{
  "id": 196454,
  "title": "how to ensemble multi trajectory predictions for the final submission?",
  "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/196454",
  "author_name": "",
  "post_date": "2020-11-11T10:34:02.126243100Z",
  "votes": 4,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hello everyone, how can we ensemble multi trajectory predictions for the final submission?<br>\n For x,y trajectory, we can average the value from different models, but I am not sure how to re-calculate the confidence vector. Any help? Thank you!</p>",
  "messages": [
    {
      "id": "1075031",
      "postDate": "11/11/2020 10:34:02",
      "content": "<p>Hello everyone, how can we ensemble multi trajectory predictions for the final submission?<br>\n For x,y trajectory, we can average the value from different models, but I am not sure how to re-calculate the confidence vector. Any help? Thank you!</p>",
      "rawMarkdown": "Hello everyone, how can we ensemble multi trajectory predictions for the final submission?\n For x,y trajectory, we can average the value from different models, but I am not sure how to re-calculate the confidence vector. Any help? Thank you!",
      "votes": null
    },
    {
      "id": "1075669",
      "postDate": "11/11/2020 21:40:13",
      "content": "<p>it's hard question. i amstuck in this too. because you cannot average it. 1 model use 3 mode, difference trajectory in each mode too. So you can try to stack it and add so layer for this</p>",
      "rawMarkdown": "it's hard question. i amstuck in this too. because you cannot average it. 1 model use 3 mode, difference trajectory in each mode too. So you can try to stack it and add so layer for this",
      "votes": null
    },
    {
      "id": "1075798",
      "postDate": "11/12/2020 01:38:11",
      "content": "<p>Thank you for your reply. But when we stacking the model, how could we calculate the confidence vector?</p>",
      "rawMarkdown": "Thank you for your reply. But when we stacking the model, how could we calculate the confidence vector?",
      "votes": null
    },
    {
      "id": "1076886",
      "postDate": "11/13/2020 02:35:29",
      "content": "<p>I had a little experiments.</p>\n<ul>\n<li><p>pick top-3 trajectories by the orders of confidences. and set confidence manually. (eg. conf_0 = 0.5, conf_1 = 0.3, conf_2 = 0.2)<br>\nFailed</p></li>\n<li><p>pick the 1 trajectory of the highest confidences. evaluate in single mode. (conf_0 = 1.0, conf_2 = 0.0, conf_3 = 0.0)<br>\nFailed</p></li>\n</ul>\n<p>Now I am trying to average each trajectories and confidences.<br>\nconf_0 = average of conf_0's<br>\ncoord_1 = average of coord_1's</p>\n<p>A notebook is running. about 3 hours after, I can figure out.<br>\n<strong>Update: Failed</strong></p>\n<p>I have to try stacking model.</p>",
      "rawMarkdown": "I had a little experiments.\n\n- pick top-3 trajectories by the orders of confidences. and set confidence manually. (eg. conf_0 = 0.5, conf_1 = 0.3, conf_2 = 0.2)\nFailed\n\n- pick the 1 trajectory of the highest confidences. evaluate in single mode. (conf_0 = 1.0, conf_2 = 0.0, conf_3 = 0.0)\nFailed\n\nNow I am trying to average each trajectories and confidences.\nconf_0 = average of conf_0's\ncoord_1 = average of coord_1's\n\nA notebook is running. about 3 hours after, I can figure out.\n**Update: Failed**\n\nI have to try stacking model.",
      "votes": null
    },
    {
      "id": "1076892",
      "postDate": "11/13/2020 02:42:22",
      "content": "<p>how about training a stacking model to do the job?</p>",
      "rawMarkdown": "how about training a stacking model to do the job?",
      "votes": null
    },
    {
      "id": "1077094",
      "postDate": "11/13/2020 08:30:51",
      "content": "<p>For me, average works for two same structure models(training data different). 20.4+ with 25+  to 19.5x,   LB<br>\nI tried some public combinings, all failed. <br>\nhow about you, any progress?</p>",
      "rawMarkdown": "For me, average works for two same structure models(training data different). 20.4+ with 25+  to 19.5x,   LB\nI tried some public combinings, all failed. \nhow about you, any progress?",
      "votes": null
    },
    {
      "id": "1077133",
