Base-quantization

accomplished

• 2022.11.2
  • Base QAT, PTQ
  • Per_tensor, Per_channel
  • Minmax, EmaMinmax, Histogram, omse, adaround, bias_correction
  • dorefa, lsq

Requirements

Python > 3.6 + Pytorch >= 1.6

Usage

base quazation

such as QAT per_layer

python main.py --type QAT --level L

You can change --type and --level to choose different quazation method

adaround

python main.py --type PTQ --adaround --level L

bias_correction

python main.py --type PTQ --bias_correction --level L

Histogram

python main.py --type PTQ --Histogram --level L

omse

python main.py --type PTQ --omse --level L

dorefa

python main.py --type QAT --dorefa --level L

lsq

python main.py --type QAT --lsq --level L

Example

Models	VGG_S
QAT_8	99.4
dorefa_4_32	99.46
dorefa_6_32	99.56
LSQ_8	99.48
PTQ_bias_correction	99.43
PTQ_adaround	99.41

Note

• LSQ need pretrained model to inital scale
• Because MNIST is small, different quanzation methods all can compress the model without accuracy drop
• The code structure is simple and easy to expand other sota quanzation methods