{
  "id": 16365,
  "title": "RNN approach",
  "url": "/competitions/grasp-and-lift-eeg-detection/discussion/16365",
  "author_name": "",
  "post_date": "2015-09-07T18:22:16.130Z",
  "votes": 2,
  "comment_count": 4,
  "views": 2411,
  "content": "<p>Hi!</p>\n\n<p>I have little knowledge in EEG and signal processing but I worked on this competition, because I was curious how good RNNs can do in here. Unfortunately I missed the first-submission deadline and therefore I have no idea of the final score I would get. If any of you are interested in trying them out you can check out my code at github (it's using Torch):\n<a href=\"https://github.com/apaszke/kaggle-grasp-and-lift\">https://github.com/apaszke/kaggle-grasp-and-lift</a></p>\n\n<p>I've also written a blog post describing some LSTM implementation details:\n<a href=\"http://apaszke.github.io/lstm-explained.html\">http://apaszke.github.io/lstm-explained.html</a></p>\n\n<p>If you have any questions, suggestions, would like to try it out or just talk, feel free to message me or write here.</p>",
  "messages": [
    {
      "id": "91788",
      "postDate": "09/07/2015 18:22:16",
      "content": "<p>Hi!</p>\n\n<p>I have little knowledge in EEG and signal processing but I worked on this competition, because I was curious how good RNNs can do in here. Unfortunately I missed the first-submission deadline and therefore I have no idea of the final score I would get. If any of you are interested in trying them out you can check out my code at github (it's using Torch):\n<a href=\"https://github.com/apaszke/kaggle-grasp-and-lift\">https://github.com/apaszke/kaggle-grasp-and-lift</a></p>\n\n<p>I've also written a blog post describing some LSTM implementation details:\n<a href=\"http://apaszke.github.io/lstm-explained.html\">http://apaszke.github.io/lstm-explained.html</a></p>\n\n<p>If you have any questions, suggestions, would like to try it out or just talk, feel free to message me or write here.</p>",
      "rawMarkdown": "Hi!\r\n\r\nI have little knowledge in EEG and signal processing but I worked on this competition, because I was curious how good RNNs can do in here. Unfortunately I missed the first-submission deadline and therefore I have no idea of the final score I would get. If any of you are interested in trying them out you can check out my code at github (it's using Torch):\r\nhttps://github.com/apaszke/kaggle-grasp-and-lift\r\n\r\nI've also written a blog post describing some LSTM implementation details:\r\nhttp://apaszke.github.io/lstm-explained.html\r\n\r\nIf you have any questions, suggestions, would like to try it out or just talk, feel free to message me or write here.",
      "votes": null
    },
    {
      "id": "92000",
      "postDate": "09/10/2015 11:56:14",
      "content": "<p>Hi,</p>\n\n<p>cool thank you for posting this :)</p>\n\n<p>I am also interested in the performance of RNNs on this dataset. \nDo you have any auc/whatever-values for your model splitting a part of the training set off as a validation/test set?\nTo get some indication of how good they already are.</p>\n\n<p>Best,\nRobin</p>",
      "rawMarkdown": "Hi,\r\n\r\ncool thank you for posting this :)\r\n\r\n I am also interested in the performance of RNNs on this dataset. \r\nDo you have any auc/whatever-values for your model splitting a part of the training set off as a validation/test set?\r\nTo get some indication of how good they already are.\r\n\r\nBest,\r\nRobin",
      "votes": null
    },
    {
      "id": "92005",
      "postDate": "09/10/2015 13:15:27",
      "content": "<p>Hey,</p>\n\n<p>@apaszke, thanks for sharing! </p>\n\n<p>Too bad you missed the 1st submission deadline. But after the competition has ended you can still make submissions to see what score (and spot) you'd get on the private leaderboard :-) </p>",
      "rawMarkdown": "Hey,\r\n\r\n@apaszke, thanks for sharing! \r\n\r\nToo bad you missed the 1st submission deadline. But after the competition has ended you can still make submissions to see what score (and spot) you'd get on the private leaderboard :-)",
      "votes": null
    },
    {
      "id": "92043",
      "postDate": "09/10/2015 21:33:14",
      "content": "<p>Thanks for sharing your code.\nI used a modified version of Tim Hochberg's Script to implement a simple LSTM using 10 and 100 units.</p>\n\n<p>They scored:\npublic: 0.84157, private: 0.84085 and public: 0.85611, private: 0.85841,\nrespectively.</p>\n\n<p>Greets!</p>",
      "rawMarkdown": "Thanks for sharing your code.\r\nI used a modified version of Tim Hochberg's Script to implement a simple LSTM using 10 and 100 units.\r\n\r\nThey scored:\r\npublic: 0.84157, private: 0.84085 and public: 0.85611, private: 0.85841,\r\nrespectively.\r\n\r\n\r\nGreets!",
      "votes": null
    },
    {
      "id": "92163",
      "postDate": "09/11/2015 16:13:51",
      "content": "<p>Hi!</p>\n\n<p>@blaine I didn't know that. I've stopped working on it as soon as I've realized that it's already after the deadline :(</p>\n\n<p>I've used two LSTM layers of 128 units.</p>\n\n<p>As far as I remember I got slightly better performance than @ptrblck, but it really differed from subject to subject (for some i got AUC &gt; 0.95 in the first run, but for some it was a struggle to get around 0.75).</p>\n\n<p>I haven't played much with signal processing and trained it mainly on raw features or with a low-pass filter from one script (but I'm not completely sure which bands it passes), so it probably could get much better.</p>",
