{
  "id": 190761,
  "title": "Can anybody tell me what's going wrong?",
  "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/190761",
  "author_name": "",
  "post_date": "2020-10-13T08:52:07.784459300Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>During the training of the baseline model with Resnet50 as the backbone for 20000 iterations, the loss on each batch goes like the following picture.<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3722352%2Ffba3d7cd1ace323ab37925d5598c044d%2F10311602379160_.pic.jpg?generation=1602578524325057&amp;alt=media\" alt=\"\"></p>\n<p>But the performance of the baseline model on the test set seems incredibly bad likes higher than 10000 on LB.</p>\n<p>What kind of problem can lead to this case? Can <code>disable_traffic_light_faces</code> causes such a huge huge huge gap between train loss and LB score? </p>",
  "messages": [
    {
      "id": "1048175",
      "postDate": "10/13/2020 08:52:07",
      "content": "<p>During the training of the baseline model with Resnet50 as the backbone for 20000 iterations, the loss on each batch goes like the following picture.<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3722352%2Ffba3d7cd1ace323ab37925d5598c044d%2F10311602379160_.pic.jpg?generation=1602578524325057&amp;alt=media\" alt=\"\"></p>\n<p>But the performance of the baseline model on the test set seems incredibly bad likes higher than 10000 on LB.</p>\n<p>What kind of problem can lead to this case? Can <code>disable_traffic_light_faces</code> causes such a huge huge huge gap between train loss and LB score? </p>",
      "rawMarkdown": "During the training of the baseline model with Resnet50 as the backbone for 20000 iterations, the loss on each batch goes like the following picture.![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3722352%2Ffba3d7cd1ace323ab37925d5598c044d%2F10311602379160_.pic.jpg?generation=1602578524325057&alt=media)\n\nBut the performance of the baseline model on the test set seems incredibly bad likes higher than 10000 on LB.\n\nWhat kind of problem can lead to this case? Can `disable_traffic_light_faces` causes such a huge huge huge gap between train loss and LB score?",
      "votes": null
    },
    {
      "id": "1048316",
      "postDate": "10/13/2020 11:36:39",
      "content": "<p>One common problem is not transforming the coordinates before submitting. <a href=\"https://github.com/lyft/l5kit/blob/master/examples/agent_motion_prediction/agent_motion_prediction.ipynb\" target=\"_blank\">In this example notebook</a> search for this line</p>\n<p><code>coords_offset.append(transform_points(agent_coords, world_from_agent) - centroid[:2])</code></p>",
      "rawMarkdown": "One common problem is not transforming the coordinates before submitting. [In this example notebook](https://github.com/lyft/l5kit/blob/master/examples/agent_motion_prediction/agent_motion_prediction.ipynb) search for this line\n\n`coords_offset.append(transform_points(agent_coords, world_from_agent) - centroid[:2])`",
      "votes": null
    },
    {
      "id": "1048943",
      "postDate": "10/14/2020 01:18:21",
      "content": "<p>Thank you so much! I hadn't realized that the test should be transformed like the validation set in the sample code.</p>",
      "rawMarkdown": "Thank you so much! I hadn't realized that the test should be transformed like the validation set in the sample code.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1048316,
      "author_name": "zaharch",
      "author_url": "",
      "post_date": "10/13/2020 11:36:39",
      "content": "<p>One common problem is not transforming the coordinates before submitting. <a href=\"https://github.com/lyft/l5kit/blob/master/examples/agent_motion_prediction/agent_motion_prediction.ipynb\" target=\"_blank\">In this example notebook</a> search for this line</p>\n<p><code>coords_offset.append(transform_points(agent_coords, world_from_agent) - centroid[:2])</code></p>",
      "votes": null,
      "replies": [
        {
          "id": 1048943,
          "author_name": "spicychicken38",
          "author_url": "",
          "post_date": "10/14/2020 01:18:21",
          "content": "<p>Thank you so much! I hadn't realized that the test should be transformed like the validation set in the sample code.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1048175": "During the training of the baseline model with Resnet50 as the backbone for 20000 iterations, the loss on each batch goes like the following picture.![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3722352%2Ffba3d7cd1ace323ab37925d5598c044d%2F10311602379160_.pic.jpg?generation=1602578524325057&alt=media)\n\nBut the performance of the baseline model on the test set seems incredibly bad likes higher than 10000 on LB.\n\nWhat kind of problem can lead to this case? Can `disable_traffic_light_faces` causes such a huge huge huge gap between train loss and LB score?",
    "1048316": "One common problem is not transforming the coordinates before submitting. [In this example notebook](https://github.com/lyft/l5kit/blob/master/examples/agent_motion_prediction/agent_motion_prediction.ipynb) search for this line\n\n`coords_offset.append(transform_points(agent_coords, world_from_agent) - centroid[:2])`",
    "1048943": "Thank you so much! I hadn't realized that the test should be transformed like the validation set in the sample code."
  },
  "source": "meta"
}