{
  "id": 175939,
  "title": "Sharing my 0.6977 CenterNet training + inference notebooks",
  "url": "/competitions/global-wheat-detection/discussion/175939",
  "author_name": "",
  "post_date": "2020-08-19T22:44:32.669475200Z",
  "votes": 20,
  "comment_count": 5,
  "views": 0,
  "content": "<p><a href=\"https://www.kaggle.com/jihangz/wheat-effnet-b5-bifpn-centernet-train?scriptVersionId=41045120\" target=\"_blank\">Training</a> / <a href=\"https://www.kaggle.com/jihangz/wheat-centernet-infer-bifpn-nfolds-tta-v2?scriptVersionId=40044689\" target=\"_blank\">Inference (0.6964 private)</a> / <a href=\"https://www.kaggle.com/jihangz/wheat-centernet-pseudo-l-bifpn?scriptVersionId=40054948\" target=\"_blank\">Pseudo-label (0.6977 private)</a></p>\n<p>The only differences between the 0.6964 notebook and my final submission are:</p>\n<ol>\n<li>3-fold vs 5-fold (not much difference)</li>\n<li>inference on 1.1x + 1.25x vs 1.1x + 1.25x + 1.35x scales (huge drop by adding 1.35x)</li>\n</ol>\n<p>Lesson learned: <strong>do not inference on 1.35x</strong> 😂</p>\n<p>Solution details are <a href=\"https://www.kaggle.com/c/global-wheat-detection/discussion/172485\" target=\"_blank\">here</a></p>",
  "messages": [
    {
      "id": "978061",
      "postDate": "08/19/2020 22:44:32",
      "content": "<p><a href=\"https://www.kaggle.com/jihangz/wheat-effnet-b5-bifpn-centernet-train?scriptVersionId=41045120\" target=\"_blank\">Training</a> / <a href=\"https://www.kaggle.com/jihangz/wheat-centernet-infer-bifpn-nfolds-tta-v2?scriptVersionId=40044689\" target=\"_blank\">Inference (0.6964 private)</a> / <a href=\"https://www.kaggle.com/jihangz/wheat-centernet-pseudo-l-bifpn?scriptVersionId=40054948\" target=\"_blank\">Pseudo-label (0.6977 private)</a></p>\n<p>The only differences between the 0.6964 notebook and my final submission are:</p>\n<ol>\n<li>3-fold vs 5-fold (not much difference)</li>\n<li>inference on 1.1x + 1.25x vs 1.1x + 1.25x + 1.35x scales (huge drop by adding 1.35x)</li>\n</ol>\n<p>Lesson learned: <strong>do not inference on 1.35x</strong> 😂</p>\n<p>Solution details are <a href=\"https://www.kaggle.com/c/global-wheat-detection/discussion/172485\" target=\"_blank\">here</a></p>",
      "rawMarkdown": "[Training](https://www.kaggle.com/jihangz/wheat-effnet-b5-bifpn-centernet-train?scriptVersionId=41045120) / [Inference (0.6964 private)](https://www.kaggle.com/jihangz/wheat-centernet-infer-bifpn-nfolds-tta-v2?scriptVersionId=40044689) / [Pseudo-label (0.6977 private)](https://www.kaggle.com/jihangz/wheat-centernet-pseudo-l-bifpn?scriptVersionId=40054948)\n\nThe only differences between the 0.6964 notebook and my final submission are:\n1. 3-fold vs 5-fold (not much difference)\n2. inference on 1.1x + 1.25x vs 1.1x + 1.25x + 1.35x scales (huge drop by adding 1.35x)\n\nLesson learned: **do not inference on 1.35x** 😂\n\n\nSolution details are [here](https://www.kaggle.com/c/global-wheat-detection/discussion/172485)",
      "votes": null
    },
    {
      "id": "978161",
      "postDate": "08/20/2020 01:54:05",
      "content": "<p>Feeding a simple efficientdet to a yolov3 pseudo labeling was also in .68+ on private lb. I unfortunately didn't select it…</p>",
      "rawMarkdown": "Feeding a simple efficientdet to a yolov3 pseudo labeling was also in .68+ on private lb. I unfortunately didn't select it...",
      "votes": null
    },
    {
      "id": "978175",
      "postDate": "08/20/2020 02:22:16",
      "content": "<p>It's a lottery haha</p>",
      "rawMarkdown": "It's a lottery haha",
      "votes": null
    },
    {
      "id": "978284",
      "postDate": "08/20/2020 04:43:11",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/jihangz\" target=\"_blank\">@jihangz</a> and Thanks for sharing.</p>\n<p>Could you explain more this? :<br>\n<code>inference on 1.1x + 1.25x vs 1.1x + 1.25x + 1.35x (huge drop by adding 1.35x)</code></p>\n<p>(It seems that test_scales but I am not familiar with this part.)</p>",
