{
  "id": 166121,
  "title": "FixEfficientNetB7 Noisy-student ",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/166121",
  "author_name": "",
  "post_date": "2020-07-11T20:43:36.091770100Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I have reviewed the IMAGENET rankings. And notice with the existing notebook resources FixEfficientnetB7 is the best model ( size + performance). Has anyone tried to use this architecture?\n<a href=\"https://arxiv.org/pdf/2003.08237v4.pdf\">https://arxiv.org/pdf/2003.08237v4.pdf</a></p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3544701%2F257697a23cdc58dda537eae3cf692d76%2FScreenshot_20200711_233855.png?generation=1594500346042050&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": "925136",
      "postDate": "07/11/2020 20:43:36",
      "content": "<p>I have reviewed the IMAGENET rankings. And notice with the existing notebook resources FixEfficientnetB7 is the best model ( size + performance). Has anyone tried to use this architecture?\n<a href=\"https://arxiv.org/pdf/2003.08237v4.pdf\">https://arxiv.org/pdf/2003.08237v4.pdf</a></p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3544701%2F257697a23cdc58dda537eae3cf692d76%2FScreenshot_20200711_233855.png?generation=1594500346042050&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "I have reviewed the IMAGENET rankings. And notice with the existing notebook resources FixEfficientnetB7 is the best model ( size + performance). Has anyone tried to use this architecture?\nhttps://arxiv.org/pdf/2003.08237v4.pdf\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3544701%2F257697a23cdc58dda537eae3cf692d76%2FScreenshot_20200711_233855.png?generation=1594500346042050&amp;alt=media)",
      "votes": null
    },
    {
      "id": "925163",
      "postDate": "07/11/2020 21:17:51",
      "content": "<p>It is the same architecture, its only a finetuning method to infer at higher resolution when using RandomResizedCrop data augmentation. This paper explains it: <a href=\"https://arxiv.org/pdf/1906.06423.pdf\">https://arxiv.org/pdf/1906.06423.pdf</a></p>\n\n<p>I have tried it without the finetuning part which is not mandatory and it gives me a CV boost</p>",
      "rawMarkdown": "It is the same architecture, its only a finetuning method to infer at higher resolution when using RandomResizedCrop data augmentation. This paper explains it: https://arxiv.org/pdf/1906.06423.pdf\n\nI have tried it without the finetuning part which is not mandatory and it gives me a CV boost",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 925163,
      "author_name": "yannmajewski",
      "author_url": "",
      "post_date": "07/11/2020 21:17:51",
      "content": "<p>It is the same architecture, its only a finetuning method to infer at higher resolution when using RandomResizedCrop data augmentation. This paper explains it: <a href=\"https://arxiv.org/pdf/1906.06423.pdf\">https://arxiv.org/pdf/1906.06423.pdf</a></p>\n\n<p>I have tried it without the finetuning part which is not mandatory and it gives me a CV boost</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "925136": "I have reviewed the IMAGENET rankings. And notice with the existing notebook resources FixEfficientnetB7 is the best model ( size + performance). Has anyone tried to use this architecture?\nhttps://arxiv.org/pdf/2003.08237v4.pdf\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3544701%2F257697a23cdc58dda537eae3cf692d76%2FScreenshot_20200711_233855.png?generation=1594500346042050&amp;alt=media)",
    "925163": "It is the same architecture, its only a finetuning method to infer at higher resolution when using RandomResizedCrop data augmentation. This paper explains it: https://arxiv.org/pdf/1906.06423.pdf\n\nI have tried it without the finetuning part which is not mandatory and it gives me a CV boost"
  },
  "source": "meta"
}