{
  "id": 234429,
  "title": "Weird behaviour of LB score",
  "url": "/competitions/hotel-id-2021-fgvc8/discussion/234429",
  "author_name": "AleNic",
  "post_date": "2021-04-24T07:15:05.294000",
  "votes": 4,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hi, maybe it's my problem, bu I noticed a strong difference between my validation score and LB score. As a first solution I tried a very simple classifier and I obtained a similar mAP of public LB. After that I tried more advanced techniques like: arcmargin loss, triplet loss, multi head architectures, kmeans instance grouping, etc. and in my local validation I got much better results than the first simple solution as I expected, but I got worse on LB! <br>\nMaybe there is an error in my notebook…<br>\nHave you notice any strange behaviour?</p>",
  "messages": [
    {
      "id": 1282687,
      "postDate": "2021-04-24T07:15:05.293Z",
      "content": "<p>Hi, maybe it's my problem, bu I noticed a strong difference between my validation score and LB score. As a first solution I tried a very simple classifier and I obtained a similar mAP of public LB. After that I tried more advanced techniques like: arcmargin loss, triplet loss, multi head architectures, kmeans instance grouping, etc. and in my local validation I got much better results than the first simple solution as I expected, but I got worse on LB! <br>\nMaybe there is an error in my notebook…<br>\nHave you notice any strange behaviour?</p>",
      "rawMarkdown": "Hi, maybe it's my problem, bu I noticed a strong difference between my validation score and LB score. As a first solution I tried a very simple classifier and I obtained a similar mAP of public LB. After that I tried more advanced techniques like: arcmargin loss, triplet loss, multi head architectures, kmeans instance grouping, etc. and in my local validation I got much better results than the first simple solution as I expected, but I got worse on LB! \nMaybe there is an error in my notebook...\nHave you notice any strange behaviour?",
      "votes": 4
    },
    {
      "id": 1300952,
      "postDate": "2021-05-10T21:23:38.733Z",
      "content": "<p>Well I get 0.000 every single time, Can you share your baseline with me? tnx</p>",
      "rawMarkdown": "Well I get 0.000 every single time, Can you share your baseline with me? tnx",
      "replies": [
        {
          "id": 1300965,
          "postDate": "2021-05-10T21:37:05.167Z",
          "content": "<p>The host has posted <a href=\"https://www.kaggle.com/anjrash/batchall-baseline-hotelid\" target=\"_blank\">a baseline here</a>.</p>",
          "rawMarkdown": "The host has posted [a baseline here](https://www.kaggle.com/anjrash/batchall-baseline-hotelid)."
        },
        {
          "id": 1300982,
          "postDate": "2021-05-10T21:52:26.430Z",
          "content": "<p>Much appreciated, that will help a lot</p>",
          "rawMarkdown": "Much appreciated, that will help a lot"
        },
        {
          "id": 1301059,
          "postDate": "2021-05-10T23:49:48.917Z",
          "content": "<p>A baseline can be very simple, just train an efficientnet b0 as is (7770 classes), and get the topk scores directly from logits. I used pytorch with timm library:<br>\n<a href=\"url\" target=\"_blank\">https://pypi.org/project/timm/</a><br>\nYou can also train a model to predict the rotation of the images (some images are randomly rotated) to get a better result<br>\n<a href=\"url\" target=\"_blank\">https://www.kaggle.com/alenic/discover-rotation</a><br>\nI just splitted 85% train and 25% validation</p>",
          "rawMarkdown": "A baseline can be very simple, just train an efficientnet b0 as is (7770 classes), and get the topk scores directly from logits. I used pytorch with timm library:\n[https://pypi.org/project/timm/](url)\nYou can also train a model to predict the rotation of the images (some images are randomly rotated) to get a better result\n[https://www.kaggle.com/alenic/discover-rotation](url)\nI just splitted 85% train and 25% validation"
        },
        {
          "id": 1301250,
          "postDate": "2021-05-11T01:56:55.987Z",
          "content": "<p>Appreciate the comments, I will try doing that. Wish me luck =)</p>",
          "rawMarkdown": "Appreciate the comments, I will try doing that. Wish me luck =)"
        }
      ]
    },
    {
      "id": 1300973,
      "postDate": "2021-05-10T21:41:57.827Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1300952,
      "author_name": "Adriano Passos",
      "author_url": "",
      "post_date": "2021-05-10T21:23:38.733000",
      "content": "<p>Well I get 0.000 every single time, Can you share your baseline with me? tnx</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1300965,
          "author_name": "Sohier Dane",
          "author_url": "",
          "post_date": "2021-05-10T21:37:05.167000",
          "content": "<p>The host has posted <a href=\"https://www.kaggle.com/anjrash/batchall-baseline-hotelid\" target=\"_blank\">a baseline here</a>.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1300982,
          "author_name": "Adriano Passos",
          "author_url": "",
          "post_date": "2021-05-10T21:52:26.430000",
          "content": "<p>Much appreciated, that will help a lot</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1301059,
          "author_name": "AleNic",
          "author_url": "",
          "post_date": "2021-05-10T23:49:48.917000",
          "content": "<p>A baseline can be very simple, just train an efficientnet b0 as is (7770 classes), and get the topk scores directly from logits. I used pytorch with timm library:<br>\n<a href=\"url\" target=\"_blank\">https://pypi.org/project/timm/</a><br>\nYou can also train a model to predict the rotation of the images (some images are randomly rotated) to get a better result<br>\n<a href=\"url\" target=\"_blank\">https://www.kaggle.com/alenic/discover-rotation</a><br>\nI just splitted 85% train and 25% validation</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1301250,
          "author_name": "Adriano Passos",
          "author_url": "",
          "post_date": "2021-05-11T01:56:55.987000",
          "content": "<p>Appreciate the comments, I will try doing that. Wish me luck =)</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1300973,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-05-10T21:41:57.827000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1282687": "Hi, maybe it's my problem, bu I noticed a strong difference between my validation score and LB score. As a first solution I tried a very simple classifier and I obtained a similar mAP of public LB. After that I tried more advanced techniques like: arcmargin loss, triplet loss, multi head architectures, kmeans instance grouping, etc. and in my local validation I got much better results than the first simple solution as I expected, but I got worse on LB! \nMaybe there is an error in my notebook...\nHave you notice any strange behaviour?",
    "1300952": "Well I get 0.000 every single time, Can you share your baseline with me? tnx",
    "1300973": ""
  }
}