{
  "id": 138947,
  "title": "Does basic data augmentation really improve accuracy ?",
  "url": "/competitions/flower-classification-with-tpus/discussion/138947",
  "author_name": "",
  "post_date": "2020-03-26T21:24:29.267173Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi everyone,</p>\n\n<p>I trained 2 DenseNet-201, one with basic data augmentation (rotation, shift, flip etc ...) (model 1) and one without any data augmentation (model 2). </p>\n\n<p>After several attempts,  I noticed that model 1 performed almost the same or less good than model 2 (around +/- 0.01 difference).</p>\n\n<p>Does anyone have encounter the same problem ?</p>",
  "messages": [
    {
      "id": "787529",
      "postDate": "03/26/2020 21:24:29",
      "content": "<p>Hi everyone,</p>\n\n<p>I trained 2 DenseNet-201, one with basic data augmentation (rotation, shift, flip etc ...) (model 1) and one without any data augmentation (model 2). </p>\n\n<p>After several attempts,  I noticed that model 1 performed almost the same or less good than model 2 (around +/- 0.01 difference).</p>\n\n<p>Does anyone have encounter the same problem ?</p>",
      "rawMarkdown": "Hi everyone,\n\nI trained 2 DenseNet-201, one with basic data augmentation (rotation, shift, flip etc ...) (model 1) and one without any data augmentation (model 2). \n\nAfter several attempts,  I noticed that model 1 performed almost the same or less good than model 2 (around +/- 0.01 difference).\n\nDoes anyone have encounter the same problem ?",
      "votes": null
    },
    {
      "id": "788058",
      "postDate": "03/27/2020 10:45:52",
      "content": "<p>A couple of thoughts. Make sure that Model 1 is trained both with the unaltered data and the augmented. Also I haven't gotten to this point yet but I would assume with something like a flower you would want to resize or mess with the pixels. Rotating a round object won't change the image that much. </p>",
      "rawMarkdown": "A couple of thoughts. Make sure that Model 1 is trained both with the unaltered data and the augmented. Also I haven't gotten to this point yet but I would assume with something like a flower you would want to resize or mess with the pixels. Rotating a round object won't change the image that much.",
      "votes": null
    },
    {
      "id": "788187",
      "postDate": "03/27/2020 13:13:09",
      "content": "<p>Is there an easy way to use both original and training data simultaneously in keras when using a DataImageGenerator()?</p>",
      "rawMarkdown": "Is there an easy way to use both original and training data simultaneously in keras when using a DataImageGenerator()?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 788058,
      "author_name": "ericstraw",
      "author_url": "",
      "post_date": "03/27/2020 10:45:52",
      "content": "<p>A couple of thoughts. Make sure that Model 1 is trained both with the unaltered data and the augmented. Also I haven't gotten to this point yet but I would assume with something like a flower you would want to resize or mess with the pixels. Rotating a round object won't change the image that much. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 788187,
      "author_name": "jullang",
      "author_url": "",
      "post_date": "03/27/2020 13:13:09",
      "content": "<p>Is there an easy way to use both original and training data simultaneously in keras when using a DataImageGenerator()?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "787529": "Hi everyone,\n\nI trained 2 DenseNet-201, one with basic data augmentation (rotation, shift, flip etc ...) (model 1) and one without any data augmentation (model 2). \n\nAfter several attempts,  I noticed that model 1 performed almost the same or less good than model 2 (around +/- 0.01 difference).\n\nDoes anyone have encounter the same problem ?",
    "788058": "A couple of thoughts. Make sure that Model 1 is trained both with the unaltered data and the augmented. Also I haven't gotten to this point yet but I would assume with something like a flower you would want to resize or mess with the pixels. Rotating a round object won't change the image that much.",
    "788187": "Is there an easy way to use both original and training data simultaneously in keras when using a DataImageGenerator()?"
  },
  "source": "meta"
}