{
  "id": 22464,
  "title": "fine-tuning Caffe-net [~45 min] [L.B. 0.72636]",
  "url": "/competitions/state-farm-distracted-driver-detection/discussion/22464",
  "author_name": "alireza",
  "post_date": "2016-07-24T21:09:26.847000",
  "votes": 9,
  "comment_count": 0,
  "views": 457,
  "content": "<p>Since this is my first time fine-tuning a deep net, I have decided to use Caffe. It has elaborate examples and lets you focus on changing the parameters of the Solver.</p>\n\n<p>I drove my solution from the fine-tuning example in Caffe github repository and wrapped it in a docker container.</p>\n\n<p>With the current parameters, the model achieved 0.72636 score on the leader board. It took only 15 min to train the caffe-net and about 30 min to process the test-set on an AWS g2.2xlarge instance. The fast running time let me experiment with the parameters of the solver. However, I have not been able to get a score close to 0.2 which many on the leader board currently have. With final exams approaching and my bank account getting major hits by AWS, I am going to leave this here, hoping more people will experiment with the parameters.</p>",
  "messages": [
    {
      "id": 128856,
      "postDate": "2016-07-24T21:09:26.847Z",
      "content": "<p>Since this is my first time fine-tuning a deep net, I have decided to use Caffe. It has elaborate examples and lets you focus on changing the parameters of the Solver.</p>\n\n<p>I drove my solution from the fine-tuning example in Caffe github repository and wrapped it in a docker container.</p>\n\n<p>With the current parameters, the model achieved 0.72636 score on the leader board. It took only 15 min to train the caffe-net and about 30 min to process the test-set on an AWS g2.2xlarge instance. The fast running time let me experiment with the parameters of the solver. However, I have not been able to get a score close to 0.2 which many on the leader board currently have. With final exams approaching and my bank account getting major hits by AWS, I am going to leave this here, hoping more people will experiment with the parameters.</p>",
      "rawMarkdown": "Since this is my first time fine-tuning a deep net, I have decided to use Caffe. It has elaborate examples and lets you focus on changing the parameters of the Solver.\n\nI drove my solution from the fine-tuning example in Caffe github repository and wrapped it in a docker container.\n\nWith the current parameters, the model achieved 0.72636 score on the leader board. It took only 15 min to train the caffe-net and about 30 min to process the test-set on an AWS g2.2xlarge instance. The fast running time let me experiment with the parameters of the solver. However, I have not been able to get a score close to 0.2 which many on the leader board currently have. With final exams approaching and my bank account getting major hits by AWS, I am going to leave this here, hoping more people will experiment with the parameters.",
      "votes": 9
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "128856": "Since this is my first time fine-tuning a deep net, I have decided to use Caffe. It has elaborate examples and lets you focus on changing the parameters of the Solver.\n\nI drove my solution from the fine-tuning example in Caffe github repository and wrapped it in a docker container.\n\nWith the current parameters, the model achieved 0.72636 score on the leader board. It took only 15 min to train the caffe-net and about 30 min to process the test-set on an AWS g2.2xlarge instance. The fast running time let me experiment with the parameters of the solver. However, I have not been able to get a score close to 0.2 which many on the leader board currently have. With final exams approaching and my bank account getting major hits by AWS, I am going to leave this here, hoping more people will experiment with the parameters."
  }
}