{
  "id": 436845,
  "title": "126th Place Solution",
  "url": "/competitions/asl-fingerspelling/discussion/436845",
  "author_name": "Mark Wijkhuizen",
  "post_date": "2023-09-04T10:39:38.822000",
  "votes": 5,
  "comment_count": 2,
  "views": 0,
  "content": "<p><a href=\"https://www.kaggle.com/code/markwijkhuizen/aslfr-inference-python3-7-tpu\" target=\"_blank\">This solution</a> is an improved version of <a href=\"https://www.kaggle.com/code/markwijkhuizen/aslfr-transformer-training-inference\" target=\"_blank\">my public notebook</a> with the following modifications:</p>\n<ul>\n<li>Running on a TPU to allow for faster training and experimentation</li>\n<li>Increased number of input frames 128 → 288</li>\n<li>Batch size 64 → 512</li>\n<li>Units encoder 384 → 288</li>\n<li>Units decoder 256 → 128</li>\n</ul>\n<p>For an in depth explanation of the approach I refer to the public notebook code and comment section.</p>\n<p>The expect the performance boost in my private notebook is mostly due to the decreased model capacity and increased batch size, this should reduce overfitting.</p>",
  "messages": [
    {
      "id": 2422909,
      "postDate": "2023-09-04T10:39:38.823Z",
      "content": "<p><a href=\"https://www.kaggle.com/code/markwijkhuizen/aslfr-inference-python3-7-tpu\" target=\"_blank\">This solution</a> is an improved version of <a href=\"https://www.kaggle.com/code/markwijkhuizen/aslfr-transformer-training-inference\" target=\"_blank\">my public notebook</a> with the following modifications:</p>\n<ul>\n<li>Running on a TPU to allow for faster training and experimentation</li>\n<li>Increased number of input frames 128 → 288</li>\n<li>Batch size 64 → 512</li>\n<li>Units encoder 384 → 288</li>\n<li>Units decoder 256 → 128</li>\n</ul>\n<p>For an in depth explanation of the approach I refer to the public notebook code and comment section.</p>\n<p>The expect the performance boost in my private notebook is mostly due to the decreased model capacity and increased batch size, this should reduce overfitting.</p>",
      "rawMarkdown": "[This solution](https://www.kaggle.com/code/markwijkhuizen/aslfr-inference-python3-7-tpu) is an improved version of [my public notebook](https://www.kaggle.com/code/markwijkhuizen/aslfr-transformer-training-inference) with the following modifications:\n\n* Running on a TPU to allow for faster training and experimentation\n* Increased number of input frames 128 → 288\n* Batch size 64 → 512\n* Units encoder 384 → 288\n* Units decoder 256 → 128\n\nFor an in depth explanation of the approach I refer to the public notebook code and comment section.\n\nThe expect the performance boost in my private notebook is mostly due to the decreased model capacity and increased batch size, this should reduce overfitting.",
      "votes": 5
    },
    {
      "id": 2422915,
      "postDate": "2023-09-04T10:44:03.523Z",
      "content": "<p>Thankyou for this <a href=\"https://www.kaggle.com/markwijkhuizen\" target=\"_blank\">@markwijkhuizen</a> </p>",
      "rawMarkdown": "Thankyou for this @markwijkhuizen ",
      "votes": 1
    },
    {
      "id": 2426086,
      "postDate": "2023-09-06T11:55:21.470Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 2422915,
      "author_name": "Laksika Tharmalingam",
      "author_url": "",
      "post_date": "2023-09-04T10:44:03.523000",
      "content": "<p>Thankyou for this <a href=\"https://www.kaggle.com/markwijkhuizen\" target=\"_blank\">@markwijkhuizen</a> </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 2426086,
      "author_name": "",
      "author_url": "",
      "post_date": "2023-09-06T11:55:21.470000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2422909": "[This solution](https://www.kaggle.com/code/markwijkhuizen/aslfr-inference-python3-7-tpu) is an improved version of [my public notebook](https://www.kaggle.com/code/markwijkhuizen/aslfr-transformer-training-inference) with the following modifications:\n\n* Running on a TPU to allow for faster training and experimentation\n* Increased number of input frames 128 → 288\n* Batch size 64 → 512\n* Units encoder 384 → 288\n* Units decoder 256 → 128\n\nFor an in depth explanation of the approach I refer to the public notebook code and comment section.\n\nThe expect the performance boost in my private notebook is mostly due to the decreased model capacity and increased batch size, this should reduce overfitting.",
    "2422915": "Thankyou for this @markwijkhuizen ",
    "2426086": ""
  }
}