{
  "id": 38332,
  "title": "92% accuracy on Cifar-10 using Keras",
  "url": "/competitions/cifar-10/discussion/38332",
  "author_name": "",
  "post_date": "2017-08-19T18:21:54.613973500Z",
  "votes": 6,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hello,</p>\n\n<p>It is my first post in Kaggle!</p>\n\n<p>I tried to implement <a href=\"https://arxiv.org/pdf/1605.07146.pdf\">this paper</a> in Keras. According to the paper, one should be able to achieve accuracy of 96% which would be state of the art result for Cifar-10 data set. \nI tried to implement this in Keras but best I have achieved is 92%. </p>\n\n<p>More experienced users (and starting users) could help figure out why this Keras code does not produce similar results shown in the paper. With the paper they also released a pytorch code available <a href=\"https://github.com/szagoruyko/wide-residual-networks/tree/master/pytorch\">here</a>.</p>\n\n<p><a href=\"https://github.com/olympus999/cifar-10-wrn\">Github link</a> for the Keras code.</p>",
  "messages": [
    {
      "id": "215091",
      "postDate": "08/19/2017 18:21:54",
      "content": "<p>Hello,</p>\n\n<p>It is my first post in Kaggle!</p>\n\n<p>I tried to implement <a href=\"https://arxiv.org/pdf/1605.07146.pdf\">this paper</a> in Keras. According to the paper, one should be able to achieve accuracy of 96% which would be state of the art result for Cifar-10 data set. \nI tried to implement this in Keras but best I have achieved is 92%. </p>\n\n<p>More experienced users (and starting users) could help figure out why this Keras code does not produce similar results shown in the paper. With the paper they also released a pytorch code available <a href=\"https://github.com/szagoruyko/wide-residual-networks/tree/master/pytorch\">here</a>.</p>\n\n<p><a href=\"https://github.com/olympus999/cifar-10-wrn\">Github link</a> for the Keras code.</p>",
      "rawMarkdown": "Hello,\n\nIt is my first post in Kaggle!\n\nI tried to implement [this paper][1] in Keras. According to the paper, one should be able to achieve accuracy of 96% which would be state of the art result for Cifar-10 data set. \nI tried to implement this in Keras but best I have achieved is 92%. \n\nMore experienced users (and starting users) could help figure out why this Keras code does not produce similar results shown in the paper. With the paper they also released a pytorch code available [here][2].\n\n[Github link][3] for the Keras code.\n\n\n  [1]: https://arxiv.org/pdf/1605.07146.pdf\n  [2]: https://github.com/szagoruyko/wide-residual-networks/tree/master/pytorch\n  [3]: https://github.com/olympus999/cifar-10-wrn",
      "votes": null
    },
    {
      "id": "247364",
      "postDate": "11/22/2017 22:30:48",
      "content": "<p>Hmmm, have you managed to improve the validation accuracy? It seems an overfitting problem so I would try regularization or dropout to address the issue... By the way... this model is so huge that I can't run it on my computer (2gb ddr5 graphics)... Where are you training it?</p>",
      "rawMarkdown": "Hmmm, have you managed to improve the validation accuracy? It seems an overfitting problem so I would try regularization or dropout to address the issue... By the way... this model is so huge that I can't run it on my computer (2gb ddr5 graphics)... Where are you training it?",
      "votes": null
    },
    {
      "id": "247545",
      "postDate": "11/23/2017 10:23:19",
      "content": "<p>I have not dealt with it since.</p>\n\n<p>But I did notice that the pytorch version of the code has been updated (previously missing library has been added). I will try to rerun pytorch code later.</p>\n\n<p>I ran the code on 1070 8gb ram. Computer vision learning tends to take quite a bit of video memory.</p>",
      "rawMarkdown": "I have not dealt with it since.\n\nBut I did notice that the pytorch version of the code has been updated (previously missing library has been added). I will try to rerun pytorch code later.\n\nI ran the code on 1070 8gb ram. Computer vision learning tends to take quite a bit of video memory.",
      "votes": null
    },
    {
      "id": "252385",
      "postDate": "12/02/2017 20:31:47",
      "content": "<p>Tried WRN-16-4 and 16-8. Got 91.x for both of them. Also run DenseNets and NASNets, still trying to get above 92.5 with something in Keras. I don't know why but it seems impossible for me. I'll try some pytorch, mxnet or caffe implementations.</p>",
      "rawMarkdown": "Tried WRN-16-4 and 16-8. Got 91.x for both of them. Also run DenseNets and NASNets, still trying to get above 92.5 with something in Keras. I don't know why but it seems impossible for me. I'll try some pytorch, mxnet or caffe implementations.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 247364,
      "author_name": "ericalcaide9834",
      "author_url": "",
      "post_date": "11/22/2017 22:30:48",
      "content": "<p>Hmmm, have you managed to improve the validation accuracy? It seems an overfitting problem so I would try regularization or dropout to address the issue... By the way... this model is so huge that I can't run it on my computer (2gb ddr5 graphics)... Where are you training it?</p>",
      "votes": null,
      "replies": [
        {
          "id": 247545,
          "author_name": "hazzle",
          "author_url": "",
          "post_date": "11/23/2017 10:23:19",
          "content": "<p>I have not dealt with it since.</p>\n\n<p>But I did notice that the pytorch version of the code has been updated (previously missing library has been added). I will try to rerun pytorch code later.</p>\n\n<p>I ran the code on 1070 8gb ram. Computer vision learning tends to take quite a bit of video memory.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 252385,
      "author_name": "ericalcaide9834",
      "author_url": "",
      "post_date": "12/02/2017 20:31:47",
      "content": "<p>Tried WRN-16-4 and 16-8. Got 91.x for both of them. Also run DenseNets and NASNets, still trying to get above 92.5 with something in Keras. I don't know why but it seems impossible for me. I'll try some pytorch, mxnet or caffe implementations.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "215091": "Hello,\n\nIt is my first post in Kaggle!\n\nI tried to implement [this paper][1] in Keras. According to the paper, one should be able to achieve accuracy of 96% which would be state of the art result for Cifar-10 data set. \nI tried to implement this in Keras but best I have achieved is 92%. \n\nMore experienced users (and starting users) could help figure out why this Keras code does not produce similar results shown in the paper. With the paper they also released a pytorch code available [here][2].\n\n[Github link][3] for the Keras code.\n\n\n  [1]: https://arxiv.org/pdf/1605.07146.pdf\n  [2]: https://github.com/szagoruyko/wide-residual-networks/tree/master/pytorch\n  [3]: https://github.com/olympus999/cifar-10-wrn",
    "247364": "Hmmm, have you managed to improve the validation accuracy? It seems an overfitting problem so I would try regularization or dropout to address the issue... By the way... this model is so huge that I can't run it on my computer (2gb ddr5 graphics)... Where are you training it?",
    "247545": "I have not dealt with it since.\n\nBut I did notice that the pytorch version of the code has been updated (previously missing library has been added). I will try to rerun pytorch code later.\n\nI ran the code on 1070 8gb ram. Computer vision learning tends to take quite a bit of video memory.",
    "252385": "Tried WRN-16-4 and 16-8. Got 91.x for both of them. Also run DenseNets and NASNets, still trying to get above 92.5 with something in Keras. I don't know why but it seems impossible for me. I'll try some pytorch, mxnet or caffe implementations."
  },
  "source": "meta"
}