{
  "id": 56435,
  "title": "multiple filters in CNN gives 0.0004 boost to 0.9780 ",
  "url": "/competitions/talkingdata-adtracking-fraud-detection/discussion/56435",
  "author_name": "Yibing Wu",
  "post_date": "2018-05-09T17:45:19.380000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>My rank doesn't support me to speak too much :-) but here is my tiny share from experiments. It increased public LB from 0.9776 to 0.9780, by changing:</p>\n\n<p>conv = Conv1D(num_filters, kernel_size = 4, activation='relu', padding='same')(s_dout) <br>\nto:</p>\n\n<p>convs = []</p>\n\n<p>filter_sizes = [3,4,5,6]</p>\n\n<p>for fsize in filter_sizes:</p>\n\n<pre><code>  x = Conv1D(num_filters, kernel_size = fsize, activation='relu', padding='same')(s_dout)\n\n  convs.append(x)\n</code></pre>\n\n<p>conv = Concatenate()([convs[0], convs[1], convs[2],convs[3]])</p>",
  "messages": [
    {
      "id": 326437,
      "postDate": "2018-05-09T17:45:19.380Z",
      "content": "<p>My rank doesn't support me to speak too much :-) but here is my tiny share from experiments. It increased public LB from 0.9776 to 0.9780, by changing:</p>\n\n<p>conv = Conv1D(num_filters, kernel_size = 4, activation='relu', padding='same')(s_dout) <br>\nto:</p>\n\n<p>convs = []</p>\n\n<p>filter_sizes = [3,4,5,6]</p>\n\n<p>for fsize in filter_sizes:</p>\n\n<pre><code>  x = Conv1D(num_filters, kernel_size = fsize, activation='relu', padding='same')(s_dout)\n\n  convs.append(x)\n</code></pre>\n\n<p>conv = Concatenate()([convs[0], convs[1], convs[2],convs[3]])</p>",
      "rawMarkdown": "My rank doesn't support me to speak too much :-) but here is my tiny share from experiments. It increased public LB from 0.9776 to 0.9780, by changing:\n\nconv = Conv1D(num_filters, kernel_size = 4, activation='relu', padding='same')(s_dout)  \nto:\n\nconvs = []\n\nfilter_sizes = [3,4,5,6]\n\nfor fsize in filter_sizes:\n\n      x = Conv1D(num_filters, kernel_size = fsize, activation='relu', padding='same')(s_dout)\n\n      convs.append(x)\n\nconv = Concatenate()([convs[0], convs[1], convs[2],convs[3]])\n",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "326437": "My rank doesn't support me to speak too much :-) but here is my tiny share from experiments. It increased public LB from 0.9776 to 0.9780, by changing:\n\nconv = Conv1D(num_filters, kernel_size = 4, activation='relu', padding='same')(s_dout)  \nto:\n\nconvs = []\n\nfilter_sizes = [3,4,5,6]\n\nfor fsize in filter_sizes:\n\n      x = Conv1D(num_filters, kernel_size = fsize, activation='relu', padding='same')(s_dout)\n\n      convs.append(x)\n\nconv = Concatenate()([convs[0], convs[1], convs[2],convs[3]])\n"
  }
}