{
  "id": 91110,
  "title": "When Extracting Image Features",
  "url": "/competitions/imet-2019-fgvc6/discussion/91110",
  "author_name": "",
  "post_date": "2019-05-01T03:53:04.260456Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Is it right to extract Image features before doing data augmentation?\nTo be clear, I'm doing some feature engineering on my images and storing all in a single DataFrame.\nBut in my DataGenerator, I feed to the CNN each time the image after Augmentation and the features that were extracted before.</p>\n\n<p>Note that, When I try to extract image features in the generator, it becomes very slow.</p>\n\n<p>Does anyone experience these same issues?</p>",
  "messages": [
    {
      "id": "525454",
      "postDate": "05/01/2019 03:53:04",
      "content": "<p>Is it right to extract Image features before doing data augmentation?\nTo be clear, I'm doing some feature engineering on my images and storing all in a single DataFrame.\nBut in my DataGenerator, I feed to the CNN each time the image after Augmentation and the features that were extracted before.</p>\n\n<p>Note that, When I try to extract image features in the generator, it becomes very slow.</p>\n\n<p>Does anyone experience these same issues?</p>",
      "rawMarkdown": "Is it right to extract Image features before doing data augmentation?\nTo be clear, I'm doing some feature engineering on my images and storing all in a single DataFrame.\nBut in my DataGenerator, I feed to the CNN each time the image after Augmentation and the features that were extracted before.\n\nNote that, When I try to extract image features in the generator, it becomes very slow.\n\nDoes anyone experience these same issues?",
      "votes": null
    },
    {
      "id": "526589",
      "postDate": "05/03/2019 10:36:53",
      "content": "<p>I am not an expert in image features, but in general, you first do augmentation and then extract features, that's the whole point of increasing your training dataset size. \nIf it's too slow try to augment just several times and then extract features and then see the change in CV and LB. </p>",
      "rawMarkdown": "I am not an expert in image features, but in general, you first do augmentation and then extract features, that's the whole point of increasing your training dataset size. \nIf it's too slow try to augment just several times and then extract features and then see the change in CV and LB.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 526589,
      "author_name": "blondinka",
      "author_url": "",
      "post_date": "05/03/2019 10:36:53",
      "content": "<p>I am not an expert in image features, but in general, you first do augmentation and then extract features, that's the whole point of increasing your training dataset size. \nIf it's too slow try to augment just several times and then extract features and then see the change in CV and LB. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "525454": "Is it right to extract Image features before doing data augmentation?\nTo be clear, I'm doing some feature engineering on my images and storing all in a single DataFrame.\nBut in my DataGenerator, I feed to the CNN each time the image after Augmentation and the features that were extracted before.\n\nNote that, When I try to extract image features in the generator, it becomes very slow.\n\nDoes anyone experience these same issues?",
    "526589": "I am not an expert in image features, but in general, you first do augmentation and then extract features, that's the whole point of increasing your training dataset size. \nIf it's too slow try to augment just several times and then extract features and then see the change in CV and LB."
  },
  "source": "meta"
}