Auto-Tune with Templates and AutoTVM

AutoTVM offers a way to tune models and operators by providing a template schedule, and searcing the parameter space defined by the template. These how-tos demonstrate how to write template schedules and optimize them for a variety of different hardware platforms.

• Tuning High Performance Convolution on NVIDIA GPUs
  • Install dependencies
  • Step 1: Define the search space
  • Step 2: Search through the space
• Auto-tuning a Convolutional Network for NVIDIA GPU
  • Install dependencies
  • Define Network
  • Set Tuning Options
  • Begin Tuning
  • Sample Output
  • Scale up measurement by using multiple devices
• Auto-tuning a Convolutional Network for x86 CPU
  • Define network
  • Configure tensor tuning settings and create tasks
  • Sample Output
• Auto-tuning a Convolutional Network for ARM CPU
  • Install dependencies
  • Define network
  • Start RPC Tracker
  • Register Devices to RPC Tracker
  • Set Tuning Options
  • Begin Tuning
  • Sample Output
• Auto-tuning a Convolutional Network for Mobile GPU
  • Install dependencies
  • Define network
  • Start RPC Tracker
  • Register Devices to RPC Tracker
  • Set Tuning Options
  • Begin Tuning
  • Sample Output