{
  "id": 303723,
  "title": "SNIPER for Large Scale Training",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/303723",
  "author_name": "",
  "post_date": "2022-01-29T02:14:18.845469600Z",
  "votes": 13,
  "comment_count": 5,
  "views": 0,
  "content": "<p>There is no doubt that large scale is the key to this competition，we can train in yolo5s with size 3000 and inference with size 9000~10000。if we can train with size 9000~10000 and inference with size 9000~10000 too,I believe the score will be further improved.@Remek Kinas introduce<br>\nSAHI to slice image,it's a great tool.but how to reasonably slice image in training stage,and select positive and negative samples？ the paper of SNIPER proposed the concept of chips，it can solve the problem of large-scale training.<br>\nthe paper link:<a href=\"url\" target=\"_blank\">https://arxiv.org/abs/1805.09300</a><br>\nthe code link:<a href=\"url\" target=\"_blank\">https://github.com/mahyarnajibi/SNIPER</a></p>",
  "messages": [
    {
      "id": "1667856",
      "postDate": "01/29/2022 02:14:18",
      "content": "<p>There is no doubt that large scale is the key to this competition，we can train in yolo5s with size 3000 and inference with size 9000~10000。if we can train with size 9000~10000 and inference with size 9000~10000 too,I believe the score will be further improved.@Remek Kinas introduce<br>\nSAHI to slice image,it's a great tool.but how to reasonably slice image in training stage,and select positive and negative samples？ the paper of SNIPER proposed the concept of chips，it can solve the problem of large-scale training.<br>\nthe paper link:<a href=\"url\" target=\"_blank\">https://arxiv.org/abs/1805.09300</a><br>\nthe code link:<a href=\"url\" target=\"_blank\">https://github.com/mahyarnajibi/SNIPER</a></p>",
      "rawMarkdown": "There is no doubt that large scale is the key to this competition，we can train in yolo5s with size 3000 and inference with size 9000~10000。if we can train with size 9000~10000 and inference with size 9000~10000 too,I believe the score will be further improved.@Remek Kinas introduce\nSAHI to slice image,it's a great tool.but how to reasonably slice image in training stage,and select positive and negative samples？ the paper of SNIPER proposed the concept of chips，it can solve the problem of large-scale training.\nthe paper link:[https://arxiv.org/abs/1805.09300](url)\nthe code link:[https://github.com/mahyarnajibi/SNIPER](url)",
      "votes": null
    },
    {
      "id": "1668060",
      "postDate": "01/29/2022 09:04:55",
      "content": "<p>some additional choice maybe..</p>\n<ol>\n<li>focal transformer: transformer for large scale input </li>\n<li>long transformer:  same as 1.</li>\n<li>sparse rcnn: efficient rpn focus on spare region of interest of input </li>\n</ol>",
      "rawMarkdown": "some additional choice maybe..\n\n1. focal transformer: transformer for large scale input \n2. long transformer:  same as 1.\n3. sparse rcnn: efficient rpn focus on spare region of interest of input",
      "votes": null
    },
    {
      "id": "1668064",
      "postDate": "01/29/2022 09:11:18",
      "content": "<p>thanks Drzhuzhe！</p>",
      "rawMarkdown": "thanks Drzhuzhe！",
      "votes": null
    },
    {
      "id": "1670679",
      "postDate": "01/31/2022 18:35:00",
      "content": "<p>SNIPER is example of bad paper, in this paper was created multiscale models and it call best single model.<br>\nIts paper was before FPN<br>\nIn this competition there is no problem with large-scale the problem in lack of training data (almost same images) and traing distribution have difference from test. For me it is unknown reason that resolution 10000 works</p>",
      "rawMarkdown": "SNIPER is example of bad paper, in this paper was created multiscale models and it call best single model.\nIts paper was before FPN\nIn this competition there is no problem with large-scale the problem in lack of training data (almost same images) and traing distribution have difference from test. For me it is unknown reason that resolution 10000 works",
      "votes": null
    },
    {
      "id": "1671046",
      "postDate": "02/01/2022 05:43:44",
      "content": "<blockquote>\n  <p>no problem with large-scale the problem in lack of training data (almost same images) and traing distribution have difference from test. For me it is unknown reason that resolution 10000 works</p>\n</blockquote>\n<p>ennnn<br>\nmaybe you can share some additional provement of this conclusion?</p>",
      "rawMarkdown": "> no problem with large-scale the problem in lack of training data (almost same images) and traing distribution have difference from test. For me it is unknown reason that resolution 10000 works\n\nennnn\nmaybe you can share some additional provement of this conclusion?",
