{
  "id": 430160,
  "title": "Quantization Aware Training - Gaps in Scores",
  "url": "/competitions/asl-fingerspelling/discussion/430160",
  "author_name": "Aaryam Sharma",
  "post_date": "2023-08-08T12:46:45.694000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Context: Using TensorFlow quantization</p>\n<p>I'm wondering if others have used quantization in their models to make and submit larger models.</p>\n<p>Personally, I'm finding a large gap between my validation scores and the scores on the leaderboard after quantization. By my understanding, I shouldn't see a difference as large as 0.1</p>\n<p>I wonder if anyone is using quantization, and was able to make it work. If so, could you share which type of quantization is working for you?</p>\n<p>I'm not asking the model, just whether or not someone found success with quantization.</p>\n<p>Also of note is that I'm not getting a large enough speed-up with quantization aware training.</p>\n<p>I am also finding that data augmentations are not helping me.</p>\n<p>It would be really helpful if someone could share their experience with both of these :)<br>\nIt'll help me find mistakes in my own work, or at least learn something new.</p>\n<p>Thank you!</p>",
  "messages": [
    {
      "id": 2380158,
      "postDate": "2023-08-08T12:46:45.693Z",
      "content": "<p>Context: Using TensorFlow quantization</p>\n<p>I'm wondering if others have used quantization in their models to make and submit larger models.</p>\n<p>Personally, I'm finding a large gap between my validation scores and the scores on the leaderboard after quantization. By my understanding, I shouldn't see a difference as large as 0.1</p>\n<p>I wonder if anyone is using quantization, and was able to make it work. If so, could you share which type of quantization is working for you?</p>\n<p>I'm not asking the model, just whether or not someone found success with quantization.</p>\n<p>Also of note is that I'm not getting a large enough speed-up with quantization aware training.</p>\n<p>I am also finding that data augmentations are not helping me.</p>\n<p>It would be really helpful if someone could share their experience with both of these :)<br>\nIt'll help me find mistakes in my own work, or at least learn something new.</p>\n<p>Thank you!</p>",
      "rawMarkdown": "Context: Using TensorFlow quantization\n\nI'm wondering if others have used quantization in their models to make and submit larger models.\n\nPersonally, I'm finding a large gap between my validation scores and the scores on the leaderboard after quantization. By my understanding, I shouldn't see a difference as large as 0.1\n\nI wonder if anyone is using quantization, and was able to make it work. If so, could you share which type of quantization is working for you?\n\nI'm not asking the model, just whether or not someone found success with quantization.\n\nAlso of note is that I'm not getting a large enough speed-up with quantization aware training.\n\nI am also finding that data augmentations are not helping me.\n\nIt would be really helpful if someone could share their experience with both of these :)\nIt'll help me find mistakes in my own work, or at least learn something new.\n\nThank you!",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2380158": "Context: Using TensorFlow quantization\n\nI'm wondering if others have used quantization in their models to make and submit larger models.\n\nPersonally, I'm finding a large gap between my validation scores and the scores on the leaderboard after quantization. By my understanding, I shouldn't see a difference as large as 0.1\n\nI wonder if anyone is using quantization, and was able to make it work. If so, could you share which type of quantization is working for you?\n\nI'm not asking the model, just whether or not someone found success with quantization.\n\nAlso of note is that I'm not getting a large enough speed-up with quantization aware training.\n\nI am also finding that data augmentations are not helping me.\n\nIt would be really helpful if someone could share their experience with both of these :)\nIt'll help me find mistakes in my own work, or at least learn something new.\n\nThank you!"
  }
}