{
  "id": 415490,
  "title": "A request for pre and post processing layer examples",
  "url": "/competitions/asl-fingerspelling/discussion/415490",
  "author_name": "Steven Gubkin",
  "post_date": "2023-06-06T17:11:20.444000",
  "votes": 0,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi everyone!  This is my first Kaggle competition and my first time playing with keras/tensorflow, and I am hoping you might be able to help me out.  Currently I have a model which can correctly identify the first character of the phrase with about 50% accuracy.</p>\n<p>Anyone have a basic version of custom pre and post processing layers in keras that actually work?</p>\n<p>I am not looking for you to reveal state secrets here.  Even just something like:</p>\n<p>pre-processing:</p>\n<ul>\n<li>remove NaNs</li>\n<li>select only hand columns</li>\n<li>translate all hand columns by (x_0, y_0, z_0) to compute relative positions.</li>\n</ul>\n<p>For post-processing assume the model is assigning a character to each frame, rather than associating a phrase to the whole video.</p>\n<p>post-processing:</p>\n<ul>\n<li>Get character associated with each frame.</li>\n<li>return string of unique predicted characters in order.</li>\n</ul>\n<p>Just having a basic working example of custom layers which accomplish this would be extremely valuable to me!  The information on writing custom layers in the keras documentation is a little too opaque for me…</p>",
  "messages": [
    {
      "id": 2290281,
      "postDate": "2023-06-06T17:11:20.443Z",
      "content": "<p>Hi everyone!  This is my first Kaggle competition and my first time playing with keras/tensorflow, and I am hoping you might be able to help me out.  Currently I have a model which can correctly identify the first character of the phrase with about 50% accuracy.</p>\n<p>Anyone have a basic version of custom pre and post processing layers in keras that actually work?</p>\n<p>I am not looking for you to reveal state secrets here.  Even just something like:</p>\n<p>pre-processing:</p>\n<ul>\n<li>remove NaNs</li>\n<li>select only hand columns</li>\n<li>translate all hand columns by (x_0, y_0, z_0) to compute relative positions.</li>\n</ul>\n<p>For post-processing assume the model is assigning a character to each frame, rather than associating a phrase to the whole video.</p>\n<p>post-processing:</p>\n<ul>\n<li>Get character associated with each frame.</li>\n<li>return string of unique predicted characters in order.</li>\n</ul>\n<p>Just having a basic working example of custom layers which accomplish this would be extremely valuable to me!  The information on writing custom layers in the keras documentation is a little too opaque for me…</p>",
      "rawMarkdown": "Hi everyone!  This is my first Kaggle competition and my first time playing with keras/tensorflow, and I am hoping you might be able to help me out.  Currently I have a model which can correctly identify the first character of the phrase with about 50% accuracy.\n\nAnyone have a basic version of custom pre and post processing layers in keras that actually work?\n\nI am not looking for you to reveal state secrets here.  Even just something like:\n\npre-processing:\n - remove NaNs\n - select only hand columns\n - translate all hand columns by (x_0, y_0, z_0) to compute relative positions.\n\nFor post-processing assume the model is assigning a character to each frame, rather than associating a phrase to the whole video.\n\npost-processing:\n   - Get character associated with each frame.\n   - return string of unique predicted characters in order.\n\nJust having a basic working example of custom layers which accomplish this would be extremely valuable to me!  The information on writing custom layers in the keras documentation is a little too opaque for me..."
    },
    {
      "id": 2291305,
      "postDate": "2023-06-07T13:11:10.210Z",
      "content": "<p>Dear <a href=\"https://www.kaggle.com/stevengubkin\" target=\"_blank\">@stevengubkin</a> ,</p>\n<p>Figuring out how to implement pre-and postprocessing inside a TfLite model is one of the core elements (and learning opportunities) of this (and the GISLR) competition. We've all spent a lot of time on these aspects and you can find example solutions in the notebooks that have been shared here: <a href=\"https://www.kaggle.com/competitions/asl-fingerspelling/discussion/409438\" target=\"_blank\">https://www.kaggle.com/competitions/asl-fingerspelling/discussion/409438</a></p>",
      "rawMarkdown": "Dear @stevengubkin ,\n\nFiguring out how to implement pre-and postprocessing inside a TfLite model is one of the core elements (and learning opportunities) of this (and the GISLR) competition. We've all spent a lot of time on these aspects and you can find example solutions in the notebooks that have been shared here: https://www.kaggle.com/competitions/asl-fingerspelling/discussion/409438"
    }
  ],
  "comments": [
    {
      "id": 2291305,
      "author_name": "Wondering Alice",
      "author_url": "",
      "post_date": "2023-06-07T13:11:10.210000",
      "content": "<p>Dear <a href=\"https://www.kaggle.com/stevengubkin\" target=\"_blank\">@stevengubkin</a> ,</p>\n<p>Figuring out how to implement pre-and postprocessing inside a TfLite model is one of the core elements (and learning opportunities) of this (and the GISLR) competition. We've all spent a lot of time on these aspects and you can find example solutions in the notebooks that have been shared here: <a href=\"https://www.kaggle.com/competitions/asl-fingerspelling/discussion/409438\" target=\"_blank\">https://www.kaggle.com/competitions/asl-fingerspelling/discussion/409438</a></p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2290281": "Hi everyone!  This is my first Kaggle competition and my first time playing with keras/tensorflow, and I am hoping you might be able to help me out.  Currently I have a model which can correctly identify the first character of the phrase with about 50% accuracy.\n\nAnyone have a basic version of custom pre and post processing layers in keras that actually work?\n\nI am not looking for you to reveal state secrets here.  Even just something like:\n\npre-processing:\n - remove NaNs\n - select only hand columns\n - translate all hand columns by (x_0, y_0, z_0) to compute relative positions.\n\nFor post-processing assume the model is assigning a character to each frame, rather than associating a phrase to the whole video.\n\npost-processing:\n   - Get character associated with each frame.\n   - return string of unique predicted characters in order.\n\nJust having a basic working example of custom layers which accomplish this would be extremely valuable to me!  The information on writing custom layers in the keras documentation is a little too opaque for me...",
    "2291305": "Dear @stevengubkin ,\n\nFiguring out how to implement pre-and postprocessing inside a TfLite model is one of the core elements (and learning opportunities) of this (and the GISLR) competition. We've all spent a lot of time on these aspects and you can find example solutions in the notebooks that have been shared here: https://www.kaggle.com/competitions/asl-fingerspelling/discussion/409438"
  }
}