{
  "id": 45715,
  "title": "Methods that should've worked (but didn't)",
  "url": "/competitions/cdiscount-image-classification-challenge/discussion/45715",
  "author_name": "",
  "post_date": "2017-12-15T00:27:31.251172200Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>It seemed like this competition would have been ideal to use snapshot ensembles <a href=\"https://arxiv.org/pdf/1704.00109.pdf\">SNAPSHOT ENSEMBLES: TRAIN 1, GET M FOR FREE</a> however we weren't able to get it to work using a densenet implementation, perhaps we were too impatient though.  I was also surprised that additional augmentation beyond standard flips seemed to hurt performance.  Any other methods people tried that were surprising?</p>",
  "messages": [
    {
      "id": "257792",
      "postDate": "12/15/2017 00:27:31",
      "content": "<p>It seemed like this competition would have been ideal to use snapshot ensembles <a href=\"https://arxiv.org/pdf/1704.00109.pdf\">SNAPSHOT ENSEMBLES: TRAIN 1, GET M FOR FREE</a> however we weren't able to get it to work using a densenet implementation, perhaps we were too impatient though.  I was also surprised that additional augmentation beyond standard flips seemed to hurt performance.  Any other methods people tried that were surprising?</p>",
      "rawMarkdown": "It seemed like this competition would have been ideal to use snapshot ensembles [SNAPSHOT ENSEMBLES: TRAIN 1, GET M FOR FREE][1] however we weren't able to get it to work using a densenet implementation, perhaps we were too impatient though.  I was also surprised that additional augmentation beyond standard flips seemed to hurt performance.  Any other methods people tried that were surprising?\n\n\n  [1]: https://arxiv.org/pdf/1704.00109.pdf",
      "votes": null
    },
    {
      "id": "257805",
      "postDate": "12/15/2017 00:38:12",
      "content": "<p>I didn't find any type of affine transform augmentation to hurt performance, so long as I kept it within sane bounds and so long as I tuned on unaugmented data for the last couple of epochs. </p>",
      "rawMarkdown": "I didn't find any type of affine transform augmentation to hurt performance, so long as I kept it within sane bounds and so long as I tuned on unaugmented data for the last couple of epochs.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 257805,
      "author_name": "eachshadow",
      "author_url": "",
      "post_date": "12/15/2017 00:38:12",
      "content": "<p>I didn't find any type of affine transform augmentation to hurt performance, so long as I kept it within sane bounds and so long as I tuned on unaugmented data for the last couple of epochs. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "257792": "It seemed like this competition would have been ideal to use snapshot ensembles [SNAPSHOT ENSEMBLES: TRAIN 1, GET M FOR FREE][1] however we weren't able to get it to work using a densenet implementation, perhaps we were too impatient though.  I was also surprised that additional augmentation beyond standard flips seemed to hurt performance.  Any other methods people tried that were surprising?\n\n\n  [1]: https://arxiv.org/pdf/1704.00109.pdf",
    "257805": "I didn't find any type of affine transform augmentation to hurt performance, so long as I kept it within sane bounds and so long as I tuned on unaugmented data for the last couple of epochs."
  },
  "source": "meta"
}