{
  "id": 132894,
  "title": "[Keras Users] EfficientNet Noisy Student weights released",
  "url": "/competitions/bengaliai-cv19/discussion/132894",
  "author_name": "Pavel Iakubovskii (qubvel)",
  "post_date": "2020-02-28T13:34:30.517000",
  "votes": 89,
  "comment_count": 15,
  "views": 0,
  "content": "<p>Hi everyone, in case you are using <code>efficientnet</code> (<a href=\"https://github.com/qubvel/efficientnet\">https://github.com/qubvel/efficientnet</a>) pypi package you can now use not only <code>imagenet</code> weights option but <code>noisy-student</code> weights also for all B0-B7 models:\n<code>\nimport efficientnet.keras as eff\nmodel = eff.EfficientNetB0(weights='noisy-student')\n</code></p>\n\n<p>reed more about noisy student: \n<a href=\"https://arxiv.org/abs/1911.04252\">https://arxiv.org/abs/1911.04252</a>\n<a href=\"https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet\">https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet</a></p>",
  "messages": [
    {
      "id": 759033,
      "postDate": "2020-02-28T13:34:30.517Z",
      "content": "<p>Hi everyone, in case you are using <code>efficientnet</code> (<a href=\"https://github.com/qubvel/efficientnet\">https://github.com/qubvel/efficientnet</a>) pypi package you can now use not only <code>imagenet</code> weights option but <code>noisy-student</code> weights also for all B0-B7 models:\n<code>\nimport efficientnet.keras as eff\nmodel = eff.EfficientNetB0(weights='noisy-student')\n</code></p>\n\n<p>reed more about noisy student: \n<a href=\"https://arxiv.org/abs/1911.04252\">https://arxiv.org/abs/1911.04252</a>\n<a href=\"https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet\">https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet</a></p>",
      "rawMarkdown": "Hi everyone, in case you are using `efficientnet` (https://github.com/qubvel/efficientnet) pypi package you can now use not only `imagenet` weights option but `noisy-student` weights also for all B0-B7 models:\n```\nimport efficientnet.keras as eff\nmodel = eff.EfficientNetB0(weights='noisy-student')\n```\n\nreed more about noisy student: \nhttps://arxiv.org/abs/1911.04252\nhttps://github.com/tensorflow/tpu/tree/master/models/official/efficientnet",
      "votes": 89
    },
    {
      "id": 759285,
      "postDate": "2020-02-28T20:56:17.883Z",
      "content": "<p>Thanks qubvel ( <a href=\"/pavel92\">@pavel92</a> ) for letting us know this. And thanks for all your fantastic GitHub repositories: efficientnet, classification models, segmentation models, and tta_wrapper. I use them all the time for comps.</p>",
      "rawMarkdown": "Thanks qubvel ( @pavel92 ) for letting us know this. And thanks for all your fantastic GitHub repositories: efficientnet, classification models, segmentation models, and tta_wrapper. I use them all the time for comps.",
      "votes": 8
    },
    {
      "id": 1130738,
      "postDate": "2020-12-29T09:36:43.957Z",
      "content": "<p>Awesome work, I noticed the noisy students weights only fit your implementation of efficientnet (it has -1 layers comparing to the tf implementation)<br>\nCan you point at which layer is missing?<br>\nThe implementation of tf has a normalization layer inside it (it says so in the docs), does your implementation has it there too? can you provide some details about it? (since I havent found any by searching the net)<br>\nThank you so much for your work!</p>",
      "rawMarkdown": "Awesome work, I noticed the noisy students weights only fit your implementation of efficientnet (it has -1 layers comparing to the tf implementation)\nCan you point at which layer is missing?\nThe implementation of tf has a normalization layer inside it (it says so in the docs), does your implementation has it there too? can you provide some details about it? (since I havent found any by searching the net)\nThank you so much for your work!",
      "votes": 1,
      "replies": [
        {
          "id": 1140287,
          "postDate": "2021-01-05T22:57:00.663Z",
          "content": "<p>I had the same issue, could not find out why this happened but you can fix this issue if you create the model kind of like this.</p>\n<p><code>model = models.Sequential()</code></p>\n<p><code>model.add(efn.EfficientNetB0(include_top = False, weights = None)</code></p>",
          "rawMarkdown": "I had the same issue, could not find out why this happened but you can fix this issue if you create the model kind of like this.\n\n\n\n`model = models.Sequential()`\n    \n`model.add(efn.EfficientNetB0(include_top = False, weights = None)`\n\n",
          "votes": 1
        }
      ]
    },
    {
      "id": 760205,
      "postDate": "2020-03-01T00:57:25.343Z",
      "content": "<p>with proper credits:\n<a href=\"https://www.kaggle.com/ipythonx/efficientnet-keras-noisystudent-weights-b0b7\">https://www.kaggle.com/ipythonx/efficientnet-keras-noisystudent-weights-b0b7</a>\nthanks a lot <a href=\"/pavel92\">@pavel92</a> </p>",
      "rawMarkdown": "with proper credits:\nhttps://www.kaggle.com/ipythonx/efficientnet-keras-noisystudent-weights-b0b7\nthanks a lot @pavel92 ",
      "votes": 3,
      "replies": [
        {
          "id": 830478,
          "postDate": "2020-05-02T16:13:09.717Z",
          "rawMarkdown": "",
          "votes": 1,
          "isDeleted": true
        }
      ]
    },
    {
      "id": 863421,
      "postDate": "2020-05-27T10:06:06.797Z",
      "content": "<p>Can someone please share the json files containing the architectures of efficiNet models between b0 and b7?</p>",
      "rawMarkdown": "Can someone please share the json files containing the architectures of efficiNet models between b0 and b7?"
