{
  "id": 163760,
  "title": "Training pipeline to get > 0.76 LB score",
  "url": "/competitions/global-wheat-detection/discussion/163760",
  "author_name": "",
  "post_date": "2020-07-03T11:18:31.729496800Z",
  "votes": 24,
  "comment_count": 9,
  "views": 0,
  "content": "<p>Hey guys, I created a notebook to explain how to get a 0.76 score from A to Z. I ve explained how to train with coco pretrained weights + some tips on how to train on colab without having runtime issues. \ncheck it out <a href=\"https://www.kaggle.com/nacirbouazizi/well-explained-training-pipeline-to-get-lb-0-76\">here</a>.\nDon't forget to upvote if you like it and thank you 👐 </p>",
  "messages": [
    {
      "id": "913713",
      "postDate": "07/03/2020 11:18:31",
      "content": "<p>Hey guys, I created a notebook to explain how to get a 0.76 score from A to Z. I ve explained how to train with coco pretrained weights + some tips on how to train on colab without having runtime issues. \ncheck it out <a href=\"https://www.kaggle.com/nacirbouazizi/well-explained-training-pipeline-to-get-lb-0-76\">here</a>.\nDon't forget to upvote if you like it and thank you 👐 </p>",
      "rawMarkdown": "Hey guys, I created a notebook to explain how to get a 0.76 score from A to Z. I ve explained how to train with coco pretrained weights + some tips on how to train on colab without having runtime issues. \ncheck it out [here](https://www.kaggle.com/nacirbouazizi/well-explained-training-pipeline-to-get-lb-0-76).\nDon't forget to upvote if you like it and thank you 👐",
      "votes": null
    },
    {
      "id": "914029",
      "postDate": "07/03/2020 15:23:58",
      "content": "<p>Thanks. Just wondering why all yolo public kernels use <code>index = list(set(df.image_id))</code> instead of stratified k fold</p>",
      "rawMarkdown": "Thanks. Just wondering why all yolo public kernels use `index = list(set(df.image_id))` instead of stratified k fold",
      "votes": null
    },
    {
      "id": "914070",
      "postDate": "07/03/2020 15:55:08",
      "content": "<p>Regarding your reported 0.73x score of single model, is it without TTA or with TTA?</p>",
      "rawMarkdown": "Regarding your reported 0.73x score of single model, is it without TTA or with TTA?",
      "votes": null
    },
    {
      "id": "914082",
      "postDate": "07/03/2020 16:07:04",
      "content": "<p>Yeah I got the same question, I m going to test it and see. </p>",
      "rawMarkdown": "Yeah I got the same question, I m going to test it and see.",
      "votes": null
    },
    {
      "id": "914084",
      "postDate": "07/03/2020 16:07:54",
      "content": "<p>With TTA from this notebook : <a href=\"https://www.kaggle.com/wasupandceacar/yolov5-single-model-with-more-tta-lb-0-745\">yolov5-single-model-with-more-tta-lb-0-745</a></p>",
      "rawMarkdown": "With TTA from this notebook : [yolov5-single-model-with-more-tta-lb-0-745](https://www.kaggle.com/wasupandceacar/yolov5-single-model-with-more-tta-lb-0-745)",
      "votes": null
    },
    {
      "id": "914094",
      "postDate": "07/03/2020 16:13:17",
      "content": "<p>Got it, this is very helpful for someone not familiar with yolov5 (like me)!</p>",
      "rawMarkdown": "Got it, this is very helpful for someone not familiar with yolov5 (like me)!",
      "votes": null
    },
    {
      "id": "914097",
      "postDate": "07/03/2020 16:17:59",
      "content": "<p>Thanks ! Btw are you using EfficientDet D5 ? </p>",
      "rawMarkdown": "Thanks ! Btw are you using EfficientDet D5 ?",
      "votes": null
    },
    {
      "id": "915803",
      "postDate": "07/05/2020 05:23:12",
      "content": "<p>In my opiniton, we don't need to use stratified k fold because the number of labels are only one in this competition.</p>\n\n<p>If you want more randomness in training data, please try this code.\n```\nimport random</p>\n\n<p>index = list(set(df.image_id))\nrandom.seed(123)\nindex = random.sample(index, len(index))\n```</p>",
      "rawMarkdown": "In my opiniton, we don't need to use stratified k fold because the number of labels are only one in this competition.\n\nIf you want more randomness in training data, please try this code.\n```\nimport random\n\nindex = list(set(df.image_id))\nrandom.seed(123)\nindex = random.sample(index, len(index))\n```",
      "votes": null
    },
    {
      "id": "915957",
      "postDate": "07/05/2020 08:36:01",
      "content": "<p>Agree but I was thinking about sources, not classes</p>",
      "rawMarkdown": "Agree but I was thinking about sources, not classes",
      "votes": null
    },
    {
      "id": "916163",
      "postDate": "07/05/2020 12:41:19",
