{
  "id": 477460,
  "title": "EEGNet Modeling Question",
  "url": "/competitions/hms-harmful-brain-activity-classification/discussion/477460",
  "author_name": "snehal",
  "post_date": "2024-02-16T07:37:39.710000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I had a question about <a href=\"https://www.kaggle.com/nischaydnk\" target=\"_blank\">@nischaydnk</a> 's EEG net model. In the model after parallel convolution block with multiple kernel sizes you chose to concatenate along axis 2 which was the time axis of the signal as opposed to maybe padding the output of all parallel convolutions to same length and concatenating along axis 1 (channel axis). I am referring to this code: <br>\n<code>for i in range(len(self.kernels)):\nsep = self.parallel_conv[i](x)\nout_sep.append(sep)\nout = torch.cat(out_sep, dim=2)</code><br>\nI tried concatenating along channel axis because i thought it made more sense but it performs significantly worse.</p>\n<p>also not sure how to add newlines to the code in my discussion post</p>",
  "messages": [
    {
      "id": 2654531,
      "postDate": "2024-02-16T07:37:39.710Z",
      "content": "<p>I had a question about <a href=\"https://www.kaggle.com/nischaydnk\" target=\"_blank\">@nischaydnk</a> 's EEG net model. In the model after parallel convolution block with multiple kernel sizes you chose to concatenate along axis 2 which was the time axis of the signal as opposed to maybe padding the output of all parallel convolutions to same length and concatenating along axis 1 (channel axis). I am referring to this code: <br>\n<code>for i in range(len(self.kernels)):\nsep = self.parallel_conv[i](x)\nout_sep.append(sep)\nout = torch.cat(out_sep, dim=2)</code><br>\nI tried concatenating along channel axis because i thought it made more sense but it performs significantly worse.</p>\n<p>also not sure how to add newlines to the code in my discussion post</p>",
      "rawMarkdown": "I had a question about @nischaydnk 's EEG net model. In the model after parallel convolution block with multiple kernel sizes you chose to concatenate along axis 2 which was the time axis of the signal as opposed to maybe padding the output of all parallel convolutions to same length and concatenating along axis 1 (channel axis). I am referring to this code: \n```for i in range(len(self.kernels)):\nsep = self.parallel_conv[i](x)\nout_sep.append(sep)\nout = torch.cat(out_sep, dim=2)```\nI tried concatenating along channel axis because i thought it made more sense but it performs significantly worse.\n\nalso not sure how to add newlines to the code in my discussion post",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2654531": "I had a question about @nischaydnk 's EEG net model. In the model after parallel convolution block with multiple kernel sizes you chose to concatenate along axis 2 which was the time axis of the signal as opposed to maybe padding the output of all parallel convolutions to same length and concatenating along axis 1 (channel axis). I am referring to this code: \n```for i in range(len(self.kernels)):\nsep = self.parallel_conv[i](x)\nout_sep.append(sep)\nout = torch.cat(out_sep, dim=2)```\nI tried concatenating along channel axis because i thought it made more sense but it performs significantly worse.\n\nalso not sure how to add newlines to the code in my discussion post"
  }
}