{
  "id": 178245,
  "title": "Experience: Deep of Backbone",
  "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/178245",
  "author_name": "",
  "post_date": "2020-08-29T08:11:18.054437600Z",
  "votes": 17,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I tried some model and have that experience:</p>\n<ul>\n<li>with EfficientNet worse than the result of Resnet18</li>\n<li>with Resnet50 worse than the result of Resnet18<br>\nSeem like in this competition, deep model is not the key<br>\nNice to see another experience</li>\n</ul>",
  "messages": [
    {
      "id": "989923",
      "postDate": "08/29/2020 08:11:18",
      "content": "<p>I tried some model and have that experience:</p>\n<ul>\n<li>with EfficientNet worse than the result of Resnet18</li>\n<li>with Resnet50 worse than the result of Resnet18<br>\nSeem like in this competition, deep model is not the key<br>\nNice to see another experience</li>\n</ul>",
      "rawMarkdown": "I tried some model and have that experience:\n- with EfficientNet worse than the result of Resnet18\n- with Resnet50 worse than the result of Resnet18\nSeem like in this competition, deep model is not the key\nNice to see another experience",
      "votes": null
    },
    {
      "id": "990534",
      "postDate": "08/29/2020 17:16:50",
      "content": "<p>Same here…</p>\n<p>Very poor results with Efficientnet. </p>\n<p>May be you need to tune carefully raster/pixels sizes</p>",
      "rawMarkdown": "Same here...\n\nVery poor results with Efficientnet. \n\n\n\nMay be you need to tune carefully raster/pixels sizes",
      "votes": null
    },
    {
      "id": "990562",
      "postDate": "08/29/2020 17:37:29",
      "content": "<p>History_frame is very good data. Use this can improve LB</p>",
      "rawMarkdown": "History_frame is very good data. Use this can improve LB",
      "votes": null
    },
    {
      "id": "991669",
      "postDate": "08/30/2020 15:14:14",
      "content": "<p>Same here! I got worst results on Efficient Net.</p>",
      "rawMarkdown": "Same here! I got worst results on Efficient Net.",
      "votes": null
    },
    {
      "id": "1004131",
      "postDate": "09/09/2020 13:39:44",
      "content": "<p><a href=\"https://www.kaggle.com/doanquanvietnamca\" target=\"_blank\">@doanquanvietnamca</a> Even I have arrived at same conclusion. Res50 with 100k Iteration is worse than Res18 with 25k Iteration for Single Trajectory prediction. Its not just about the  loss but performance wise as well.  </p>",
      "rawMarkdown": "doanquanvietnamca Even I have arrived at same conclusion. Res50 with 100k Iteration is worse than Res18 with 25k Iteration for Single Trajectory prediction. Its not just about the  loss but performance wise as well.",
      "votes": null
    },
    {
      "id": "1074705",
      "postDate": "11/11/2020 00:44:43",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/doanquanvietnamca\" target=\"_blank\">@doanquanvietnamca</a> , by history_frame, are you referring to history frames in data['image'] variable or data['history_position'] ?</p>",
      "rawMarkdown": "Hi @doanquanvietnamca , by history_frame, are you referring to history frames in data['image'] variable or data['history_position'] ?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 990534,
      "author_name": "serigne",
      "author_url": "",
      "post_date": "08/29/2020 17:16:50",
      "content": "<p>Same here…</p>\n<p>Very poor results with Efficientnet. </p>\n<p>May be you need to tune carefully raster/pixels sizes</p>",
      "votes": null,
      "replies": [
        {
          "id": 990562,
          "author_name": "doanquanvietnamca",
          "author_url": "",
          "post_date": "08/29/2020 17:37:29",
          "content": "<p>History_frame is very good data. Use this can improve LB</p>",
          "votes": null,
          "replies": [
            {
              "id": 1074705,
              "author_name": "suryajrrafl",
              "author_url": "",
              "post_date": "11/11/2020 00:44:43",
              "content": "<p>Hi <a href=\"https://www.kaggle.com/doanquanvietnamca\" target=\"_blank\">@doanquanvietnamca</a> , by history_frame, are you referring to history frames in data['image'] variable or data['history_position'] ?</p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 991669,
      "author_name": "kaushal2896",
      "author_url": "",
      "post_date": "08/30/2020 15:14:14",
      "content": "<p>Same here! I got worst results on Efficient Net.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1004131,
      "author_name": "deepakrajpurushothaman",
      "author_url": "",
      "post_date": "09/09/2020 13:39:44",
      "content": "<p><a href=\"https://www.kaggle.com/doanquanvietnamca\" target=\"_blank\">@doanquanvietnamca</a> Even I have arrived at same conclusion. Res50 with 100k Iteration is worse than Res18 with 25k Iteration for Single Trajectory prediction. Its not just about the  loss but performance wise as well.  </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "989923": "I tried some model and have that experience:\n- with EfficientNet worse than the result of Resnet18\n- with Resnet50 worse than the result of Resnet18\nSeem like in this competition, deep model is not the key\nNice to see another experience",
    "990534": "Same here...\n\nVery poor results with Efficientnet. \n\n\n\nMay be you need to tune carefully raster/pixels sizes",
    "990562": "History_frame is very good data. Use this can improve LB",
    "991669": "Same here! I got worst results on Efficient Net.",
    "1004131": "doanquanvietnamca Even I have arrived at same conclusion. Res50 with 100k Iteration is worse than Res18 with 25k Iteration for Single Trajectory prediction. Its not just about the  loss but performance wise as well.",
    "1074705": "Hi @doanquanvietnamca , by history_frame, are you referring to history frames in data['image'] variable or data['history_position'] ?"
  },
  "source": "meta"
}