{
  "id": 233448,
  "title": "Scaling Up Models",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/233448",
  "author_name": "",
  "post_date": "2021-04-19T11:11:03.381176300Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>When working with the EfficientNet family, I can easily achieve a decent CV(92.8% CV) using a B0, but the second I scale up the model(B2, B3, B4), the CV drops immediately. Is anyone else facing this issue. If not, how did you combat this issue of larger EfficientNets failing to train?</p>",
  "messages": [
    {
      "id": "1277909",
      "postDate": "04/19/2021 11:11:03",
      "content": "<p>When working with the EfficientNet family, I can easily achieve a decent CV(92.8% CV) using a B0, but the second I scale up the model(B2, B3, B4), the CV drops immediately. Is anyone else facing this issue. If not, how did you combat this issue of larger EfficientNets failing to train?</p>",
      "rawMarkdown": "When working with the EfficientNet family, I can easily achieve a decent CV(92.8% CV) using a B0, but the second I scale up the model(B2, B3, B4), the CV drops immediately. Is anyone else facing this issue. If not, how did you combat this issue of larger EfficientNets failing to train?",
      "votes": null
    },
    {
      "id": "1279257",
      "postDate": "04/20/2021 18:38:30",
      "content": "<p>I'm actually getting very similar cvs for b0 and b1, but still b1 is the model i'm getting the best results. For higher models i'm also getting lower cvs</p>",
      "rawMarkdown": "I'm actually getting very similar cvs for b0 and b1, but still b1 is the model i'm getting the best results. For higher models i'm also getting lower cvs",
      "votes": null
    },
    {
      "id": "1280487",
      "postDate": "04/22/2021 03:09:33",
      "content": "<p>lighter model will be easier overfiting?</p>",
      "rawMarkdown": "lighter model will be easier overfiting?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1279257,
      "author_name": "victorasso",
      "author_url": "",
      "post_date": "04/20/2021 18:38:30",
      "content": "<p>I'm actually getting very similar cvs for b0 and b1, but still b1 is the model i'm getting the best results. For higher models i'm also getting lower cvs</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1280487,
      "author_name": "southsakura",
      "author_url": "",
      "post_date": "04/22/2021 03:09:33",
      "content": "<p>lighter model will be easier overfiting?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1277909": "When working with the EfficientNet family, I can easily achieve a decent CV(92.8% CV) using a B0, but the second I scale up the model(B2, B3, B4), the CV drops immediately. Is anyone else facing this issue. If not, how did you combat this issue of larger EfficientNets failing to train?",
    "1279257": "I'm actually getting very similar cvs for b0 and b1, but still b1 is the model i'm getting the best results. For higher models i'm also getting lower cvs",
    "1280487": "lighter model will be easier overfiting?"
  },
  "source": "meta"
}