{
  "id": 188987,
  "title": "recent submissions are very bad!",
  "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/188987",
  "author_name": "",
  "post_date": "2020-10-06T06:46:44.301737600Z",
  "votes": null,
  "comment_count": 5,
  "views": 0,
  "content": "<p>my recent submission scores are above <code>10, 000</code>, it should be around <code>100</code> as prior submission because i did't change my code a lot and <code>the evaluation score (less than 100) is ok</code>. Did I miss something important?</p>\n<p>I know that <code>l5kit</code> has been updated to version <code>1.1.0</code>, the coord system has changed.  I've try to transfer the agent's coord to real world during inference, <code>but it didn't work</code>.  </p>\n<p>Do you guys had similar experenice? How did you solved this problem? Need you guys help!</p>",
  "messages": [
    {
      "id": "1038909",
      "postDate": "10/06/2020 06:46:44",
      "content": "<p>my recent submission scores are above <code>10, 000</code>, it should be around <code>100</code> as prior submission because i did't change my code a lot and <code>the evaluation score (less than 100) is ok</code>. Did I miss something important?</p>\n<p>I know that <code>l5kit</code> has been updated to version <code>1.1.0</code>, the coord system has changed.  I've try to transfer the agent's coord to real world during inference, <code>but it didn't work</code>.  </p>\n<p>Do you guys had similar experenice? How did you solved this problem? Need you guys help!</p>",
      "rawMarkdown": "my recent submission scores are above `10, 000`, it should be around `100` as prior submission because i did't change my code a lot and `the evaluation score (less than 100) is ok`. Did I miss something important?\n\nI know that `l5kit` has been updated to version `1.1.0`, the coord system has changed.  I've try to transfer the agent's coord to real world during inference, `but it didn't work`.  \n\nDo you guys had similar experenice? How did you solved this problem? Need you guys help!",
      "votes": null
    },
    {
      "id": "1039061",
      "postDate": "10/06/2020 09:40:29",
      "content": "<p>Hey <a href=\"https://www.kaggle.com/wantsu\" target=\"_blank\">@wantsu</a> did you take a look at <a href=\"https://www.kaggle.com/lucabergamini/lyft-baseline-09-02\" target=\"_blank\">my example kernel</a>? I got a nice improvement after the coord system fix. Can you post here the code you're using to transform coordinates between spaces?</p>",
      "rawMarkdown": "Hey @wantsu did you take a look at [my example kernel](https://www.kaggle.com/lucabergamini/lyft-baseline-09-02)? I got a nice improvement after the coord system fix. Can you post here the code you're using to transform coordinates between spaces?",
      "votes": null
    },
    {
      "id": "1039097",
      "postDate": "10/06/2020 10:24:43",
      "content": "<p>Hi,Luca. yes, I 've seen your notebook before, version7, and modify my notebook base on it.<br>\nThis is my code. The output is multimodal.</p>\n<pre><code>        world_from_agents = data[\"world_from_agent\"].numpy()\n        centroids = data['centroid'].numpy()\n        # convert into world coordinates and compute offsets\n        for idx in range(len(outputs)):\n            outputs[idx][:,0] = transform_points(outputs[idx][:,0], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][:,1] = transform_points(outputs[idx][:,1], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][:,2] = transform_points(outputs[idx][:,2], world_from_agents[idx]) - centroids[idx]\n\n        future_coords_offsets_pd.append(outputs.copy())\n        timestamps.append(data[\"timestamp\"].numpy().copy())\n        agent_ids.append(data[\"track_id\"].numpy().copy())\n        confs.append(confidences.cpu().numpy().copy())\n</code></pre>",
      "rawMarkdown": "Hi,Luca. yes, I 've seen your notebook before, version7, and modify my notebook base on it.\nThis is my code. The output is multimodal.\n```python\n        world_from_agents = data[\"world_from_agent\"].numpy()\n        centroids = data['centroid'].numpy()\n        # convert into world coordinates and compute offsets\n        for idx in range(len(outputs)):\n            outputs[idx][:,0] = transform_points(outputs[idx][:,0], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][:,1] = transform_points(outputs[idx][:,1], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][:,2] = transform_points(outputs[idx][:,2], world_from_agents[idx]) - centroids[idx]\n            \n        future_coords_offsets_pd.append(outputs.copy())\n        timestamps.append(data[\"timestamp\"].numpy().copy())\n        agent_ids.append(data[\"track_id\"].numpy().copy())\n        confs.append(confidences.cpu().numpy().copy())\n```",
