{
  "id": 104138,
  "title": "How unstable are GANs?",
  "url": "/competitions/generative-dog-images/discussion/104138",
  "author_name": "Daniil Barysevich",
  "post_date": "2019-08-14T16:36:03.811000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I noticed that training the same GAN with slightly different parameters (another fixed random seed or number of epochs) can cause LB score to change up to ~25. Maybe, I have done something wrong, but this behavior occurs weird to me and fiddling with random seed until a good score does not feel like a good strategy.</p>\n\n<p>What is your experience with reproducing results and conducting experiments not to be fooled by too much randomness? </p>",
  "messages": [
    {
      "id": 599194,
      "postDate": "2019-08-14T16:36:03.810Z",
      "content": "<p>I noticed that training the same GAN with slightly different parameters (another fixed random seed or number of epochs) can cause LB score to change up to ~25. Maybe, I have done something wrong, but this behavior occurs weird to me and fiddling with random seed until a good score does not feel like a good strategy.</p>\n\n<p>What is your experience with reproducing results and conducting experiments not to be fooled by too much randomness? </p>",
      "rawMarkdown": "I noticed that training the same GAN with slightly different parameters (another fixed random seed or number of epochs) can cause LB score to change up to ~25. Maybe, I have done something wrong, but this behavior occurs weird to me and fiddling with random seed until a good score does not feel like a good strategy.\n\nWhat is your experience with reproducing results and conducting experiments not to be fooled by too much randomness? ",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "599194": "I noticed that training the same GAN with slightly different parameters (another fixed random seed or number of epochs) can cause LB score to change up to ~25. Maybe, I have done something wrong, but this behavior occurs weird to me and fiddling with random seed until a good score does not feel like a good strategy.\n\nWhat is your experience with reproducing results and conducting experiments not to be fooled by too much randomness? "
  }
}