{
  "id": 94325,
  "title": "Private LB 2.43622 kernel (Wavenet + LSTM)",
  "url": "/competitions/LANL-Earthquake-Prediction/discussion/94325",
  "author_name": "",
  "post_date": "2019-06-04T01:37:52.348276600Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p><a href=\"https://www.kaggle.com/wimwim/stacked-raw-lstm?scriptVersionId=12415596\">https://www.kaggle.com/wimwim/stacked-raw-lstm?scriptVersionId=12415596</a></p>\n\n<p>This wavenet + LSTM based single model is enough to get in top 50. Unfortunately I didn't  choose this as my submission because the Public LB of this model is very bad (1.67242)...</p>",
  "messages": [
    {
      "id": "542568",
      "postDate": "06/04/2019 01:37:52",
      "content": "<p><a href=\"https://www.kaggle.com/wimwim/stacked-raw-lstm?scriptVersionId=12415596\">https://www.kaggle.com/wimwim/stacked-raw-lstm?scriptVersionId=12415596</a></p>\n\n<p>This wavenet + LSTM based single model is enough to get in top 50. Unfortunately I didn't  choose this as my submission because the Public LB of this model is very bad (1.67242)...</p>",
      "rawMarkdown": "https://www.kaggle.com/wimwim/stacked-raw-lstm?scriptVersionId=12415596\n\nThis wavenet + LSTM based single model is enough to get in top 50. Unfortunately I didn't  choose this as my submission because the Public LB of this model is very bad (1.67242)...",
      "votes": null
    },
    {
      "id": "542589",
      "postDate": "06/04/2019 01:51:28",
      "content": "<p>Your kernel looks interesting but it is difficult to grasp just reading the code. Can you explain the ideas in a few paragraphs?</p>",
      "rawMarkdown": "Your kernel looks interesting but it is difficult to grasp just reading the code. Can you explain the ideas in a few paragraphs?",
      "votes": null
    },
    {
      "id": "542604",
      "postDate": "06/04/2019 02:01:30",
      "content": "<p>Sure, the model is consist of two wavenet blocks (<a href=\"https://arxiv.org/pdf/1609.03499.pdf\">https://arxiv.org/pdf/1609.03499.pdf</a>) and 1 LSTM.  I use 1d Convolutional layer in the middle to reduce the seqenuce length so that LSTM can handle it. Attention layer, Dropout and FC layer are then applied after LSTM.</p>",
      "rawMarkdown": "Sure, the model is consist of two wavenet blocks (https://arxiv.org/pdf/1609.03499.pdf) and 1 LSTM.  I use 1d Convolutional layer in the middle to reduce the seqenuce length so that LSTM can handle it. Attention layer, Dropout and FC layer are then applied after LSTM.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 542589,
      "author_name": "petewills",
      "author_url": "",
      "post_date": "06/04/2019 01:51:28",
      "content": "<p>Your kernel looks interesting but it is difficult to grasp just reading the code. Can you explain the ideas in a few paragraphs?</p>",
      "votes": null,
      "replies": [
        {
          "id": 542604,
          "author_name": "wimwim",
          "author_url": "",
          "post_date": "06/04/2019 02:01:30",
          "content": "<p>Sure, the model is consist of two wavenet blocks (<a href=\"https://arxiv.org/pdf/1609.03499.pdf\">https://arxiv.org/pdf/1609.03499.pdf</a>) and 1 LSTM.  I use 1d Convolutional layer in the middle to reduce the seqenuce length so that LSTM can handle it. Attention layer, Dropout and FC layer are then applied after LSTM.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "542568": "https://www.kaggle.com/wimwim/stacked-raw-lstm?scriptVersionId=12415596\n\nThis wavenet + LSTM based single model is enough to get in top 50. Unfortunately I didn't  choose this as my submission because the Public LB of this model is very bad (1.67242)...",
    "542589": "Your kernel looks interesting but it is difficult to grasp just reading the code. Can you explain the ideas in a few paragraphs?",
    "542604": "Sure, the model is consist of two wavenet blocks (https://arxiv.org/pdf/1609.03499.pdf) and 1 LSTM.  I use 1d Convolutional layer in the middle to reduce the seqenuce length so that LSTM can handle it. Attention layer, Dropout and FC layer are then applied after LSTM."
  },
  "source": "meta"
}