{
  "id": 46183,
  "title": "Question about the impact of Dropouut techinque.",
  "url": "/competitions/dog-breed-identification/discussion/46183",
  "author_name": "",
  "post_date": "2017-12-22T02:28:53.174550900Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>This are my accuracy that i implemented in state of the art models, but i have a question, In your view why my accuracy of model is reduced when i put the dropout in the models?</p>\n\n<p><strong>link of the table:</strong>\n<a href=\"https://drive.google.com/file/d/1rz_ZUSaDCeRPF7_zSIjUD3gqQRWCU5KA/view?usp=sharing\">https://drive.google.com/file/d/1rz_ZUSaDCeRPF7_zSIjUD3gqQRWCU5KA/view?usp=sharing</a></p>",
  "messages": [
    {
      "id": "261218",
      "postDate": "12/22/2017 02:28:53",
      "content": "<p>This are my accuracy that i implemented in state of the art models, but i have a question, In your view why my accuracy of model is reduced when i put the dropout in the models?</p>\n\n<p><strong>link of the table:</strong>\n<a href=\"https://drive.google.com/file/d/1rz_ZUSaDCeRPF7_zSIjUD3gqQRWCU5KA/view?usp=sharing\">https://drive.google.com/file/d/1rz_ZUSaDCeRPF7_zSIjUD3gqQRWCU5KA/view?usp=sharing</a></p>",
      "rawMarkdown": "This are my accuracy that i implemented in state of the art models, but i have a question, In your view why my accuracy of model is reduced when i put the dropout in the models?\n\n**link of the table:**\nhttps://drive.google.com/file/d/1rz_ZUSaDCeRPF7_zSIjUD3gqQRWCU5KA/view?usp=sharing",
      "votes": null
    },
    {
      "id": "261230",
      "postDate": "12/22/2017 03:09:22",
      "content": "<p>Actually, maybe you determine the probability of the block to the lower than 0.5, but all in all it is very unbelievable.</p>",
      "rawMarkdown": "Actually, maybe you determine the probability of the block to the lower than 0.5, but all in all it is very unbelievable.",
      "votes": null
    },
    {
      "id": "277427",
      "postDate": "02/03/2018 04:32:41",
      "content": "<p>Dropout can be tricky and should be deployed carefully. It is also important to understand where and when to use it. Without directly seeing your model it is hard to understand what your issues might be. My general rules before tuning:\n(1) FC layers are around 0.5\n(2) CNN layers start around 0.2\n(3) Do not mix batch normalization and dropout\n(4) Also, only 1 dropout per pooling layer; this means you could potentially have several ConvNet layers before a dropout.\nOnce you start with that, you can slowly tune your model with respect to dropout rates. This is usually my STARTING position - sometimes models need more help than others, don't be afraid to play.</p>",
      "rawMarkdown": "Dropout can be tricky and should be deployed carefully. It is also important to understand where and when to use it. Without directly seeing your model it is hard to understand what your issues might be. My general rules before tuning:\n(1) FC layers are around 0.5\n(2) CNN layers start around 0.2\n(3) Do not mix batch normalization and dropout\n(4) Also, only 1 dropout per pooling layer; this means you could potentially have several ConvNet layers before a dropout.\nOnce you start with that, you can slowly tune your model with respect to dropout rates. This is usually my STARTING position - sometimes models need more help than others, don't be afraid to play.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 261230,
      "author_name": "beyoghlu",
      "author_url": "",
      "post_date": "12/22/2017 03:09:22",
      "content": "<p>Actually, maybe you determine the probability of the block to the lower than 0.5, but all in all it is very unbelievable.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 277427,
      "author_name": "cbryant",
      "author_url": "",
      "post_date": "02/03/2018 04:32:41",
      "content": "<p>Dropout can be tricky and should be deployed carefully. It is also important to understand where and when to use it. Without directly seeing your model it is hard to understand what your issues might be. My general rules before tuning:\n(1) FC layers are around 0.5\n(2) CNN layers start around 0.2\n(3) Do not mix batch normalization and dropout\n(4) Also, only 1 dropout per pooling layer; this means you could potentially have several ConvNet layers before a dropout.\nOnce you start with that, you can slowly tune your model with respect to dropout rates. This is usually my STARTING position - sometimes models need more help than others, don't be afraid to play.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "261218": "This are my accuracy that i implemented in state of the art models, but i have a question, In your view why my accuracy of model is reduced when i put the dropout in the models?\n\n**link of the table:**\nhttps://drive.google.com/file/d/1rz_ZUSaDCeRPF7_zSIjUD3gqQRWCU5KA/view?usp=sharing",
    "261230": "Actually, maybe you determine the probability of the block to the lower than 0.5, but all in all it is very unbelievable.",
    "277427": "Dropout can be tricky and should be deployed carefully. It is also important to understand where and when to use it. Without directly seeing your model it is hard to understand what your issues might be. My general rules before tuning:\n(1) FC layers are around 0.5\n(2) CNN layers start around 0.2\n(3) Do not mix batch normalization and dropout\n(4) Also, only 1 dropout per pooling layer; this means you could potentially have several ConvNet layers before a dropout.\nOnce you start with that, you can slowly tune your model with respect to dropout rates. This is usually my STARTING position - sometimes models need more help than others, don't be afraid to play."
  },
  "source": "meta"
}