{
  "id": 94807,
  "title": "what's the best score your NN based model got?",
  "url": "/competitions/LANL-Earthquake-Prediction/discussion/94807",
  "author_name": "Z. Liu",
  "post_date": "2019-06-07T02:51:41.912000",
  "votes": 3,
  "comment_count": 7,
  "views": 0,
  "content": "<p>I am curious how good a NN (e.g., CNN, RNN, Dense) based model can get for this problem. </p>\n\n<p>My LSTM model (moving window preprocessed ) got 2.50563(Private), 1.58003(Public). </p>",
  "messages": [
    {
      "id": 546924,
      "postDate": "2019-06-07T02:51:41.913Z",
      "content": "<p>I am curious how good a NN (e.g., CNN, RNN, Dense) based model can get for this problem. </p>\n\n<p>My LSTM model (moving window preprocessed ) got 2.50563(Private), 1.58003(Public). </p>",
      "rawMarkdown": "I am curious how good a NN (e.g., CNN, RNN, Dense) based model can get for this problem. \n\nMy LSTM model (moving window preprocessed ) got 2.50563(Private), 1.58003(Public). \n",
      "votes": 3
    },
    {
      "id": 547172,
      "postDate": "2019-06-07T11:13:54.553Z",
      "content": "<p><strong>Scaled based on test data set leak</strong>\nprivate 2.34085, public 1.80018</p>\n\n<p><strong>Not scaled based on test data set leak</strong>\nprivate 2.77165, public 1.65713</p>\n\n<p>I decimated the data by a factor of 5 and computed the magnitude spectrogram. My neural network processed the spectrogram as an array of 1D vectors as opposed to a 2D image. I modifed the ResNet architecture described <a href=\"https://github.com/hfawaz/dl-4-tsc\">here</a>. I changed the single final output to use a linear activation function since this is was a regression problem. I added a fully connected layer prior to the output node. I also removed the first layer of each block and all batch normalization (not sure why this actually improved results).</p>",
      "rawMarkdown": "**Scaled based on test data set leak**\nprivate 2.34085, public 1.80018\n\n**Not scaled based on test data set leak**\nprivate 2.77165, public 1.65713\n\nI decimated the data by a factor of 5 and computed the magnitude spectrogram. My neural network processed the spectrogram as an array of 1D vectors as opposed to a 2D image. I modifed the ResNet architecture described [here](https://github.com/hfawaz/dl-4-tsc). I changed the single final output to use a linear activation function since this is was a regression problem. I added a fully connected layer prior to the output node. I also removed the first layer of each block and all batch normalization (not sure why this actually improved results).",
      "votes": 1,
      "replies": [
        {
          "id": 547554,
          "postDate": "2019-06-07T21:20:04.817Z",
          "content": "<p>cool, I also tried 1d conv + LSTM.</p>",
          "rawMarkdown": "cool, I also tried 1d conv + LSTM."
        }
      ]
    },
    {
      "id": 546945,
      "postDate": "2019-06-07T03:44:44.013Z",
      "content": "<p>Our team has single model (private 2.27820, public 1.74057). My teammate will be posting the kernel soon, be on the lookout</p>",
      "rawMarkdown": "Our team has single model (private 2.27820, public 1.74057). My teammate will be posting the kernel soon, be on the lookout",
      "votes": 2,
      "replies": [
        {
          "id": 547551,
          "postDate": "2019-06-07T21:19:20.120Z",
          "content": "<p>cool, cannot wait to see your kennel </p>",
          "rawMarkdown": "cool, cannot wait to see your kennel "
        }
      ]
    },
    {
      "id": 547035,
      "postDate": "2019-06-07T07:17:44.613Z",
      "content": "<p>CNN: 1.63 public, 2.40 private (unfortunately not selected...) directly on the (squared) acoustic data. <a href=\"https://www.kaggle.com/friedchips/simple-cnn-would-have-been-top-25\">Here is the kernel with a full explanation</a>.</p>",
      "rawMarkdown": "CNN: 1.63 public, 2.40 private (unfortunately not selected...) directly on the (squared) acoustic data. [Here is the kernel with a full explanation](https://www.kaggle.com/friedchips/simple-cnn-would-have-been-top-25).",
      "replies": [
        {
          "id": 547552,
          "postDate": "2019-06-07T21:19:31.037Z",
          "content": "<p>thanks ...</p>",
          "rawMarkdown": "thanks ..."