      "votes": null
    },
    {
      "id": "1671733",
      "postDate": "02/01/2022 17:31:36",
      "content": "<p>image is 1280x720, when it is upscaling to 10000 does not add more information.<br>\nWhy it works at public LB I dont know.<br>\nFinding to train and test at massive upscale is wrong way<br>\nNo proof -- just my common sense</p>",
      "rawMarkdown": "image is 1280x720, when it is upscaling to 10000 does not add more information.\nWhy it works at public LB I dont know.\nFinding to train and test at massive upscale is wrong way\nNo proof -- just my common sense",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1668060,
      "author_name": "drzhuzhe",
      "author_url": "",
      "post_date": "01/29/2022 09:04:55",
      "content": "<p>some additional choice maybe..</p>\n<ol>\n<li>focal transformer: transformer for large scale input </li>\n<li>long transformer:  same as 1.</li>\n<li>sparse rcnn: efficient rpn focus on spare region of interest of input </li>\n</ol>",
      "votes": null,
      "replies": [
        {
          "id": 1668064,
          "author_name": "hanson0910",
          "author_url": "",
          "post_date": "01/29/2022 09:11:18",
          "content": "<p>thanks Drzhuzhe！</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1670679,
      "author_name": "nicksergievskiy",
      "author_url": "",
      "post_date": "01/31/2022 18:35:00",
      "content": "<p>SNIPER is example of bad paper, in this paper was created multiscale models and it call best single model.<br>\nIts paper was before FPN<br>\nIn this competition there is no problem with large-scale the problem in lack of training data (almost same images) and traing distribution have difference from test. For me it is unknown reason that resolution 10000 works</p>",
      "votes": null,
      "replies": [
        {
          "id": 1671046,
          "author_name": "drzhuzhe",
          "author_url": "",
          "post_date": "02/01/2022 05:43:44",
          "content": "<blockquote>\n  <p>no problem with large-scale the problem in lack of training data (almost same images) and traing distribution have difference from test. For me it is unknown reason that resolution 10000 works</p>\n</blockquote>\n<p>ennnn<br>\nmaybe you can share some additional provement of this conclusion?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1671733,
          "author_name": "nicksergievskiy",
          "author_url": "",
          "post_date": "02/01/2022 17:31:36",
          "content": "<p>image is 1280x720, when it is upscaling to 10000 does not add more information.<br>\nWhy it works at public LB I dont know.<br>\nFinding to train and test at massive upscale is wrong way<br>\nNo proof -- just my common sense</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1667856": "There is no doubt that large scale is the key to this competition，we can train in yolo5s with size 3000 and inference with size 9000~10000。if we can train with size 9000~10000 and inference with size 9000~10000 too,I believe the score will be further improved.@Remek Kinas introduce\nSAHI to slice image,it's a great tool.but how to reasonably slice image in training stage,and select positive and negative samples？ the paper of SNIPER proposed the concept of chips，it can solve the problem of large-scale training.\nthe paper link:[https://arxiv.org/abs/1805.09300](url)\nthe code link:[https://github.com/mahyarnajibi/SNIPER](url)",
    "1668060": "some additional choice maybe..\n\n1. focal transformer: transformer for large scale input \n2. long transformer:  same as 1.\n3. sparse rcnn: efficient rpn focus on spare region of interest of input",
    "1668064": "thanks Drzhuzhe！",
    "1670679": "SNIPER is example of bad paper, in this paper was created multiscale models and it call best single model.\nIts paper was before FPN\nIn this competition there is no problem with large-scale the problem in lack of training data (almost same images) and traing distribution have difference from test. For me it is unknown reason that resolution 10000 works",
    "1671046": "> no problem with large-scale the problem in lack of training data (almost same images) and traing distribution have difference from test. For me it is unknown reason that resolution 10000 works\n\nennnn\nmaybe you can share some additional provement of this conclusion?",
    "1671733": "image is 1280x720, when it is upscaling to 10000 does not add more information.\nWhy it works at public LB I dont know.\nFinding to train and test at massive upscale is wrong way\nNo proof -- just my common sense"
  },
  "source": "meta"
}