{
  "id": 90250,
  "title": "Image Classification Tricks",
  "url": "/competitions/imet-2019-fgvc6/discussion/90250",
  "author_name": "seefun",
  "post_date": "2019-04-22T04:34:07.105000",
  "votes": 30,
  "comment_count": 13,
  "views": 0,
  "content": "<p>This is an excellent paper about training tricks in image classification. <a href=\"https://arxiv.org/abs/1812.01187\">Bag of Tricks for Image Classification with Convolutional Neural Networks</a></p>\n\n<p>There are many tricks include: Cosine Learning Rate Decay, Label Smoothing, Knowledge Distillation, Mixup Training, Learning rate warmup, No bias decay in BN, Zero gamma in ResNet block and so on. Very intersting paper from Mu Li et. al.</p>",
  "messages": [
    {
      "id": 520961,
      "postDate": "2019-04-22T04:34:07.107Z",
      "content": "<p>This is an excellent paper about training tricks in image classification. <a href=\"https://arxiv.org/abs/1812.01187\">Bag of Tricks for Image Classification with Convolutional Neural Networks</a></p>\n\n<p>There are many tricks include: Cosine Learning Rate Decay, Label Smoothing, Knowledge Distillation, Mixup Training, Learning rate warmup, No bias decay in BN, Zero gamma in ResNet block and so on. Very intersting paper from Mu Li et. al.</p>",
      "rawMarkdown": "This is an excellent paper about training tricks in image classification. [Bag of Tricks for Image Classification with Convolutional Neural Networks](https://arxiv.org/abs/1812.01187)\n\nThere are many tricks include: Cosine Learning Rate Decay, Label Smoothing, Knowledge Distillation, Mixup Training, Learning rate warmup, No bias decay in BN, Zero gamma in ResNet block and so on. Very intersting paper from Mu Li et. al.",
      "votes": 30
    },
    {
      "id": 537428,
      "postDate": "2019-05-27T02:35:49.017Z",
      "content": "<p>Interesting article, the implementations are done in MXNet, which was invented by Mu Li.</p>",
      "rawMarkdown": "Interesting article, the implementations are done in MXNet, which was invented by Mu Li.",
      "votes": 1
    },
    {
      "id": 523459,
      "postDate": "2019-04-26T10:40:14.520Z",
      "content": "<p>Thx for sharing, methods of the paper turely illuminating</p>",
      "rawMarkdown": "Thx for sharing, methods of the paper turely illuminating",
      "votes": 1
    },
    {
      "id": 1699969,
      "postDate": "2022-02-21T15:10:51.317Z",
      "content": "<p>Interesting paper!, Do you know if is there any notebook in Kaggle implementing one of these ideias</p>",
      "rawMarkdown": "Interesting paper!, Do you know if is there any notebook in Kaggle implementing one of these ideias"
    },
    {
      "id": 956845,
      "postDate": "2020-08-03T20:43:47.373Z",
      "content": "<p>Hello Guys. I choose ResNet for İmage Classification. I thınk you should check thıs code\n<a href=\"https://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning\">https://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning</a></p>",
      "rawMarkdown": "Hello Guys. I choose ResNet for İmage Classification. I thınk you should check thıs code\nhttps://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning"
    },
    {
      "id": 531592,
      "postDate": "2019-05-15T07:19:22.997Z",
      "content": "<p>Hello，have you tried to Low-precision training?how to code?</p>",
      "rawMarkdown": "Hello，have you tried to Low-precision training?how to code?",
      "replies": [
        {
          "id": 531728,
          "postDate": "2019-05-15T12:39:01.550Z",
          "content": "<p>I have not try it. You can use Mixed Precision Training.  <a href=\"https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html\">https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html</a></p>",
          "rawMarkdown": "I have not try it. You can use Mixed Precision Training.  https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html",
          "votes": 2
        },
        {
          "id": 531736,
          "postDate": "2019-05-15T12:50:41.880Z",
          "content": "<p>Thanks!</p>",
          "rawMarkdown": "Thanks!"
        },
        {
          "id": 531746,
          "postDate": "2019-05-15T13:19:33.047Z",
          "content": "<p>I've tried for my pytorch model but loss didn't converge somehow...\nAnd I also tried fastai's very simple callbacks(<a href=\"https://docs.fast.ai/callbacks.fp16.html\">https://docs.fast.ai/callbacks.fp16.html</a>).\nIt works great for resnet, but I couldn't apply it to SENet.\nDid someone success to manage it??</p>",
          "rawMarkdown": "I've tried for my pytorch model but loss didn't converge somehow...\nAnd I also tried fastai's very simple callbacks(https://docs.fast.ai/callbacks.fp16.html).\nIt works great for resnet, but I couldn't apply it to SENet.\nDid someone success to manage it??",
          "votes": 2
        },
        {
          "id": 537058,
          "postDate": "2019-05-26T05:21:05.143Z",
          "content": "<p>Hi, <a href=\"/bamps53\">@bamps53</a> ，what loss function do you use in fastai?</p>",
          "rawMarkdown": "Hi, @bamps53 ，what loss function do you use in fastai?"
