{
  "id": 410598,
  "title": "New to TF-Lite, can someone help explain the caveats to expect?",
  "url": "/competitions/asl-fingerspelling/discussion/410598",
  "author_name": "Rahul Namboodiri",
  "post_date": "2023-05-15T23:13:25.686000",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hey community!</p>\n<p>I'm a bit new to TF-Lite model deployments, and have done a cursory reading on TF-Lite being ML/DL for edge devices. I've also done some reading on how to optimise TF models for TF-lite from the <a href=\"https://www.tensorflow.org/lite/performance/best_practices\" target=\"_blank\">tensorflow documentation</a>.</p>\n<p>I'd like to know a more comprehensive list of caveats and tradeoffs that need to be considered when building TF-lite models, I'm assuming most of these are based on power and VRAM sizes, but I'm not entirely sure. Also, what would be the best way to test model deployment efficiency as well for TF-lite models?</p>",
  "messages": [
    {
      "id": 2260827,
      "postDate": "2023-05-15T23:13:25.687Z",
      "content": "<p>Hey community!</p>\n<p>I'm a bit new to TF-Lite model deployments, and have done a cursory reading on TF-Lite being ML/DL for edge devices. I've also done some reading on how to optimise TF models for TF-lite from the <a href=\"https://www.tensorflow.org/lite/performance/best_practices\" target=\"_blank\">tensorflow documentation</a>.</p>\n<p>I'd like to know a more comprehensive list of caveats and tradeoffs that need to be considered when building TF-lite models, I'm assuming most of these are based on power and VRAM sizes, but I'm not entirely sure. Also, what would be the best way to test model deployment efficiency as well for TF-lite models?</p>",
      "rawMarkdown": "Hey community!\n\nI'm a bit new to TF-Lite model deployments, and have done a cursory reading on TF-Lite being ML/DL for edge devices. I've also done some reading on how to optimise TF models for TF-lite from the [tensorflow documentation](https://www.tensorflow.org/lite/performance/best_practices).\n\nI'd like to know a more comprehensive list of caveats and tradeoffs that need to be considered when building TF-lite models, I'm assuming most of these are based on power and VRAM sizes, but I'm not entirely sure. Also, what would be the best way to test model deployment efficiency as well for TF-lite models?\n",
      "votes": 4
    },
    {
      "id": 2261747,
      "postDate": "2023-05-16T14:20:40Z",
      "content": "<p>Hi Rahul,</p>\n<p>Your best starting point would be the discussion and example notebooks for the previous ASL (Isolated sign recognition) competition. This has multiple examples of how to generate TfLite models</p>",
      "rawMarkdown": "Hi Rahul,\n\nYour best starting point would be the discussion and example notebooks for the previous ASL (Isolated sign recognition) competition. This has multiple examples of how to generate TfLite models\n",
      "votes": 2,
      "replies": [
        {
          "id": 2262436,
          "postDate": "2023-05-16T23:33:38.587Z",
          "content": "<p>thank you Alice, It's less the issue of generating TfLite models, but to understand what are the tradeoffs when making them! model sizes, complexity,etc and to know how it'll limit the scale of the model to be developed for this task (I'm concerned what I'll end up training would be too large to be converted into a TfLite model or be accepted in the competition!</p>",
          "rawMarkdown": "thank you Alice, It's less the issue of generating TfLite models, but to understand what are the tradeoffs when making them! model sizes, complexity,etc and to know how it'll limit the scale of the model to be developed for this task (I'm concerned what I'll end up training would be too large to be converted into a TfLite model or be accepted in the competition!"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2261747,
      "author_name": "Wondering Alice",
      "author_url": "",
      "post_date": "2023-05-16T14:20:40",
      "content": "<p>Hi Rahul,</p>\n<p>Your best starting point would be the discussion and example notebooks for the previous ASL (Isolated sign recognition) competition. This has multiple examples of how to generate TfLite models</p>",
      "votes": 2,
      "replies": [
        {
          "id": 2262436,
          "author_name": "Rahul Namboodiri",
          "author_url": "",
          "post_date": "2023-05-16T23:33:38.587000",
          "content": "<p>thank you Alice, It's less the issue of generating TfLite models, but to understand what are the tradeoffs when making them! model sizes, complexity,etc and to know how it'll limit the scale of the model to be developed for this task (I'm concerned what I'll end up training would be too large to be converted into a TfLite model or be accepted in the competition!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2260827": "Hey community!\n\nI'm a bit new to TF-Lite model deployments, and have done a cursory reading on TF-Lite being ML/DL for edge devices. I've also done some reading on how to optimise TF models for TF-lite from the [tensorflow documentation](https://www.tensorflow.org/lite/performance/best_practices).\n\nI'd like to know a more comprehensive list of caveats and tradeoffs that need to be considered when building TF-lite models, I'm assuming most of these are based on power and VRAM sizes, but I'm not entirely sure. Also, what would be the best way to test model deployment efficiency as well for TF-lite models?\n",
    "2261747": "Hi Rahul,\n\nYour best starting point would be the discussion and example notebooks for the previous ASL (Isolated sign recognition) competition. This has multiple examples of how to generate TfLite models\n"
  }
}