      "votes": null
    },
    {
      "id": "1039137",
      "postDate": "10/06/2020 11:05:48",
      "content": "<p>I just realized that the slice operation is incorrect. Thanks for your comment.👍</p>",
      "rawMarkdown": "I just realized that the slice operation is incorrect. Thanks for your comment.👍",
      "votes": null
    },
    {
      "id": "1039138",
      "postDate": "10/06/2020 11:08:27",
      "content": "<p><a href=\"https://www.kaggle.com/wantsu\" target=\"_blank\">@wantsu</a> . Can you post the corrected code ?<br>\nI think I am also having slicing problems doing that</p>",
      "rawMarkdown": "wantsu . Can you post the corrected code ?\nI think I am also having slicing problems doing that",
      "votes": null
    },
    {
      "id": "1039148",
      "postDate": "10/06/2020 11:20:01",
      "content": "<p>shape of output is (batch, mode, length, 2)</p>\n<pre><code>for idx in range(len(outputs)):\n            outputs[idx][0] = transform_points(outputs[idx][0], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][1] = transform_points(outputs[idx][1], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][2] = transform_points(outputs[idx][2], world_from_agents[idx]) - centroids[idx]\n</code></pre>\n<p>or</p>\n<pre><code>for idx in range(len(outputs)):\n            outputs[:, 0][idx] = transform_points(outputs[:, 0][idx], world_from_agents[idx]) - centroids[idx]\n            outputs[:, 1][idx] = transform_points(outputs[:, 1][idx], world_from_agents[idx]) - centroids[idx]\n            outputs[:, 2][idx] = transform_points(outputs[:, 2][idx], world_from_agents[idx]) - centroids[idx]\n</code></pre>",
      "rawMarkdown": "shape of output is (batch, mode, length, 2)\n```python\nfor idx in range(len(outputs)):\n            outputs[idx][0] = transform_points(outputs[idx][0], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][1] = transform_points(outputs[idx][1], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][2] = transform_points(outputs[idx][2], world_from_agents[idx]) - centroids[idx]\n```\nor\n```python\nfor idx in range(len(outputs)):\n            outputs[:, 0][idx] = transform_points(outputs[:, 0][idx], world_from_agents[idx]) - centroids[idx]\n            outputs[:, 1][idx] = transform_points(outputs[:, 1][idx], world_from_agents[idx]) - centroids[idx]\n            outputs[:, 2][idx] = transform_points(outputs[:, 2][idx], world_from_agents[idx]) - centroids[idx]\n```",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1039061,
      "author_name": "lucabergamini",
      "author_url": "",
      "post_date": "10/06/2020 09:40:29",
      "content": "<p>Hey <a href=\"https://www.kaggle.com/wantsu\" target=\"_blank\">@wantsu</a> did you take a look at <a href=\"https://www.kaggle.com/lucabergamini/lyft-baseline-09-02\" target=\"_blank\">my example kernel</a>? I got a nice improvement after the coord system fix. Can you post here the code you're using to transform coordinates between spaces?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1039097,
          "author_name": "wantsu",
          "author_url": "",
          "post_date": "10/06/2020 10:24:43",
          "content": "<p>Hi,Luca. yes, I 've seen your notebook before, version7, and modify my notebook base on it.<br>\nThis is my code. The output is multimodal.</p>\n<pre><code>        world_from_agents = data[\"world_from_agent\"].numpy()\n        centroids = data['centroid'].numpy()\n        # convert into world coordinates and compute offsets\n        for idx in range(len(outputs)):\n            outputs[idx][:,0] = transform_points(outputs[idx][:,0], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][:,1] = transform_points(outputs[idx][:,1], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][:,2] = transform_points(outputs[idx][:,2], world_from_agents[idx]) - centroids[idx]\n\n        future_coords_offsets_pd.append(outputs.copy())\n        timestamps.append(data[\"timestamp\"].numpy().copy())\n        agent_ids.append(data[\"track_id\"].numpy().copy())\n        confs.append(confidences.cpu().numpy().copy())\n</code></pre>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1039137,
