Custom Network Quantization Method for Lightweight CNN Acceleration on FPGAs

Algorithm 1

Framework of custom network quantization.

Input: Data, quantizers, pre-trained FP network with convolutional layers

Output: The quantized network inference model

1: Add quantizers before convolution operators;

2: for do

3:   Forward propagation by to weights of the network and by to activations of the network ;

4:   Backward propagation by STE to update network parameters;

5: end for

6: Add quantizers before non-convolution operators;

7: Re-train the network and subgraph fusion;

8:return quantized network inference model;