery-feature-detection/discussion/28042",
  "author_name": "",
  "post_date": "2017-01-23T17:43:16.187Z",
  "votes": 1,
  "comment_count": 1,
  "views": 166,
  "content": "<p>Hi,\nI just made mask over all training images and then looked at pixle distribution acrosse all 10 classes (see image), and is not even at all. \n<img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5832/figure_1.png\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>So I am wondering how are you dealing with it? Do you just downsample to particular class to get even distribution.\nOr are you training models for each class separately (which would also come handy as some classes overlap)?</p>",
  "messages": [
    {
      "id": "157775",
      "postDate": "01/23/2017 17:43:16",
      "content": "<p>Hi,\nI just made mask over all training images and then looked at pixle distribution acrosse all 10 classes (see image), and is not even at all. \n<img src=\"https://www.kaggle.com/blobs/download/forum-message-attachment-files/5832/figure_1.png\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>So I am wondering how are you dealing with it? Do you just downsample to particular class to get even distribution.\nOr are you training models for each class separately (which would also come handy as some classes overlap)?</p>",
      "rawMarkdown": "Hi,\r\nI just made mask over all training images and then looked at pixle distribution acrosse all 10 classes (see image), and is not even at all. \r\n![enter image description here][1]\r\n\r\nSo I am wondering how are you dealing with it? Do you just downsample to particular class to get even distribution.\r\nOr are you training models for each class separately (which would also come handy as some classes overlap)?\r\n\r\n\r\n  [1]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5832/figure_1.png",
      "votes": null
    },
    {
      "id": "157843",
      "postDate": "01/23/2017 22:49:52",
      "content": "<p>As you mention, I'm planning to train models separately for each class due primarily to the overlap issue you mention.  I've also seen a few other participants on the forum mentioning they attempted to train a model to predict all classes at once and that it did not work very well.</p>",
      "rawMarkdown": "As you mention, I'm planning to train models separately for each class due primarily to the overlap issue you mention.  I've also seen a few other participants on the forum mentioning they attempted to train a model to predict all classes at once and that it did not work very well.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 157843,
      "author_name": "michaeltoth1",
      "author_url": "",
      "post_date": "01/23/2017 22:49:52",
      "content": "<p>As you mention, I'm planning to train models separately for each class due primarily to the overlap issue you mention.  I've also seen a few other participants on the forum mentioning they attempted to train a model to predict all classes at once and that it did not work very well.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "157775": "Hi,\r\nI just made mask over all training images and then looked at pixle distribution acrosse all 10 classes (see image), and is not even at all. \r\n![enter image description here][1]\r\n\r\nSo I am wondering how are you dealing with it? Do you just downsample to particular class to get even distribution.\r\nOr are you training models for each class separately (which would also come handy as some classes overlap)?\r\n\r\n\r\n  [1]: https://www.kaggle.com/blobs/download/forum-message-attachment-files/5832/figure_1.png",
    "157843": "As you mention, I'm planning to train models separately for each class due primarily to the overlap issue you mention.  I've also seen a few other participants on the forum mentioning they attempted to train a model to predict all classes at once and that it did not work very well."
  },
  "source": "meta"
}