{
  "id": 124912,
  "title": "seems pseudo labels works...",
  "url": "/competitions/pku-autonomous-driving/discussion/124912",
  "author_name": "",
  "post_date": "2020-01-07T13:06:17.136807300Z",
  "votes": 17,
  "comment_count": 12,
  "views": 0,
  "content": "<p>My training schema as follows:</p>\n\n<ol>\n<li>train single model for 4 random_seed, lb range from 0.070~0.075, ensemble to get lb=0.08</li>\n<li>use ensemble to predict on testset, sort by confidence, select highest 1000 samples to add up to gt trainset</li>\n<li>train again on gt+pseudo labels, lb boost 0.074--&gt;0.079, 0.005+</li>\n</ol>",
  "messages": [
    {
      "id": "712623",
      "postDate": "01/07/2020 13:06:17",
      "content": "<p>My training schema as follows:</p>\n\n<ol>\n<li>train single model for 4 random_seed, lb range from 0.070~0.075, ensemble to get lb=0.08</li>\n<li>use ensemble to predict on testset, sort by confidence, select highest 1000 samples to add up to gt trainset</li>\n<li>train again on gt+pseudo labels, lb boost 0.074--&gt;0.079, 0.005+</li>\n</ol>",
      "rawMarkdown": "My training schema as follows:\n\n1. train single model for 4 random_seed, lb range from 0.070~0.075, ensemble to get lb=0.08\n2. use ensemble to predict on testset, sort by confidence, select highest 1000 samples to add up to gt trainset\n3. train again on gt+pseudo labels, lb boost 0.074--&gt;0.079, 0.005+",
      "votes": null
    },
    {
      "id": "712695",
      "postDate": "01/07/2020 14:09:17",
      "content": "<p>May I ask how you do the ensemble part? Averaging the output feature maps before decoding?\nI'm still playing with different decoder architectures...have no chance to try the ensemble.\n(You don't have to answer the question if you think this is your special trick, no harm feeling)</p>",
      "rawMarkdown": "May I ask how you do the ensemble part? Averaging the output feature maps before decoding?\nI'm still playing with different decoder architectures...have no chance to try the ensemble.\n(You don't have to answer the question if you think this is your special trick, no harm feeling)",
      "votes": null
    },
    {
      "id": "712710",
      "postDate": "01/07/2020 14:23:56",
      "content": "<p>Do you augment any of the train images? Do you use the masks to preprocess train and test images for training and predicting? Thanks for sharing your experiments!</p>",
      "rawMarkdown": "Do you augment any of the train images? Do you use the masks to preprocess train and test images for training and predicting? Thanks for sharing your experiments!",
      "votes": null
    },
    {
      "id": "712750",
      "postDate": "01/07/2020 14:55:53",
      "content": "<p>no trick, average raw predictions from 4 checkpoints.</p>",
      "rawMarkdown": "no trick, average raw predictions from 4 checkpoints.",
      "votes": null
    },
    {
      "id": "712751",
      "postDate": "01/07/2020 14:57:43",
      "content": "<p>Yes, augmentations</p>",
      "rawMarkdown": "Yes, augmentations",
      "votes": null
    },
    {
      "id": "713143",
      "postDate": "01/08/2020 00:21:03",
      "content": "<p>Thanks for the confirmation!</p>",
      "rawMarkdown": "Thanks for the confirmation!",
      "votes": null
    },
    {
      "id": "713991",
      "postDate": "01/08/2020 22:36:02",
      "content": "<p>may I ask your a silly question？what‘s the gt means?</p>",
      "rawMarkdown": "may I ask your a silly question？what‘s the gt means?",
      "votes": null
    },
    {
      "id": "714049",
      "postDate": "01/09/2020 01:44:07",
      "content": "<p>*<strong>G</strong>*round <strong>*T</strong>*ruth</p>",
      "rawMarkdown": "***G***round ***T***ruth",
      "votes": null
    },
    {
      "id": "714781",
      "postDate": "01/09/2020 18:29:55",
      "content": "<p>Did you save the raw predictions? Predicting with all 4 models sounds like a lot of GRAM usage.</p>",
      "rawMarkdown": "Did you save the raw predictions? Predicting with all 4 models sounds like a lot of GRAM usage.",
      "votes": null
    },
    {
      "id": "714960",
      "postDate": "01/10/2020 01:27:17",
      "content": "<p><a href=\"/tonychenxyz\">@tonychenxyz</a> I think maybe you can just predict with very small batch? Like 1 or 2</p>",
      "rawMarkdown": "tonychenxyz I think maybe you can just predict with very small batch? Like 1 or 2",
      "votes": null
    },
    {
      "id": "715517",
      "postDate": "01/10/2020 16:00:49",
      "content": "<p>maybe it is a silly question？I don't know how to average raw predictions from 4 checkpoints, or set a threshold ，Then </p>\n\n<p>take the intersection of four checkpoints.</p>",
      "rawMarkdown": "maybe it is a silly question？I don't know how to average raw predictions from 4 checkpoints, or set a threshold ，Then \n\ntake the intersection of four checkpoints.",
      "votes": null
    },
    {
      "id": "723755",
      "postDate": "01/20/2020 12:17:32",
      "content": "<p>Thanks for sharing,  I'm curious about the implementation of \"select highest k samples\". Here you mean by cars or by images?  For example Image A has 3 cars with confidence 0.9, 0.8, 0.7, Image B has 1 car with confidence 0.6, if select 2 highest samples will you get Both A and B or A only?</p>",
      "rawMarkdown": "Thanks for sharing,  I'm curious about the implementation of \"select highest k samples\". Here you mean by cars or by images?  For example Image A has 3 cars with confidence 0.9, 0.8, 0.7, Image B has 1 car with confidence 0.6, if select 2 highest samples will you get Both A and B or A only?",
      "votes": null
    },
    {
      "id": "724315",
      "postDate": "01/21/2020 03:15:06",