      "rawMarkdown": "Hi!\r\n\r\n@blaine I didn't know that. I've stopped working on it as soon as I've realized that it's already after the deadline :(\r\n\r\nI've used two LSTM layers of 128 units.\r\n\r\nAs far as I remember I got slightly better performance than @ptrblck, but it really differed from subject to subject (for some i got AUC > 0.95 in the first run, but for some it was a struggle to get around 0.75).\r\n\r\nI haven't played much with signal processing and trained it mainly on raw features or with a low-pass filter from one script (but I'm not completely sure which bands it passes), so it probably could get much better.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 92000,
      "author_name": "robintibor",
      "author_url": "",
      "post_date": "09/10/2015 11:56:14",
      "content": "<p>Hi,</p>\n\n<p>cool thank you for posting this :)</p>\n\n<p>I am also interested in the performance of RNNs on this dataset. \nDo you have any auc/whatever-values for your model splitting a part of the training set off as a validation/test set?\nTo get some indication of how good they already are.</p>\n\n<p>Best,\nRobin</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 92005,
      "author_name": "rafalcycon",
      "author_url": "",
      "post_date": "09/10/2015 13:15:27",
      "content": "<p>Hey,</p>\n\n<p>@apaszke, thanks for sharing! </p>\n\n<p>Too bad you missed the 1st submission deadline. But after the competition has ended you can still make submissions to see what score (and spot) you'd get on the private leaderboard :-) </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 92043,
      "author_name": "ptrblck",
      "author_url": "",
      "post_date": "09/10/2015 21:33:14",
      "content": "<p>Thanks for sharing your code.\nI used a modified version of Tim Hochberg's Script to implement a simple LSTM using 10 and 100 units.</p>\n\n<p>They scored:\npublic: 0.84157, private: 0.84085 and public: 0.85611, private: 0.85841,\nrespectively.</p>\n\n<p>Greets!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 92163,
      "author_name": "apaszke",
      "author_url": "",
      "post_date": "09/11/2015 16:13:51",
      "content": "<p>Hi!</p>\n\n<p>@blaine I didn't know that. I've stopped working on it as soon as I've realized that it's already after the deadline :(</p>\n\n<p>I've used two LSTM layers of 128 units.</p>\n\n<p>As far as I remember I got slightly better performance than @ptrblck, but it really differed from subject to subject (for some i got AUC &gt; 0.95 in the first run, but for some it was a struggle to get around 0.75).</p>\n\n<p>I haven't played much with signal processing and trained it mainly on raw features or with a low-pass filter from one script (but I'm not completely sure which bands it passes), so it probably could get much better.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "91788": "Hi!\r\n\r\nI have little knowledge in EEG and signal processing but I worked on this competition, because I was curious how good RNNs can do in here. Unfortunately I missed the first-submission deadline and therefore I have no idea of the final score I would get. If any of you are interested in trying them out you can check out my code at github (it's using Torch):\r\nhttps://github.com/apaszke/kaggle-grasp-and-lift\r\n\r\nI've also written a blog post describing some LSTM implementation details:\r\nhttp://apaszke.github.io/lstm-explained.html\r\n\r\nIf you have any questions, suggestions, would like to try it out or just talk, feel free to message me or write here.",
    "92000": "Hi,\r\n\r\ncool thank you for posting this :)\r\n\r\n I am also interested in the performance of RNNs on this dataset. \r\nDo you have any auc/whatever-values for your model splitting a part of the training set off as a validation/test set?\r\nTo get some indication of how good they already are.\r\n\r\nBest,\r\nRobin",
    "92005": "Hey,\r\n\r\n@apaszke, thanks for sharing! \r\n\r\nToo bad you missed the 1st submission deadline. But after the competition has ended you can still make submissions to see what score (and spot) you'd get on the private leaderboard :-)",
    "92043": "Thanks for sharing your code.\r\nI used a modified version of Tim Hochberg's Script to implement a simple LSTM using 10 and 100 units.\r\n\r\nThey scored:\r\npublic: 0.84157, private: 0.84085 and public: 0.85611, private: 0.85841,\r\nrespectively.\r\n\r\n\r\nGreets!",
    "92163": "Hi!\r\n\r\n@blaine I didn't know that. I've stopped working on it as soon as I've realized that it's already after the deadline :(\r\n\r\nI've used two LSTM layers of 128 units.\r\n\r\nAs far as I remember I got slightly better performance than @ptrblck, but it really differed from subject to subject (for some i got AUC > 0.95 in the first run, but for some it was a struggle to get around 0.75).\r\n\r\nI haven't played much with signal processing and trained it mainly on raw features or with a low-pass filter from one script (but I'm not completely sure which bands it passes), so it probably could get much better."
  },
  "source": "meta"
}