        }
      ]
    },
    {
      "id": 531943,
      "postDate": "2019-05-15T21:44:23.560Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 521223,
      "postDate": "2019-04-22T15:17:56.813Z",
      "content": "<p>Thank you for sharing !</p>",
      "rawMarkdown": "Thank you for sharing !",
      "votes": 1
    },
    {
      "id": 536627,
      "postDate": "2019-05-24T21:22:00.440Z",
      "content": "<p>Thank you!</p>",
      "rawMarkdown": "Thank you!"
    },
    {
      "id": 530180,
      "postDate": "2019-05-12T02:19:54.107Z",
      "content": "<p>Thanks for sharing, seems interesting.</p>",
      "rawMarkdown": "Thanks for sharing, seems interesting."
    }
  ],
  "comments": [
    {
      "id": 537428,
      "author_name": "Eric R",
      "author_url": "",
      "post_date": "2019-05-27T02:35:49.017000",
      "content": "<p>Interesting article, the implementations are done in MXNet, which was invented by Mu Li.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 523459,
      "author_name": "xutao",
      "author_url": "",
      "post_date": "2019-04-26T10:40:14.520000",
      "content": "<p>Thx for sharing, methods of the paper turely illuminating</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1699969,
      "author_name": "Paulo Junqueira",
      "author_url": "",
      "post_date": "2022-02-21T15:10:51.317000",
      "content": "<p>Interesting paper!, Do you know if is there any notebook in Kaggle implementing one of these ideias</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 956845,
      "author_name": "batu bayraktar",
      "author_url": "",
      "post_date": "2020-08-03T20:43:47.373000",
      "content": "<p>Hello Guys. I choose ResNet for İmage Classification. I thınk you should check thıs code\n<a href=\"https://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning\">https://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 531592,
      "author_name": "Gugu",
      "author_url": "",
      "post_date": "2019-05-15T07:19:22.997000",
      "content": "<p>Hello，have you tried to Low-precision training?how to code?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 531728,
          "author_name": "seefun",
          "author_url": "",
          "post_date": "2019-05-15T12:39:01.550000",
          "content": "<p>I have not try it. You can use Mixed Precision Training.  <a href=\"https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html\">https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html</a></p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 531736,
          "author_name": "Gugu",
          "author_url": "",
          "post_date": "2019-05-15T12:50:41.880000",
          "content": "<p>Thanks!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 531746,
          "author_name": "Camaro",
          "author_url": "",
          "post_date": "2019-05-15T13:19:33.047000",
          "content": "<p>I've tried for my pytorch model but loss didn't converge somehow...\nAnd I also tried fastai's very simple callbacks(<a href=\"https://docs.fast.ai/callbacks.fp16.html\">https://docs.fast.ai/callbacks.fp16.html</a>).\nIt works great for resnet, but I couldn't apply it to SENet.\nDid someone success to manage it??</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 537058,
          "author_name": "GUTAR",
          "author_url": "",
          "post_date": "2019-05-26T05:21:05.143000",
          "content": "<p>Hi, <a href=\"/bamps53\">@bamps53</a> ，what loss function do you use in fastai?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 531943,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-05-15T21:44:23.560000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 521223,
      "author_name": "Tawara",
      "author_url": "",
      "post_date": "2019-04-22T15:17:56.813000",
      "content": "<p>Thank you for sharing !</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 536627,
      "author_name": "Zhanseri Ikram",
      "author_url": "",
      "post_date": "2019-05-24T21:22:00.440000",
      "content": "<p>Thank you!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 530180,
      "author_name": "DimitreOliveira",
      "author_url": "",
      "post_date": "2019-05-12T02:19:54.107000",
      "content": "<p>Thanks for sharing, seems interesting.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "520961": "This is an excellent paper about training tricks in image classification. [Bag of Tricks for Image Classification with Convolutional Neural Networks](https://arxiv.org/abs/1812.01187)\n\nThere are many tricks include: Cosine Learning Rate Decay, Label Smoothing, Knowledge Distillation, Mixup Training, Learning rate warmup, No bias decay in BN, Zero gamma in ResNet block and so on. Very intersting paper from Mu Li et. al.",
    "537428": "Interesting article, the implementations are done in MXNet, which was invented by Mu Li.",
    "523459": "Thx for sharing, methods of the paper turely illuminating",
    "1699969": "Interesting paper!, Do you know if is there any notebook in Kaggle implementing one of these ideias",
    "956845": "Hello Guys. I choose ResNet for İmage Classification. I thınk you should check thıs code\nhttps://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning",
    "531592": "Hello，have you tried to Low-precision training?how to code?",
    "531943": "",
    "521223": "Thank you for sharing !",
    "536627": "Thank you!",
    "530180": "Thanks for sharing, seems interesting."
  }
}