{
  "id": 93709,
  "title": "fast.ai ootb cont...",
  "url": "/competitions/inaturalist-2019-fgvc6/discussion/93709",
  "author_name": "khursani",
  "post_date": "2019-05-29T11:52:54.747000",
  "votes": 5,
  "comment_count": 0,
  "views": 0,
  "content": "<p>After I don't have any idea for improving my kernel score, I decide to train my model offline for a bit longer. Below is what I did so far:</p>\n\n<ol>\n<li>Train from scratch with 3 epochs for model head and 14 epochs for unfreeze model (0.39265)</li>\n<li>Continue for 14 epochs w/o mixup (0.38429)</li>\n<li>Train 2 epochs with full dataset instead of 20% of the dataset (0.33699) </li>\n</ol>\n\n<p>Will update this post in future.</p>",
  "messages": [
    {
      "id": 539008,
      "postDate": "2019-05-29T11:52:54.747Z",
      "content": "<p>After I don't have any idea for improving my kernel score, I decide to train my model offline for a bit longer. Below is what I did so far:</p>\n\n<ol>\n<li>Train from scratch with 3 epochs for model head and 14 epochs for unfreeze model (0.39265)</li>\n<li>Continue for 14 epochs w/o mixup (0.38429)</li>\n<li>Train 2 epochs with full dataset instead of 20% of the dataset (0.33699) </li>\n</ol>\n\n<p>Will update this post in future.</p>",
      "rawMarkdown": "After I don't have any idea for improving my kernel score, I decide to train my model offline for a bit longer. Below is what I did so far:\n\n1. Train from scratch with 3 epochs for model head and 14 epochs for unfreeze model (0.39265)\n2. Continue for 14 epochs w/o mixup (0.38429)\n3. Train 2 epochs with full dataset instead of 20% of the dataset (0.33699) \n\nWill update this post in future.",
      "votes": 5
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "539008": "After I don't have any idea for improving my kernel score, I decide to train my model offline for a bit longer. Below is what I did so far:\n\n1. Train from scratch with 3 epochs for model head and 14 epochs for unfreeze model (0.39265)\n2. Continue for 14 epochs w/o mixup (0.38429)\n3. Train 2 epochs with full dataset instead of 20% of the dataset (0.33699) \n\nWill update this post in future."
  }
}