Optimize your TensorFlow Lite models
summary
TensorFlow Lite claims production use on over 4 billion devices and supplies concrete tools: Model Optimization Toolkit with post‑training and quantization‑aware APIs, weight pruning, and a benchmark binary that now reports memory and startup latency. Quantized MobileNet V2 shows a 39% speedup from runtime kernels, and Accent Pack yields about a 40% inference win on float MobileNet B2. Delegates include GPU OpenCL (beats OpenGL on Qualcomm), NNAPI, Hexagon, Core ML and Edge TPU, and pruned models can be quantized for extra gains.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:53What we enable
- 1:05Workflow at a glance
- 1:25Model discovery - TensorFlow Hub
- 2:54Post Training Quantization Example
- 3:21Quantization Results
- 3:56Weight Pruning (aka Sparsity)
- 4:52Benefits of Weight Pruning
- 5:47To get the Best Performance
- 6:37Android Profiler
- 7:28Benchmark Tool TFLite Command Line Benchmark Tool
- 9:07CPU: XNNPACK Deep Dive
- 10:33GPU Delegation
- 10:58Inference w/ GPU
- 11:10NNAPI Delegation
related talks
34mTensorFlow Lite
17mDeploy a custom Machine Learning model to mobile
32mAI for Mobile and IoT Devices: TensorFlow Lite
21mThings I learned running neural networks on microcontrollers
01h 01mHow to Optimize, Validate and Deploy ML Models on Device
59mBuilding models for TPU devices
42mWhat's New with TensorFlow?
39mCreating smaller, faster, production-worthy mobile machine learning models for Android