    },
    {
      "id": 861527,
      "postDate": "2020-05-26T05:05:53.850Z",
      "content": "<p>nice</p>",
      "rawMarkdown": "nice\n"
    },
    {
      "id": 780225,
      "postDate": "2020-03-20T04:39:08.853Z",
      "content": "<p><a href=\"/pavel92\">@pavel92</a>: please include these EfficientDetection models ( <a href=\"https://arxiv.org/pdf/1911.09070.pdf\">https://arxiv.org/pdf/1911.09070.pdf</a>) in your library. I'm a fan of your repository. Thanks!</p>",
      "rawMarkdown": "@pavel92: please include these EfficientDetection models ( https://arxiv.org/pdf/1911.09070.pdf) in your library. I'm a fan of your repository. Thanks!"
    },
    {
      "id": 760726,
      "postDate": "2020-03-01T16:53:59.277Z",
      "content": "<p>Thanks @qubvel Very big fan of your packages...We will put them through some more very serious testing ;-)</p>\n\n<p>Keep up at the good work!</p>",
      "rawMarkdown": "Thanks @qubvel Very big fan of your packages...We will put them through some more very serious testing ;-)\n\nKeep up at the good work!"
    },
    {
      "id": 771422,
      "postDate": "2020-03-14T06:28:18.043Z",
      "rawMarkdown": "",
      "isDeleted": true
    },
    {
      "id": 759287,
      "postDate": "2020-02-28T21:03:26.633Z",
      "rawMarkdown": "",
      "votes": 2,
      "isDeleted": true,
      "replies": [
        {
          "id": 759480,
          "postDate": "2020-02-29T05:01:11.870Z",
          "content": "<p><a href=\"/pavel92\">@pavel92</a> </p>",
          "rawMarkdown": "@pavel92 "
        },
        {
          "id": 759496,
          "postDate": "2020-02-29T05:34:45.947Z",
          "content": "<p>Info <a href=\"https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/113195\">here</a>. If the answer isn't in the main post, read some of the comments before for more info.</p>",
          "rawMarkdown": "Info [here][1]. If the answer isn't in the main post, read some of the comments before for more info.\n\n[1]: https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/113195"
        }
      ]
    },
    {
      "id": 766601,
      "postDate": "2020-03-08T12:50:07.687Z",
      "content": "<p>Great!. \nThank you <a href=\"/pavel92\">@pavel92</a>!.</p>",
      "rawMarkdown": "Great!. \nThank you @pavel92!."
    },
    {
      "id": 759817,
      "postDate": "2020-02-29T13:50:27.707Z",
      "content": "<p>I'm fan of your repo. Thank you!</p>",
      "rawMarkdown": "I'm fan of your repo. Thank you!"
