{
  "id": 164006,
  "title": "TTA for fasterrcnn",
  "url": "/competitions/global-wheat-detection/discussion/164006",
  "author_name": "",
  "post_date": "2020-07-04T11:37:29.978890500Z",
  "votes": 7,
  "comment_count": 11,
  "views": 0,
  "content": "<p>Hello everyone, I have just public a <a href=\"https://www.kaggle.com/whurobin/wbf-over-tta-single-model-fasterrcnn-resnest?\">kernel</a> using fasterrcnn and TTA, it reaches LB 74.11, I just use flip and rotate90 as TTA tricks.\nMy training code can be find in  this <a href=\"https://github.com/CVerRobin/wheatdetection\">repo</a>\nMy code is based on resnest101e and torchvision.\nSingle model without TTA reaches LB 73.61.\nIf it's helpful, please star it!</p>",
  "messages": [
    {
      "id": "914955",
      "postDate": "07/04/2020 11:37:29",
      "content": "<p>Hello everyone, I have just public a <a href=\"https://www.kaggle.com/whurobin/wbf-over-tta-single-model-fasterrcnn-resnest?\">kernel</a> using fasterrcnn and TTA, it reaches LB 74.11, I just use flip and rotate90 as TTA tricks.\nMy training code can be find in  this <a href=\"https://github.com/CVerRobin/wheatdetection\">repo</a>\nMy code is based on resnest101e and torchvision.\nSingle model without TTA reaches LB 73.61.\nIf it's helpful, please star it!</p>",
      "rawMarkdown": "Hello everyone, I have just public a [kernel](https://www.kaggle.com/whurobin/wbf-over-tta-single-model-fasterrcnn-resnest?) using fasterrcnn and TTA, it reaches LB 74.11, I just use flip and rotate90 as TTA tricks.\nMy training code can be find in  this [repo](https://github.com/CVerRobin/wheatdetection)\nMy code is based on resnest101e and torchvision.\nSingle model without TTA reaches LB 73.61.\nIf it's helpful, please star it!",
      "votes": null
    },
    {
      "id": "915411",
      "postDate": "07/04/2020 17:53:23",
      "content": "<p>That's awesome.</p>",
      "rawMarkdown": "That's awesome.",
      "votes": null
    },
    {
      "id": "915627",
      "postDate": "07/05/2020 00:20:29",
      "content": "<p>Updated.\nI made a inference notebook with my whole project.\nPlease look at <a href=\"https://www.kaggle.com/whurobin/fastercnn-resnest-inference?scriptVersionId=38099400\">here</a>\nCode can be git from my <a href=\"https://github.com/CVerRobin/wheatdetection\">repo</a></p>",
      "rawMarkdown": "Updated.\nI made a inference notebook with my whole project.\nPlease look at [here](https://www.kaggle.com/whurobin/fastercnn-resnest-inference?scriptVersionId=38099400)\nCode can be git from my [repo](https://github.com/CVerRobin/wheatdetection)",
      "votes": null
    },
    {
      "id": "915645",
      "postDate": "07/05/2020 01:09:59",
      "content": "<p>Thanks!</p>",
      "rawMarkdown": "Thanks!",
      "votes": null
    },
    {
      "id": "915663",
      "postDate": "07/05/2020 02:05:16",
      "content": "<p>Thank you and great job!</p>",
      "rawMarkdown": "Thank you and great job!",
      "votes": null
    },
    {
      "id": "915683",
      "postDate": "07/05/2020 02:46:52",
      "content": "<p>Thanks!\nI have no idea how to do multi-scale, is it just input multi resized images?</p>",
      "rawMarkdown": "Thanks!\nI have no idea how to do multi-scale, is it just input multi resized images?",
      "votes": null
    },
    {
      "id": "915736",
      "postDate": "07/05/2020 04:04:30",
      "content": "<p>Not sure if you mean multi-scale in YoloV5.</p>\n\n<p>If yes, according to the YoloV5 training logs I can observed (I didn't read codes in details so I might be wrong). I think \"yes\", it resizes the images in every epoch between 0.5x to 1.5x</p>",
      "rawMarkdown": "Not sure if you mean multi-scale in YoloV5.\n\nIf yes, according to the YoloV5 training logs I can observed (I didn't read codes in details so I might be wrong). I think \"yes\", it resizes the images in every epoch between 0.5x to 1.5x",
