{
  "id": 73758,
  "title": "best lstm results using point coordinates (not cnn features)",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/73758",
  "author_name": "hengck23",
  "post_date": "2018-12-05T11:55:19.057000",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>after reading all the forum, this is what i find:</p>\n\n<p>@jeandebleau :  \"I reached 0.935 with a single recurrent network\" see, <a href=\"https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/73738\">https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/73738</a></p>\n\n<p>Please let me know if anyone can go higher then this. Thanks.</p>\n\n<p>I reach about LB 0.928.</p>",
  "messages": [
    {
      "id": 434053,
      "postDate": "2018-12-05T20:31:30.287Z",
      "content": "<p>I hit 0.94188 private, 0.94139 public LB with a geom mean of two RNN models here: <a href=\"https://gist.github.com/rwightman/893723e023ffccca4abf67f92cb76526\">https://gist.github.com/rwightman/893723e023ffccca4abf67f92cb76526</a></p>\n\n<p>Technically they both used convolutions, but 1d convolutions on the input sequence to expand the feature dim before feeding into a stack of 4 bidirectional LSTM. I did reinterpolate the input sequences so they were a fairly consistent length despite the data set having a wide range.</p>\n\n<p>I would have liked to experiment with these more but ran out of time. Had to ensemble with some CNN models to bump the score higher.</p>",
      "rawMarkdown": "I hit 0.94188 private, 0.94139 public LB with a geom mean of two RNN models here: https://gist.github.com/rwightman/893723e023ffccca4abf67f92cb76526\n\nTechnically they both used convolutions, but 1d convolutions on the input sequence to expand the feature dim before feeding into a stack of 4 bidirectional LSTM. I did reinterpolate the input sequences so they were a fairly consistent length despite the data set having a wide range.\n\nI would have liked to experiment with these more but ran out of time. Had to ensemble with some CNN models to bump the score higher.",
      "votes": 1
    },
    {
      "id": 433723,
      "postDate": "2018-12-05T11:55:19.057Z",
      "content": "<p>after reading all the forum, this is what i find:</p>\n\n<p>@jeandebleau :  \"I reached 0.935 with a single recurrent network\" see, <a href=\"https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/73738\">https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/73738</a></p>\n\n<p>Please let me know if anyone can go higher then this. Thanks.</p>\n\n<p>I reach about LB 0.928.</p>",
      "rawMarkdown": "after reading all the forum, this is what i find:\n\n@jeandebleau :  \"I reached 0.935 with a single recurrent network\" see, https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/73738\n\nPlease let me know if anyone can go higher then this. Thanks.\n\nI reach about LB 0.928.",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 434053,
      "author_name": "RossWightman",
      "author_url": "",
      "post_date": "2018-12-05T20:31:30.287000",
      "content": "<p>I hit 0.94188 private, 0.94139 public LB with a geom mean of two RNN models here: <a href=\"https://gist.github.com/rwightman/893723e023ffccca4abf67f92cb76526\">https://gist.github.com/rwightman/893723e023ffccca4abf67f92cb76526</a></p>\n\n<p>Technically they both used convolutions, but 1d convolutions on the input sequence to expand the feature dim before feeding into a stack of 4 bidirectional LSTM. I did reinterpolate the input sequences so they were a fairly consistent length despite the data set having a wide range.</p>\n\n<p>I would have liked to experiment with these more but ran out of time. Had to ensemble with some CNN models to bump the score higher.</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "434053": "I hit 0.94188 private, 0.94139 public LB with a geom mean of two RNN models here: https://gist.github.com/rwightman/893723e023ffccca4abf67f92cb76526\n\nTechnically they both used convolutions, but 1d convolutions on the input sequence to expand the feature dim before feeding into a stack of 4 bidirectional LSTM. I did reinterpolate the input sequences so they were a fairly consistent length despite the data set having a wide range.\n\nI would have liked to experiment with these more but ran out of time. Had to ensemble with some CNN models to bump the score higher.",
    "433723": "after reading all the forum, this is what i find:\n\n@jeandebleau :  \"I reached 0.935 with a single recurrent network\" see, https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/73738\n\nPlease let me know if anyone can go higher then this. Thanks.\n\nI reach about LB 0.928."
  }
}