    }
  ],
  "comments": [
    {
      "id": 759285,
      "author_name": "Chris Deotte",
      "author_url": "",
      "post_date": "2020-02-28T20:56:17.883000",
      "content": "<p>Thanks qubvel ( <a href=\"/pavel92\">@pavel92</a> ) for letting us know this. And thanks for all your fantastic GitHub repositories: efficientnet, classification models, segmentation models, and tta_wrapper. I use them all the time for comps.</p>",
      "votes": 8,
      "replies": []
    },
    {
      "id": 1130738,
      "author_name": "Omer A.",
      "author_url": "",
      "post_date": "2020-12-29T09:36:43.957000",
      "content": "<p>Awesome work, I noticed the noisy students weights only fit your implementation of efficientnet (it has -1 layers comparing to the tf implementation)<br>\nCan you point at which layer is missing?<br>\nThe implementation of tf has a normalization layer inside it (it says so in the docs), does your implementation has it there too? can you provide some details about it? (since I havent found any by searching the net)<br>\nThank you so much for your work!</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1140287,
          "author_name": "Bartley",
          "author_url": "",
          "post_date": "2021-01-05T22:57:00.663000",
          "content": "<p>I had the same issue, could not find out why this happened but you can fix this issue if you create the model kind of like this.</p>\n<p><code>model = models.Sequential()</code></p>\n<p><code>model.add(efn.EfficientNetB0(include_top = False, weights = None)</code></p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 760205,
      "author_name": "Akash Shingha",
      "author_url": "",
      "post_date": "2020-03-01T00:57:25.343000",
      "content": "<p>with proper credits:\n<a href=\"https://www.kaggle.com/ipythonx/efficientnet-keras-noisystudent-weights-b0b7\">https://www.kaggle.com/ipythonx/efficientnet-keras-noisystudent-weights-b0b7</a>\nthanks a lot <a href=\"/pavel92\">@pavel92</a> </p>",
      "votes": 3,
      "replies": [
        {
          "id": 830478,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-05-02T16:13:09.717000",
          "content": "",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 863421,
      "author_name": "COŞKU ÖKSÜZ",
      "author_url": "",
      "post_date": "2020-05-27T10:06:06.797000",
      "content": "<p>Can someone please share the json files containing the architectures of efficiNet models between b0 and b7?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 861527,
      "author_name": "Richie Rich",
      "author_url": "",
      "post_date": "2020-05-26T05:05:53.850000",
      "content": "<p>nice</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 780225,
      "author_name": "FGPC",
      "author_url": "",
      "post_date": "2020-03-20T04:39:08.853000",
      "content": "<p><a href=\"/pavel92\">@pavel92</a>: please include these EfficientDetection models ( <a href=\"https://arxiv.org/pdf/1911.09070.pdf\">https://arxiv.org/pdf/1911.09070.pdf</a>) in your library. I'm a fan of your repository. Thanks!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 760726,
      "author_name": "Robin Smits",
      "author_url": "",
      "post_date": "2020-03-01T16:53:59.277000",
      "content": "<p>Thanks @qubvel Very big fan of your packages...We will put them through some more very serious testing ;-)</p>\n\n<p>Keep up at the good work!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 771422,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-03-14T06:28:18.043000",
      "content": "",
      "votes": 0,
      "replies": []
    },
    {
      "id": 759287,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-02-28T21:03:26.633000",
      "content": "",
      "votes": 2,
      "replies": [
        {
          "id": 759480,
          "author_name": "Akash Shingha",
          "author_url": "",
          "post_date": "2020-02-29T05:01:11.870000",
          "content": "<p><a href=\"/pavel92\">@pavel92</a> </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 759496,
          "author_name": "Chris Deotte",
          "author_url": "",
          "post_date": "2020-02-29T05:34:45.947000",
          "content": "<p>Info <a href=\"https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/113195\">here</a>. If the answer isn't in the main post, read some of the comments before for more info.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 766601,
      "author_name": "CASFRANCO",
      "author_url": "",
      "post_date": "2020-03-08T12:50:07.687000",
      "content": "<p>Great!. \nThank you <a href=\"/pavel92\">@pavel92</a>!.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 759817,
      "author_name": "Soonhwan Kwon",
      "author_url": "",
      "post_date": "2020-02-29T13:50:27.707000",
      "content": "<p>I'm fan of your repo. Thank you!</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "759033": "Hi everyone, in case you are using `efficientnet` (https://github.com/qubvel/efficientnet) pypi package you can now use not only `imagenet` weights option but `noisy-student` weights also for all B0-B7 models:\n```\nimport efficientnet.keras as eff\nmodel = eff.EfficientNetB0(weights='noisy-student')\n```\n\nreed more about noisy student: \nhttps://arxiv.org/abs/1911.04252\nhttps://github.com/tensorflow/tpu/tree/master/models/official/efficientnet",
    "759285": "Thanks qubvel ( @pavel92 ) for letting us know this. And thanks for all your fantastic GitHub repositories: efficientnet, classification models, segmentation models, and tta_wrapper. I use them all the time for comps.",
    "1130738": "Awesome work, I noticed the noisy students weights only fit your implementation of efficientnet (it has -1 layers comparing to the tf implementation)\nCan you point at which layer is missing?\nThe implementation of tf has a normalization layer inside it (it says so in the docs), does your implementation has it there too? can you provide some details about it? (since I havent found any by searching the net)\nThank you so much for your work!",
    "760205": "with proper credits:\nhttps://www.kaggle.com/ipythonx/efficientnet-keras-noisystudent-weights-b0b7\nthanks a lot @pavel92 ",
    "863421": "Can someone please share the json files containing the architectures of efficiNet models between b0 and b7?",
    "861527": "nice\n",
    "780225": "@pavel92: please include these EfficientDetection models ( https://arxiv.org/pdf/1911.09070.pdf) in your library. I'm a fan of your repository. Thanks!",
    "760726": "Thanks @qubvel Very big fan of your packages...We will put them through some more very serious testing ;-)\n\nKeep up at the good work!",
    "771422": "",
    "759287": "",
    "766601": "Great!. \nThank you @pavel92!.",
    "759817": "I'm fan of your repo. Thank you!"
  }
}