Overview

Description

DRP-AI for RZ/V2H supports a feature for efficiently calculating the pruned AI model. The DRP-AI Extension Pack provides a pruning function optimized for RZ/V2H. The DRP-AI optimized pruning function can be used in combination with this tool and PyTorch or TensorFlow training code.

What is pruning?

Nodes in a neural network are interconnected as shown in the figure. Methods of reducing the number of parameters by removing weights between nodes or removing nodes are referred to as “pruning”. A neural network to which pruning has not been applied is generally referred to as a dense neural network. And a neural network to which pruning has been applied is generally referred to as a sparse neural network. Applying pruning leads to a slight deterioration in the accuracy of the model but can reduce the power required by hardware and accelerate the inference process.

Image
Dense neural network; after pruning: sparse neural network

How to embed the pruned model

The pruned model can be embedded using DRP-AI TVM. Refer to the DRP-AI TVM page on GitHub for details on TVM.
https://github.com/renesas-rz/rzv_drp-ai_tvm

Note: As shown in the figure, pruning is an optional function. (Dense model also can be embedded.)

Image
DRP-AI Development Environment

Features

  • Pruning functions optimized for RZ/V2H
  • Pruning ratio can be specified for balance between accuracy and power efficiency
  • Supports 2 pruning modes for improving accuracy (One Shot/Gradual)

Target Devices

Downloads

Type Title Date
Software & Tools - Software Log in to Download GZ 3.32 MB
1 item

Documentation

Type Title Date
Manual - Software PDF 1.52 MB
1 item

Design & Development

Related Boards & Kits

Boards & Kits