      "content": "<p>Did you select 1000 masks or 1000 photos?</p>",
      "rawMarkdown": "Did you select 1000 masks or 1000 photos?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 712695,
      "author_name": "xiejialun",
      "author_url": "",
      "post_date": "01/07/2020 14:09:17",
      "content": "<p>May I ask how you do the ensemble part? Averaging the output feature maps before decoding?\nI'm still playing with different decoder architectures...have no chance to try the ensemble.\n(You don't have to answer the question if you think this is your special trick, no harm feeling)</p>",
      "votes": null,
      "replies": [
        {
          "id": 712750,
          "author_name": "niuddd",
          "author_url": "",
          "post_date": "01/07/2020 14:55:53",
          "content": "<p>no trick, average raw predictions from 4 checkpoints.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 713143,
          "author_name": "xiejialun",
          "author_url": "",
          "post_date": "01/08/2020 00:21:03",
          "content": "<p>Thanks for the confirmation!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 714781,
          "author_name": "tonychenxyz",
          "author_url": "",
          "post_date": "01/09/2020 18:29:55",
          "content": "<p>Did you save the raw predictions? Predicting with all 4 models sounds like a lot of GRAM usage.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 714960,
          "author_name": "xiejialun",
          "author_url": "",
          "post_date": "01/10/2020 01:27:17",
          "content": "<p><a href=\"/tonychenxyz\">@tonychenxyz</a> I think maybe you can just predict with very small batch? Like 1 or 2</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 712710,
      "author_name": "greatgamedota",
      "author_url": "",
      "post_date": "01/07/2020 14:23:56",
      "content": "<p>Do you augment any of the train images? Do you use the masks to preprocess train and test images for training and predicting? Thanks for sharing your experiments!</p>",
      "votes": null,
      "replies": [
        {
          "id": 712751,
          "author_name": "niuddd",
          "author_url": "",
          "post_date": "01/07/2020 14:57:43",
          "content": "<p>Yes, augmentations</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 713991,
      "author_name": "",
      "author_url": "",
      "post_date": "01/08/2020 22:36:02",
      "content": "<p>may I ask your a silly question？what‘s the gt means?</p>",
      "votes": null,
      "replies": [
        {
          "id": 714049,
          "author_name": "niuddd",
          "author_url": "",
          "post_date": "01/09/2020 01:44:07",
          "content": "<p>*<strong>G</strong>*round <strong>*T</strong>*ruth</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 715517,
      "author_name": "anning2",
      "author_url": "",
      "post_date": "01/10/2020 16:00:49",
      "content": "<p>maybe it is a silly question？I don't know how to average raw predictions from 4 checkpoints, or set a threshold ，Then </p>\n\n<p>take the intersection of four checkpoints.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 723755,
      "author_name": "ccchang801023",
      "author_url": "",
      "post_date": "01/20/2020 12:17:32",
      "content": "<p>Thanks for sharing,  I'm curious about the implementation of \"select highest k samples\". Here you mean by cars or by images?  For example Image A has 3 cars with confidence 0.9, 0.8, 0.7, Image B has 1 car with confidence 0.6, if select 2 highest samples will you get Both A and B or A only?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 724315,
      "author_name": "tonychenxyz",
      "author_url": "",
      "post_date": "01/21/2020 03:15:06",
      "content": "<p>Did you select 1000 masks or 1000 photos?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "712623": "My training schema as follows:\n\n1. train single model for 4 random_seed, lb range from 0.070~0.075, ensemble to get lb=0.08\n2. use ensemble to predict on testset, sort by confidence, select highest 1000 samples to add up to gt trainset\n3. train again on gt+pseudo labels, lb boost 0.074--&gt;0.079, 0.005+",
    "712695": "May I ask how you do the ensemble part? Averaging the output feature maps before decoding?\nI'm still playing with different decoder architectures...have no chance to try the ensemble.\n(You don't have to answer the question if you think this is your special trick, no harm feeling)",
    "712710": "Do you augment any of the train images? Do you use the masks to preprocess train and test images for training and predicting? Thanks for sharing your experiments!",
    "712750": "no trick, average raw predictions from 4 checkpoints.",
    "712751": "Yes, augmentations",
    "713143": "Thanks for the confirmation!",
    "713991": "may I ask your a silly question？what‘s the gt means?",
    "714049": "***G***round ***T***ruth",
    "714781": "Did you save the raw predictions? Predicting with all 4 models sounds like a lot of GRAM usage.",
    "714960": "tonychenxyz I think maybe you can just predict with very small batch? Like 1 or 2",
    "715517": "maybe it is a silly question？I don't know how to average raw predictions from 4 checkpoints, or set a threshold ，Then \n\ntake the intersection of four checkpoints.",
    "723755": "Thanks for sharing,  I'm curious about the implementation of \"select highest k samples\". Here you mean by cars or by images?  For example Image A has 3 cars with confidence 0.9, 0.8, 0.7, Image B has 1 car with confidence 0.6, if select 2 highest samples will you get Both A and B or A only?",
    "724315": "Did you select 1000 masks or 1000 photos?"
  },
  "source": "meta"
}