{
  "id": 120999,
  "title": "Blending",
  "url": "/competitions/pku-autonomous-driving/discussion/120999",
  "author_name": "",
  "post_date": "2019-12-10T11:54:15.209464500Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Does anyone know any blending or stacking techniques possible in this competition?</p>",
  "messages": [
    {
      "id": "691679",
      "postDate": "12/10/2019 11:54:15",
      "content": "<p>Does anyone know any blending or stacking techniques possible in this competition?</p>",
      "rawMarkdown": "Does anyone know any blending or stacking techniques possible in this competition?",
      "votes": null
    },
    {
      "id": "691789",
      "postDate": "12/10/2019 13:54:37",
      "content": "<p>I'm not super familiar with blending but I would assume you could just average the predictions of multiple models together.</p>",
      "rawMarkdown": "I'm not super familiar with blending but I would assume you could just average the predictions of multiple models together.",
      "votes": null
    },
    {
      "id": "691894",
      "postDate": "12/10/2019 16:07:10",
      "content": "<p>If your models use same image size or preprocess, you just need to take average of your predictions before any post process. </p>",
      "rawMarkdown": "If your models use same image size or preprocess, you just need to take average of your predictions before any post process.",
      "votes": null
    },
    {
      "id": "692397",
      "postDate": "12/11/2019 08:06:54",
      "content": "<p>Sounds easy, but I believe there can be many minor problems. For example, if two models produce different FP / FN predictions, how would you average these FP / FN? What if one model is better for a short-range detection, while the other one produces best estimations for cars, which are far away from camera?</p>",
      "rawMarkdown": "Sounds easy, but I believe there can be many minor problems. For example, if two models produce different FP / FN predictions, how would you average these FP / FN? What if one model is better for a short-range detection, while the other one produces best estimations for cars, which are far away from camera?",
      "votes": null
    },
    {
      "id": "702714",
      "postDate": "12/25/2019 03:51:23",
      "content": "<p>I gain way much worse result on averaging raw predictions with blending than single model. Maybe SWA will work?</p>",
      "rawMarkdown": "I gain way much worse result on averaging raw predictions with blending than single model. Maybe SWA will work?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 691789,
      "author_name": "greatgamedota",
      "author_url": "",
      "post_date": "12/10/2019 13:54:37",
      "content": "<p>I'm not super familiar with blending but I would assume you could just average the predictions of multiple models together.</p>",
      "votes": null,
      "replies": [
        {
          "id": 692397,
          "author_name": "dvorobiev",
          "author_url": "",
          "post_date": "12/11/2019 08:06:54",
          "content": "<p>Sounds easy, but I believe there can be many minor problems. For example, if two models produce different FP / FN predictions, how would you average these FP / FN? What if one model is better for a short-range detection, while the other one produces best estimations for cars, which are far away from camera?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 691894,
      "author_name": "bamps53",
      "author_url": "",
      "post_date": "12/10/2019 16:07:10",
      "content": "<p>If your models use same image size or preprocess, you just need to take average of your predictions before any post process. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 702714,
      "author_name": "xiejialun",
      "author_url": "",
      "post_date": "12/25/2019 03:51:23",
      "content": "<p>I gain way much worse result on averaging raw predictions with blending than single model. Maybe SWA will work?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "691679": "Does anyone know any blending or stacking techniques possible in this competition?",
    "691789": "I'm not super familiar with blending but I would assume you could just average the predictions of multiple models together.",
    "691894": "If your models use same image size or preprocess, you just need to take average of your predictions before any post process.",
    "692397": "Sounds easy, but I believe there can be many minor problems. For example, if two models produce different FP / FN predictions, how would you average these FP / FN? What if one model is better for a short-range detection, while the other one produces best estimations for cars, which are far away from camera?",
    "702714": "I gain way much worse result on averaging raw predictions with blending than single model. Maybe SWA will work?"
  },
  "source": "meta"
}