        }
      ]
    },
    {
      "id": 546954,
      "postDate": "2019-06-07T04:05:30.220Z",
      "rawMarkdown": "",
      "votes": 2,
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 547172,
      "author_name": "Trent",
      "author_url": "",
      "post_date": "2019-06-07T11:13:54.553000",
      "content": "<p><strong>Scaled based on test data set leak</strong>\nprivate 2.34085, public 1.80018</p>\n\n<p><strong>Not scaled based on test data set leak</strong>\nprivate 2.77165, public 1.65713</p>\n\n<p>I decimated the data by a factor of 5 and computed the magnitude spectrogram. My neural network processed the spectrogram as an array of 1D vectors as opposed to a 2D image. I modifed the ResNet architecture described <a href=\"https://github.com/hfawaz/dl-4-tsc\">here</a>. I changed the single final output to use a linear activation function since this is was a regression problem. I added a fully connected layer prior to the output node. I also removed the first layer of each block and all batch normalization (not sure why this actually improved results).</p>",
      "votes": 1,
      "replies": [
        {
          "id": 547554,
          "author_name": "Z. Liu",
          "author_url": "",
          "post_date": "2019-06-07T21:20:04.817000",
          "content": "<p>cool, I also tried 1d conv + LSTM.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 546945,
      "author_name": "CoreyJamesLevinson",
      "author_url": "",
      "post_date": "2019-06-07T03:44:44.013000",
      "content": "<p>Our team has single model (private 2.27820, public 1.74057). My teammate will be posting the kernel soon, be on the lookout</p>",
      "votes": 2,
      "replies": [
        {
          "id": 547551,
          "author_name": "Z. Liu",
          "author_url": "",
          "post_date": "2019-06-07T21:19:20.120000",
          "content": "<p>cool, cannot wait to see your kennel </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 547035,
      "author_name": "Markus Frank",
      "author_url": "",
      "post_date": "2019-06-07T07:17:44.613000",
      "content": "<p>CNN: 1.63 public, 2.40 private (unfortunately not selected...) directly on the (squared) acoustic data. <a href=\"https://www.kaggle.com/friedchips/simple-cnn-would-have-been-top-25\">Here is the kernel with a full explanation</a>.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 547552,
          "author_name": "Z. Liu",
          "author_url": "",
          "post_date": "2019-06-07T21:19:31.037000",
          "content": "<p>thanks ...</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 546954,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-06-07T04:05:30.220000",
      "content": "",
      "votes": 2,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "546924": "I am curious how good a NN (e.g., CNN, RNN, Dense) based model can get for this problem. \n\nMy LSTM model (moving window preprocessed ) got 2.50563(Private), 1.58003(Public). \n",
    "547172": "**Scaled based on test data set leak**\nprivate 2.34085, public 1.80018\n\n**Not scaled based on test data set leak**\nprivate 2.77165, public 1.65713\n\nI decimated the data by a factor of 5 and computed the magnitude spectrogram. My neural network processed the spectrogram as an array of 1D vectors as opposed to a 2D image. I modifed the ResNet architecture described [here](https://github.com/hfawaz/dl-4-tsc). I changed the single final output to use a linear activation function since this is was a regression problem. I added a fully connected layer prior to the output node. I also removed the first layer of each block and all batch normalization (not sure why this actually improved results).",
    "546945": "Our team has single model (private 2.27820, public 1.74057). My teammate will be posting the kernel soon, be on the lookout",
    "547035": "CNN: 1.63 public, 2.40 private (unfortunately not selected...) directly on the (squared) acoustic data. [Here is the kernel with a full explanation](https://www.kaggle.com/friedchips/simple-cnn-would-have-been-top-25).",
    "546954": ""
  }
}