{
  "id": 20990,
  "title": "Implementing Keras (sample) script in tensorflow ",
  "url": "/competitions/state-farm-distracted-driver-detection/discussion/20990",
  "author_name": "Asymptote",
  "post_date": "2016-05-16T08:18:16.213000",
  "votes": 0,
  "comment_count": 0,
  "views": 661,
  "content": "<p>Hi\nI have tried Keras (sample) script with tensorflow backend (with out cv).\nI recently completed the Udacity course on tensorflow and I wanted to replicate results (of keras script) in tensorflow entirely. I am not getting the same results. My validation loss isn't decreasing past 2.2. I have tried to make the same network, I am not sure what am I missing.</p>\n\n<ul>\n<li>dropouts make it worse</li>\n<li>default learning rate of adadelta in keras is 1.0 while in tensorflow it is 0.01</li>\n</ul>\n\n<p>Here are the two scripts.</p>\n\n<p>(Keras)\n<a href=\"https://github.com/Asymptote/distracted-drivers-kaggle/blob/master/kr_1_simple.py\">https://github.com/Asymptote/distracted-drivers-kaggle/blob/master/kr_1_simple.py</a></p>\n\n<p>(tensorflow)\n<a href=\"https://github.com/Asymptote/distracted-drivers-kaggle/blob/master/1_tf_v1.ipynb\">https://github.com/Asymptote/distracted-drivers-kaggle/blob/master/1_tf_v1.ipynb</a></p>\n\n<p>This is just for learning purpose, I wanted to make sure I am understanding what keras script is doing. If someone can skim through the two models that would be great. I will keep exploring, will update this thread.</p>",
  "messages": [
    {
      "id": 120189,
      "postDate": "2016-05-16T08:18:16.213Z",
      "content": "<p>Hi\nI have tried Keras (sample) script with tensorflow backend (with out cv).\nI recently completed the Udacity course on tensorflow and I wanted to replicate results (of keras script) in tensorflow entirely. I am not getting the same results. My validation loss isn't decreasing past 2.2. I have tried to make the same network, I am not sure what am I missing.</p>\n\n<ul>\n<li>dropouts make it worse</li>\n<li>default learning rate of adadelta in keras is 1.0 while in tensorflow it is 0.01</li>\n</ul>\n\n<p>Here are the two scripts.</p>\n\n<p>(Keras)\n<a href=\"https://github.com/Asymptote/distracted-drivers-kaggle/blob/master/kr_1_simple.py\">https://github.com/Asymptote/distracted-drivers-kaggle/blob/master/kr_1_simple.py</a></p>\n\n<p>(tensorflow)\n<a href=\"https://github.com/Asymptote/distracted-drivers-kaggle/blob/master/1_tf_v1.ipynb\">https://github.com/Asymptote/distracted-drivers-kaggle/blob/master/1_tf_v1.ipynb</a></p>\n\n<p>This is just for learning purpose, I wanted to make sure I am understanding what keras script is doing. If someone can skim through the two models that would be great. I will keep exploring, will update this thread.</p>",
      "rawMarkdown": "Hi\r\nI have tried Keras (sample) script with tensorflow backend (with out cv).\r\nI recently completed the Udacity course on tensorflow and I wanted to replicate results (of keras script) in tensorflow entirely. I am not getting the same results. My validation loss isn't decreasing past 2.2. I have tried to make the same network, I am not sure what am I missing.\r\n\r\n - dropouts make it worse\r\n - default learning rate of adadelta in keras is 1.0 while in tensorflow it is 0.01\r\n\r\nHere are the two scripts.\r\n\r\n(Keras)\r\nhttps://github.com/Asymptote/distracted-drivers-kaggle/blob/master/kr_1_simple.py\r\n\r\n(tensorflow)\r\nhttps://github.com/Asymptote/distracted-drivers-kaggle/blob/master/1_tf_v1.ipynb\r\n\r\nThis is just for learning purpose, I wanted to make sure I am understanding what keras script is doing. If someone can skim through the two models that would be great. I will keep exploring, will update this thread."
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "120189": "Hi\r\nI have tried Keras (sample) script with tensorflow backend (with out cv).\r\nI recently completed the Udacity course on tensorflow and I wanted to replicate results (of keras script) in tensorflow entirely. I am not getting the same results. My validation loss isn't decreasing past 2.2. I have tried to make the same network, I am not sure what am I missing.\r\n\r\n - dropouts make it worse\r\n - default learning rate of adadelta in keras is 1.0 while in tensorflow it is 0.01\r\n\r\nHere are the two scripts.\r\n\r\n(Keras)\r\nhttps://github.com/Asymptote/distracted-drivers-kaggle/blob/master/kr_1_simple.py\r\n\r\n(tensorflow)\r\nhttps://github.com/Asymptote/distracted-drivers-kaggle/blob/master/1_tf_v1.ipynb\r\n\r\nThis is just for learning purpose, I wanted to make sure I am understanding what keras script is doing. If someone can skim through the two models that would be great. I will keep exploring, will update this thread."
  }
}