{
  "id": 172442,
  "title": "UTokyo ratio in the Public set?",
  "url": "/competitions/global-wheat-detection/discussion/172442",
  "author_name": "",
  "post_date": "2020-08-05T04:52:19.939456100Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>By LB probing, it was possible to get the ratio of UTokyo in the public set (e.g. by submitting empty submissions for UTokyo and examining the LB drop..) <a href=\"https://www.kaggle.com/c/global-wheat-detection/discussion/172416\">The data source classification could be done by counting of wheat heads.</a></p>\n\n<p>Did anyone try this? I realized this last week but could not do it due to submission limitations.\nWould be thankful if anyone can share this! (would save efforts for late-subs ;)</p>",
  "messages": [
    {
      "id": "958625",
      "postDate": "08/05/2020 04:52:19",
      "content": "<p>By LB probing, it was possible to get the ratio of UTokyo in the public set (e.g. by submitting empty submissions for UTokyo and examining the LB drop..) <a href=\"https://www.kaggle.com/c/global-wheat-detection/discussion/172416\">The data source classification could be done by counting of wheat heads.</a></p>\n\n<p>Did anyone try this? I realized this last week but could not do it due to submission limitations.\nWould be thankful if anyone can share this! (would save efforts for late-subs ;)</p>",
      "rawMarkdown": "By LB probing, it was possible to get the ratio of UTokyo in the public set (e.g. by submitting empty submissions for UTokyo and examining the LB drop..) [The data source classification could be done by counting of wheat heads.](https://www.kaggle.com/c/global-wheat-detection/discussion/172416)\n\nDid anyone try this? I realized this last week but could not do it due to submission limitations.\nWould be thankful if anyone can share this! (would save efforts for late-subs ;)",
      "votes": null
    },
    {
      "id": "958627",
      "postDate": "08/05/2020 04:54:28",
      "content": "<p>I tried some metric learning methods that attempted zero-shot classification (aka train a metric-learning model on training sources, then do kmeans of feature vectors on the test set), but I didn't get any significant boost in LB performance. </p>",
      "rawMarkdown": "I tried some metric learning methods that attempted zero-shot classification (aka train a metric-learning model on training sources, then do kmeans of feature vectors on the test set), but I didn't get any significant boost in LB performance.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 958627,
      "author_name": "alexanderliao",
      "author_url": "",
      "post_date": "08/05/2020 04:54:28",
      "content": "<p>I tried some metric learning methods that attempted zero-shot classification (aka train a metric-learning model on training sources, then do kmeans of feature vectors on the test set), but I didn't get any significant boost in LB performance. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "958625": "By LB probing, it was possible to get the ratio of UTokyo in the public set (e.g. by submitting empty submissions for UTokyo and examining the LB drop..) [The data source classification could be done by counting of wheat heads.](https://www.kaggle.com/c/global-wheat-detection/discussion/172416)\n\nDid anyone try this? I realized this last week but could not do it due to submission limitations.\nWould be thankful if anyone can share this! (would save efforts for late-subs ;)",
    "958627": "I tried some metric learning methods that attempted zero-shot classification (aka train a metric-learning model on training sources, then do kmeans of feature vectors on the test set), but I didn't get any significant boost in LB performance."
  },
  "source": "meta"
}