      "postDate": "11/13/2020 09:07:29",
      "content": "<p>Could you explain a little more? <a href=\"https://www.kaggle.com/doanquanvietnamca\" target=\"_blank\">@doanquanvietnamca</a> </p>",
      "rawMarkdown": "Could you explain a little more? @doanquanvietnamca",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1075669,
      "author_name": "doanquanvietnamca",
      "author_url": "",
      "post_date": "11/11/2020 21:40:13",
      "content": "<p>it's hard question. i amstuck in this too. because you cannot average it. 1 model use 3 mode, difference trajectory in each mode too. So you can try to stack it and add so layer for this</p>",
      "votes": null,
      "replies": [
        {
          "id": 1075798,
          "author_name": "shuozhang",
          "author_url": "",
          "post_date": "11/12/2020 01:38:11",
          "content": "<p>Thank you for your reply. But when we stacking the model, how could we calculate the confidence vector?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1077133,
          "author_name": "piantic",
          "author_url": "",
          "post_date": "11/13/2020 09:07:29",
          "content": "<p>Could you explain a little more? <a href=\"https://www.kaggle.com/doanquanvietnamca\" target=\"_blank\">@doanquanvietnamca</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1076886,
      "author_name": "sunghyunjun",
      "author_url": "",
      "post_date": "11/13/2020 02:35:29",
      "content": "<p>I had a little experiments.</p>\n<ul>\n<li><p>pick top-3 trajectories by the orders of confidences. and set confidence manually. (eg. conf_0 = 0.5, conf_1 = 0.3, conf_2 = 0.2)<br>\nFailed</p></li>\n<li><p>pick the 1 trajectory of the highest confidences. evaluate in single mode. (conf_0 = 1.0, conf_2 = 0.0, conf_3 = 0.0)<br>\nFailed</p></li>\n</ul>\n<p>Now I am trying to average each trajectories and confidences.<br>\nconf_0 = average of conf_0's<br>\ncoord_1 = average of coord_1's</p>\n<p>A notebook is running. about 3 hours after, I can figure out.<br>\n<strong>Update: Failed</strong></p>\n<p>I have to try stacking model.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1076892,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "11/13/2020 02:42:22",
      "content": "<p>how about training a stacking model to do the job?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1077094,
      "author_name": "mcggood",
      "author_url": "",
      "post_date": "11/13/2020 08:30:51",
      "content": "<p>For me, average works for two same structure models(training data different). 20.4+ with 25+  to 19.5x,   LB<br>\nI tried some public combinings, all failed. <br>\nhow about you, any progress?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1075031": "Hello everyone, how can we ensemble multi trajectory predictions for the final submission?\n For x,y trajectory, we can average the value from different models, but I am not sure how to re-calculate the confidence vector. Any help? Thank you!",
    "1075669": "it's hard question. i amstuck in this too. because you cannot average it. 1 model use 3 mode, difference trajectory in each mode too. So you can try to stack it and add so layer for this",
    "1075798": "Thank you for your reply. But when we stacking the model, how could we calculate the confidence vector?",
    "1076886": "I had a little experiments.\n\n- pick top-3 trajectories by the orders of confidences. and set confidence manually. (eg. conf_0 = 0.5, conf_1 = 0.3, conf_2 = 0.2)\nFailed\n\n- pick the 1 trajectory of the highest confidences. evaluate in single mode. (conf_0 = 1.0, conf_2 = 0.0, conf_3 = 0.0)\nFailed\n\nNow I am trying to average each trajectories and confidences.\nconf_0 = average of conf_0's\ncoord_1 = average of coord_1's\n\nA notebook is running. about 3 hours after, I can figure out.\n**Update: Failed**\n\nI have to try stacking model.",
    "1076892": "how about training a stacking model to do the job?",
    "1077094": "For me, average works for two same structure models(training data different). 20.4+ with 25+  to 19.5x,   LB\nI tried some public combinings, all failed. \nhow about you, any progress?",
    "1077133": "Could you explain a little more? @doanquanvietnamca"
  },
  "source": "meta"
}