{
  "id": 67986,
  "title": "RNN/LSTM architecture?",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/67986",
  "author_name": "Brian Lee",
  "post_date": "2018-10-08T05:46:22.157000",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hello Kagglers,</p>\n\n<p>I was wondering if there are certain architectures I could experiment or derive from for RNN(LSTM) models. I'm currently using forked version of a kernel(<a href=\"https://www.kaggle.com/kmader/quickdraw-baseline-lstm-reading-and-submission\">link</a>) with BN and was wondering what others had in mind.</p>",
  "messages": [
    {
      "id": 402124,
      "postDate": "2018-10-11T07:48:32.200Z",
      "content": "<p>try this:</p>\n\n<ol>\n<li><p>start with a simple lstm network. Make sure there is no bug! Print out training and validation metrics during training iterations.</p></li>\n<li><p>increase parameters (e.g. number of layers, filters number/filter size, etc...). Do until training metrics is better than validation metrics (i.e. over fitting)</p></li>\n<li><p>Do regularization, e.g. dropout or reduce parameters</p></li>\n</ol>",
      "rawMarkdown": "try this:\n\n1. start with a simple lstm network. Make sure there is no bug! Print out training and validation metrics during training iterations.\n\n2. increase parameters (e.g. number of layers, filters number/filter size, etc...). Do until training metrics is better than validation metrics (i.e. over fitting)\n\n3. Do regularization, e.g. dropout or reduce parameters\n\n\n",
      "votes": 7,
      "replies": [
        {
          "id": 402605,
          "postDate": "2018-10-12T01:41:54.063Z",
          "content": "<p>Thanks Heng! I'll try them out.</p>",
          "rawMarkdown": "Thanks Heng! I'll try them out."
        }
      ]
    },
    {
      "id": 401304,
      "postDate": "2018-10-09T19:48:32.560Z",
      "content": "<p>Hi JoonHo,</p>\n\n<p>I think your current forked kernel is with LSTM. </p>",
      "rawMarkdown": "Hi JoonHo,\n\nI think your current forked kernel is with LSTM. ",
      "votes": 1,
      "replies": [
        {
          "id": 401487,
          "postDate": "2018-10-10T06:45:04.730Z",
          "content": "<p>Hi,</p>\n\n<p>Yes I understand that. I guess my question is whether there are certain architectures that utilizes LSTM? Hope that rephrases my question better. </p>",
          "rawMarkdown": "Hi,\n\nYes I understand that. I guess my question is whether there are certain architectures that utilizes LSTM? Hope that rephrases my question better. ",
          "votes": 1
        }
      ]
    },
    {
      "id": 400325,
      "postDate": "2018-10-08T05:46:22.157Z",
      "content": "<p>Hello Kagglers,</p>\n\n<p>I was wondering if there are certain architectures I could experiment or derive from for RNN(LSTM) models. I'm currently using forked version of a kernel(<a href=\"https://www.kaggle.com/kmader/quickdraw-baseline-lstm-reading-and-submission\">link</a>) with BN and was wondering what others had in mind.</p>",
      "rawMarkdown": "Hello Kagglers,\n\nI was wondering if there are certain architectures I could experiment or derive from for RNN(LSTM) models. I'm currently using forked version of a kernel([link][1]) with BN and was wondering what others had in mind.\n\n\n  [1]: https://www.kaggle.com/kmader/quickdraw-baseline-lstm-reading-and-submission",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 402124,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2018-10-11T07:48:32.200000",
      "content": "<p>try this:</p>\n\n<ol>\n<li><p>start with a simple lstm network. Make sure there is no bug! Print out training and validation metrics during training iterations.</p></li>\n<li><p>increase parameters (e.g. number of layers, filters number/filter size, etc...). Do until training metrics is better than validation metrics (i.e. over fitting)</p></li>\n<li><p>Do regularization, e.g. dropout or reduce parameters</p></li>\n</ol>",
      "votes": 7,
      "replies": [
        {
          "id": 402605,
          "author_name": "Brian Lee",
          "author_url": "",
          "post_date": "2018-10-12T01:41:54.063000",
          "content": "<p>Thanks Heng! I'll try them out.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 401304,
      "author_name": "Yang Shi",
      "author_url": "",
      "post_date": "2018-10-09T19:48:32.560000",
      "content": "<p>Hi JoonHo,</p>\n\n<p>I think your current forked kernel is with LSTM. </p>",
      "votes": 1,
      "replies": [
        {
          "id": 401487,
          "author_name": "Brian Lee",
          "author_url": "",
          "post_date": "2018-10-10T06:45:04.730000",
          "content": "<p>Hi,</p>\n\n<p>Yes I understand that. I guess my question is whether there are certain architectures that utilizes LSTM? Hope that rephrases my question better. </p>",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "402124": "try this:\n\n1. start with a simple lstm network. Make sure there is no bug! Print out training and validation metrics during training iterations.\n\n2. increase parameters (e.g. number of layers, filters number/filter size, etc...). Do until training metrics is better than validation metrics (i.e. over fitting)\n\n3. Do regularization, e.g. dropout or reduce parameters\n\n\n",
    "401304": "Hi JoonHo,\n\nI think your current forked kernel is with LSTM. ",
    "400325": "Hello Kagglers,\n\nI was wondering if there are certain architectures I could experiment or derive from for RNN(LSTM) models. I'm currently using forked version of a kernel([link][1]) with BN and was wondering what others had in mind.\n\n\n  [1]: https://www.kaggle.com/kmader/quickdraw-baseline-lstm-reading-and-submission"
  }
}