      "votes": null
    },
    {
      "id": "915743",
      "postDate": "07/05/2020 04:11:57",
      "content": "<p>Thanks, I'll try to use it in my repo.</p>",
      "rawMarkdown": "Thanks, I'll try to use it in my repo.",
      "votes": null
    },
    {
      "id": "915822",
      "postDate": "07/05/2020 05:50:53",
      "content": "<p>Did you try mosaic augmentation like YoloV5 ? <a href=\"/whurobin\">@whurobin</a> \n(or maybe CutMix)</p>",
      "rawMarkdown": "Did you try mosaic augmentation like YoloV5 ? @whurobin \n(or maybe CutMix)",
      "votes": null
    },
    {
      "id": "915831",
      "postDate": "07/05/2020 05:56:45",
      "content": "<p>Yes, you can see my <a href=\"https://github.com/CVerRobin/wheatdetection\">repo</a>.</p>",
      "rawMarkdown": "Yes, you can see my [repo](https://github.com/CVerRobin/wheatdetection).",
      "votes": null
    },
    {
      "id": "915844",
      "postDate": "07/05/2020 06:20:39",
      "content": "<p>Great. Now I see this : \n<code>\n            if p_ratio &amp;lt; 0.2:\n                image, boxes = self.load_mosaic_image_and_boxes(index)\n            elif p_ratio &amp;lt; 0.4:\n                image, boxes = self.load_image_and_bboxes_with_cutmix(index)\n            else:\n                image, boxes = self.load_mixup_image_and_boxes(index)\n</code>\nMaybe you would try mosaic alone (not cutmix/mixup), to (maybe) get better performance (imitate YoloV5)</p>",
      "rawMarkdown": "Great. Now I see this : \n```\n            if p_ratio &lt; 0.2:\n                image, boxes = self.load_mosaic_image_and_boxes(index)\n            elif p_ratio &lt; 0.4:\n                image, boxes = self.load_image_and_bboxes_with_cutmix(index)\n            else:\n                image, boxes = self.load_mixup_image_and_boxes(index)\n```\nMaybe you would try mosaic alone (not cutmix/mixup), to (maybe) get better performance (imitate YoloV5)",
      "votes": null
    },
    {
      "id": "915916",
      "postDate": "07/05/2020 07:41:40",
      "content": "<p>Just moasic will get better performance? I'll try it, thanks!</p>",
      "rawMarkdown": "Just moasic will get better performance? I'll try it, thanks!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 915411,
      "author_name": "jonykarki",
      "author_url": "",
      "post_date": "07/04/2020 17:53:23",
      "content": "<p>That's awesome.</p>",
      "votes": null,
      "replies": [
        {
          "id": 915645,
          "author_name": "whurobin",
          "author_url": "",
          "post_date": "07/05/2020 01:09:59",
          "content": "<p>Thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 915627,
      "author_name": "whurobin",
      "author_url": "",
      "post_date": "07/05/2020 00:20:29",
      "content": "<p>Updated.\nI made a inference notebook with my whole project.\nPlease look at <a href=\"https://www.kaggle.com/whurobin/fastercnn-resnest-inference?scriptVersionId=38099400\">here</a>\nCode can be git from my <a href=\"https://github.com/CVerRobin/wheatdetection\">repo</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 915663,
      "author_name": "ratthachat",
      "author_url": "",
      "post_date": "07/05/2020 02:05:16",
      "content": "<p>Thank you and great job!</p>",
      "votes": null,
      "replies": [
        {
          "id": 915683,
          "author_name": "whurobin",
          "author_url": "",
          "post_date": "07/05/2020 02:46:52",
          "content": "<p>Thanks!\nI have no idea how to do multi-scale, is it just input multi resized images?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 915736,
          "author_name": "ratthachat",
          "author_url": "",
          "post_date": "07/05/2020 04:04:30",