      "content": "<p>Thanks <a href=\"/spshota\">@spshota</a>  ! I ll try that ! But I made a startified kfold on number of boxes per image + source </p>",
      "rawMarkdown": "Thanks @spshota  ! I ll try that ! But I made a startified kfold on number of boxes per image + source",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 914029,
      "author_name": "alexj21",
      "author_url": "",
      "post_date": "07/03/2020 15:23:58",
      "content": "<p>Thanks. Just wondering why all yolo public kernels use <code>index = list(set(df.image_id))</code> instead of stratified k fold</p>",
      "votes": null,
      "replies": [
        {
          "id": 914082,
          "author_name": "nacirbouazizi",
          "author_url": "",
          "post_date": "07/03/2020 16:07:04",
          "content": "<p>Yeah I got the same question, I m going to test it and see. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 915803,
          "author_name": "spshota",
          "author_url": "",
          "post_date": "07/05/2020 05:23:12",
          "content": "<p>In my opiniton, we don't need to use stratified k fold because the number of labels are only one in this competition.</p>\n\n<p>If you want more randomness in training data, please try this code.\n```\nimport random</p>\n\n<p>index = list(set(df.image_id))\nrandom.seed(123)\nindex = random.sample(index, len(index))\n```</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 915957,
          "author_name": "alexj21",
          "author_url": "",
          "post_date": "07/05/2020 08:36:01",
          "content": "<p>Agree but I was thinking about sources, not classes</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 916163,
          "author_name": "nacirbouazizi",
          "author_url": "",
          "post_date": "07/05/2020 12:41:19",
          "content": "<p>Thanks <a href=\"/spshota\">@spshota</a>  ! I ll try that ! But I made a startified kfold on number of boxes per image + source </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 914070,
      "author_name": "alexanderliao",
      "author_url": "",
      "post_date": "07/03/2020 15:55:08",
      "content": "<p>Regarding your reported 0.73x score of single model, is it without TTA or with TTA?</p>",
      "votes": null,
      "replies": [
        {
          "id": 914084,
          "author_name": "nacirbouazizi",
          "author_url": "",
          "post_date": "07/03/2020 16:07:54",
          "content": "<p>With TTA from this notebook : <a href=\"https://www.kaggle.com/wasupandceacar/yolov5-single-model-with-more-tta-lb-0-745\">yolov5-single-model-with-more-tta-lb-0-745</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 914094,
          "author_name": "alexanderliao",
          "author_url": "",
          "post_date": "07/03/2020 16:13:17",
          "content": "<p>Got it, this is very helpful for someone not familiar with yolov5 (like me)!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 914097,
          "author_name": "nacirbouazizi",
          "author_url": "",
          "post_date": "07/03/2020 16:17:59",
          "content": "<p>Thanks ! Btw are you using EfficientDet D5 ? </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "913713": "Hey guys, I created a notebook to explain how to get a 0.76 score from A to Z. I ve explained how to train with coco pretrained weights + some tips on how to train on colab without having runtime issues. \ncheck it out [here](https://www.kaggle.com/nacirbouazizi/well-explained-training-pipeline-to-get-lb-0-76).\nDon't forget to upvote if you like it and thank you 👐",
    "914029": "Thanks. Just wondering why all yolo public kernels use `index = list(set(df.image_id))` instead of stratified k fold",
    "914070": "Regarding your reported 0.73x score of single model, is it without TTA or with TTA?",
    "914082": "Yeah I got the same question, I m going to test it and see.",
    "914084": "With TTA from this notebook : [yolov5-single-model-with-more-tta-lb-0-745](https://www.kaggle.com/wasupandceacar/yolov5-single-model-with-more-tta-lb-0-745)",
    "914094": "Got it, this is very helpful for someone not familiar with yolov5 (like me)!",
    "914097": "Thanks ! Btw are you using EfficientDet D5 ?",
    "915803": "In my opiniton, we don't need to use stratified k fold because the number of labels are only one in this competition.\n\nIf you want more randomness in training data, please try this code.\n```\nimport random\n\nindex = list(set(df.image_id))\nrandom.seed(123)\nindex = random.sample(index, len(index))\n```",
    "915957": "Agree but I was thinking about sources, not classes",
    "916163": "Thanks @spshota  ! I ll try that ! But I made a startified kfold on number of boxes per image + source"
  },
  "source": "meta"
}