          "author_name": "wantsu",
          "author_url": "",
          "post_date": "10/06/2020 11:05:48",
          "content": "<p>I just realized that the slice operation is incorrect. Thanks for your comment.👍</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1039138,
          "author_name": "vladvdv",
          "author_url": "",
          "post_date": "10/06/2020 11:08:27",
          "content": "<p><a href=\"https://www.kaggle.com/wantsu\" target=\"_blank\">@wantsu</a> . Can you post the corrected code ?<br>\nI think I am also having slicing problems doing that</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1039148,
          "author_name": "wantsu",
          "author_url": "",
          "post_date": "10/06/2020 11:20:01",
          "content": "<p>shape of output is (batch, mode, length, 2)</p>\n<pre><code>for idx in range(len(outputs)):\n            outputs[idx][0] = transform_points(outputs[idx][0], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][1] = transform_points(outputs[idx][1], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][2] = transform_points(outputs[idx][2], world_from_agents[idx]) - centroids[idx]\n</code></pre>\n<p>or</p>\n<pre><code>for idx in range(len(outputs)):\n            outputs[:, 0][idx] = transform_points(outputs[:, 0][idx], world_from_agents[idx]) - centroids[idx]\n            outputs[:, 1][idx] = transform_points(outputs[:, 1][idx], world_from_agents[idx]) - centroids[idx]\n            outputs[:, 2][idx] = transform_points(outputs[:, 2][idx], world_from_agents[idx]) - centroids[idx]\n</code></pre>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1038909": "my recent submission scores are above `10, 000`, it should be around `100` as prior submission because i did't change my code a lot and `the evaluation score (less than 100) is ok`. Did I miss something important?\n\nI know that `l5kit` has been updated to version `1.1.0`, the coord system has changed.  I've try to transfer the agent's coord to real world during inference, `but it didn't work`.  \n\nDo you guys had similar experenice? How did you solved this problem? Need you guys help!",
    "1039061": "Hey @wantsu did you take a look at [my example kernel](https://www.kaggle.com/lucabergamini/lyft-baseline-09-02)? I got a nice improvement after the coord system fix. Can you post here the code you're using to transform coordinates between spaces?",
    "1039097": "Hi,Luca. yes, I 've seen your notebook before, version7, and modify my notebook base on it.\nThis is my code. The output is multimodal.\n```python\n        world_from_agents = data[\"world_from_agent\"].numpy()\n        centroids = data['centroid'].numpy()\n        # convert into world coordinates and compute offsets\n        for idx in range(len(outputs)):\n            outputs[idx][:,0] = transform_points(outputs[idx][:,0], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][:,1] = transform_points(outputs[idx][:,1], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][:,2] = transform_points(outputs[idx][:,2], world_from_agents[idx]) - centroids[idx]\n            \n        future_coords_offsets_pd.append(outputs.copy())\n        timestamps.append(data[\"timestamp\"].numpy().copy())\n        agent_ids.append(data[\"track_id\"].numpy().copy())\n        confs.append(confidences.cpu().numpy().copy())\n```",
    "1039137": "I just realized that the slice operation is incorrect. Thanks for your comment.👍",
    "1039138": "wantsu . Can you post the corrected code ?\nI think I am also having slicing problems doing that",
    "1039148": "shape of output is (batch, mode, length, 2)\n```python\nfor idx in range(len(outputs)):\n            outputs[idx][0] = transform_points(outputs[idx][0], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][1] = transform_points(outputs[idx][1], world_from_agents[idx]) - centroids[idx]\n            outputs[idx][2] = transform_points(outputs[idx][2], world_from_agents[idx]) - centroids[idx]\n```\nor\n```python\nfor idx in range(len(outputs)):\n            outputs[:, 0][idx] = transform_points(outputs[:, 0][idx], world_from_agents[idx]) - centroids[idx]\n            outputs[:, 1][idx] = transform_points(outputs[:, 1][idx], world_from_agents[idx]) - centroids[idx]\n            outputs[:, 2][idx] = transform_points(outputs[:, 2][idx], world_from_agents[idx]) - centroids[idx]\n```"
  },
  "source": "meta"
}