          "content": "<p>Not sure if you mean multi-scale in YoloV5.</p>\n\n<p>If yes, according to the YoloV5 training logs I can observed (I didn't read codes in details so I might be wrong). I think \"yes\", it resizes the images in every epoch between 0.5x to 1.5x</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 915743,
          "author_name": "whurobin",
          "author_url": "",
          "post_date": "07/05/2020 04:11:57",
          "content": "<p>Thanks, I'll try to use it in my repo.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 915822,
          "author_name": "ratthachat",
          "author_url": "",
          "post_date": "07/05/2020 05:50:53",
          "content": "<p>Did you try mosaic augmentation like YoloV5 ? <a href=\"/whurobin\">@whurobin</a> \n(or maybe CutMix)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 915831,
          "author_name": "whurobin",
          "author_url": "",
          "post_date": "07/05/2020 05:56:45",
          "content": "<p>Yes, you can see my <a href=\"https://github.com/CVerRobin/wheatdetection\">repo</a>.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 915844,
          "author_name": "ratthachat",
          "author_url": "",
          "post_date": "07/05/2020 06:20:39",
          "content": "<p>Great. Now I see this : \n<code>\n            if p_ratio &amp;lt; 0.2:\n                image, boxes = self.load_mosaic_image_and_boxes(index)\n            elif p_ratio &amp;lt; 0.4:\n                image, boxes = self.load_image_and_bboxes_with_cutmix(index)\n            else:\n                image, boxes = self.load_mixup_image_and_boxes(index)\n</code>\nMaybe you would try mosaic alone (not cutmix/mixup), to (maybe) get better performance (imitate YoloV5)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 915916,
          "author_name": "whurobin",
          "author_url": "",
          "post_date": "07/05/2020 07:41:40",
          "content": "<p>Just moasic will get better performance? I'll try it, thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "914955": "Hello everyone, I have just public a [kernel](https://www.kaggle.com/whurobin/wbf-over-tta-single-model-fasterrcnn-resnest?) using fasterrcnn and TTA, it reaches LB 74.11, I just use flip and rotate90 as TTA tricks.\nMy training code can be find in  this [repo](https://github.com/CVerRobin/wheatdetection)\nMy code is based on resnest101e and torchvision.\nSingle model without TTA reaches LB 73.61.\nIf it's helpful, please star it!",
    "915411": "That's awesome.",
    "915627": "Updated.\nI made a inference notebook with my whole project.\nPlease look at [here](https://www.kaggle.com/whurobin/fastercnn-resnest-inference?scriptVersionId=38099400)\nCode can be git from my [repo](https://github.com/CVerRobin/wheatdetection)",
    "915645": "Thanks!",
    "915663": "Thank you and great job!",
    "915683": "Thanks!\nI have no idea how to do multi-scale, is it just input multi resized images?",
    "915736": "Not sure if you mean multi-scale in YoloV5.\n\nIf yes, according to the YoloV5 training logs I can observed (I didn't read codes in details so I might be wrong). I think \"yes\", it resizes the images in every epoch between 0.5x to 1.5x",
    "915743": "Thanks, I'll try to use it in my repo.",
    "915822": "Did you try mosaic augmentation like YoloV5 ? @whurobin \n(or maybe CutMix)",
    "915831": "Yes, you can see my [repo](https://github.com/CVerRobin/wheatdetection).",
    "915844": "Great. Now I see this : \n```\n            if p_ratio &lt; 0.2:\n                image, boxes = self.load_mosaic_image_and_boxes(index)\n            elif p_ratio &lt; 0.4:\n                image, boxes = self.load_image_and_bboxes_with_cutmix(index)\n            else:\n                image, boxes = self.load_mixup_image_and_boxes(index)\n```\nMaybe you would try mosaic alone (not cutmix/mixup), to (maybe) get better performance (imitate YoloV5)",
    "915916": "Just moasic will get better performance? I'll try it, thanks!"
  },
  "source": "meta"
}