      "rawMarkdown": "Congrats @jihangz and Thanks for sharing.\n\nCould you explain more this? :\n`inference on 1.1x + 1.25x vs 1.1x + 1.25x + 1.35x (huge drop by adding 1.35x)`\n\n(It seems that test_scales but I am not familiar with this part.)",
      "votes": null
    },
    {
      "id": "978289",
      "postDate": "08/20/2020 04:52:38",
      "content": "<p>Yeah testing scales. Meaning inferring on 1.1/1.25/1.35x original resolution then ensemble using WBF</p>",
      "rawMarkdown": "Yeah testing scales. Meaning inferring on 1.1/1.25/1.35x original resolution then ensemble using WBF",
      "votes": null
    },
    {
      "id": "978668",
      "postDate": "08/20/2020 10:07:05",
      "content": "<p>Thank you for sharing.</p>",
      "rawMarkdown": "Thank you for sharing.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 978161,
      "author_name": "chabir",
      "author_url": "",
      "post_date": "08/20/2020 01:54:05",
      "content": "<p>Feeding a simple efficientdet to a yolov3 pseudo labeling was also in .68+ on private lb. I unfortunately didn't select it…</p>",
      "votes": null,
      "replies": [
        {
          "id": 978175,
          "author_name": "jihangz",
          "author_url": "",
          "post_date": "08/20/2020 02:22:16",
          "content": "<p>It's a lottery haha</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 978284,
      "author_name": "piantic",
      "author_url": "",
      "post_date": "08/20/2020 04:43:11",
      "content": "<p>Congrats <a href=\"https://www.kaggle.com/jihangz\" target=\"_blank\">@jihangz</a> and Thanks for sharing.</p>\n<p>Could you explain more this? :<br>\n<code>inference on 1.1x + 1.25x vs 1.1x + 1.25x + 1.35x (huge drop by adding 1.35x)</code></p>\n<p>(It seems that test_scales but I am not familiar with this part.)</p>",
      "votes": null,
      "replies": [
        {
          "id": 978289,
          "author_name": "jihangz",
          "author_url": "",
          "post_date": "08/20/2020 04:52:38",
          "content": "<p>Yeah testing scales. Meaning inferring on 1.1/1.25/1.35x original resolution then ensemble using WBF</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 978668,
      "author_name": "aliabdin1",
      "author_url": "",
      "post_date": "08/20/2020 10:07:05",
      "content": "<p>Thank you for sharing.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "978061": "[Training](https://www.kaggle.com/jihangz/wheat-effnet-b5-bifpn-centernet-train?scriptVersionId=41045120) / [Inference (0.6964 private)](https://www.kaggle.com/jihangz/wheat-centernet-infer-bifpn-nfolds-tta-v2?scriptVersionId=40044689) / [Pseudo-label (0.6977 private)](https://www.kaggle.com/jihangz/wheat-centernet-pseudo-l-bifpn?scriptVersionId=40054948)\n\nThe only differences between the 0.6964 notebook and my final submission are:\n1. 3-fold vs 5-fold (not much difference)\n2. inference on 1.1x + 1.25x vs 1.1x + 1.25x + 1.35x scales (huge drop by adding 1.35x)\n\nLesson learned: **do not inference on 1.35x** 😂\n\n\nSolution details are [here](https://www.kaggle.com/c/global-wheat-detection/discussion/172485)",
    "978161": "Feeding a simple efficientdet to a yolov3 pseudo labeling was also in .68+ on private lb. I unfortunately didn't select it...",
    "978175": "It's a lottery haha",
    "978284": "Congrats @jihangz and Thanks for sharing.\n\nCould you explain more this? :\n`inference on 1.1x + 1.25x vs 1.1x + 1.25x + 1.35x (huge drop by adding 1.35x)`\n\n(It seems that test_scales but I am not familiar with this part.)",
    "978289": "Yeah testing scales. Meaning inferring on 1.1/1.25/1.35x original resolution then ensemble using WBF",
    "978668": "Thank you for sharing."
  },
  "source": "meta"
}