{
  "id": 110368,
  "title": "What augmentation techniques were crucial in this competition?",
  "url": "/competitions/recursion-cellular-image-classification/discussion/110368",
  "author_name": "",
  "post_date": "2019-09-27T06:45:52.066074900Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I tried standard RandomCrop, HorizontalFlip, Rotate90, and also normalized pictures, using pixel_stats.csv means and stds. But I see that some of the competitors were using brightness and contrast augmentations. Was it really helpful?</p>",
  "messages": [
    {
      "id": "635097",
      "postDate": "09/27/2019 06:45:52",
      "content": "<p>I tried standard RandomCrop, HorizontalFlip, Rotate90, and also normalized pictures, using pixel_stats.csv means and stds. But I see that some of the competitors were using brightness and contrast augmentations. Was it really helpful?</p>",
      "rawMarkdown": "I tried standard RandomCrop, HorizontalFlip, Rotate90, and also normalized pictures, using pixel_stats.csv means and stds. But I see that some of the competitors were using brightness and contrast augmentations. Was it really helpful?",
      "votes": null
    },
    {
      "id": "635129",
      "postDate": "09/27/2019 07:24:14",
      "content": "<p>Hi, for our solution hard augmentations with p=0.8 for brightness/contrast were crucial. Also RandomGamma and ChannelDropout seemed to be helpful. </p>",
      "rawMarkdown": "Hi, for our solution hard augmentations with p=0.8 for brightness/contrast were crucial. Also RandomGamma and ChannelDropout seemed to be helpful.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 635129,
      "author_name": "yaroshevskiy",
      "author_url": "",
      "post_date": "09/27/2019 07:24:14",
      "content": "<p>Hi, for our solution hard augmentations with p=0.8 for brightness/contrast were crucial. Also RandomGamma and ChannelDropout seemed to be helpful. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "635097": "I tried standard RandomCrop, HorizontalFlip, Rotate90, and also normalized pictures, using pixel_stats.csv means and stds. But I see that some of the competitors were using brightness and contrast augmentations. Was it really helpful?",
    "635129": "Hi, for our solution hard augmentations with p=0.8 for brightness/contrast were crucial. Also RandomGamma and ChannelDropout seemed to be helpful."
  },
  "source": "meta"
}