fx

Step-by-Step

This document describes the step-by-step instructions for reproducing PyTorch PeleeNet quantization and benchmarking results with Intel® Neural Compressor.

Note

• PyTorch quantization implementation in imperative path has limitations on automatically execution. It's required to manually add QuantStub and DequantStub for quantizable ops, also to manually do fusion operation.
• Intel® Neural Compressor supposes user have done these two steps before invoking Intel® Neural Compressor interface. For details, please refer to https://pytorch.org/docs/stable/quantization.html

Prerequisite

1. Environment

Python 3.6 or higher version is recommended. The dependent packages are all in requirements, please install as following.

cd examples/pytorch/image_recognition/peleenet/quantization/ptq/fx
pip install -r requirements.txt

2. Prepare Dataset

Download ImageNet Raw image to dir: /path/to/imagenet. The dir includes below folder:

ls /path/to/imagenet
train  val

3. Prepare Pretrained Model

Download weights to examples/pytorch/image_recognition/peleenet/quantization/ptq/fx/weights.

Run

1. Quantization

cd examples/pytorch/image_recognition/peleenet/quantization/ptq/fx
python main.py --tune --pretrained -j 1 /path/to/imagenet --weights weights/peleenet_acc7208.pth.tar
or
```shell
sh run_quant.sh --dataset_location=/path/to/imagenet --input_model=weights/peleenet_acc7208.pth.tar

2. Benchmark

# int8
sh run_benchmark.sh --dataset_location=/path/to/imagenet --mode=performance --int8=true
# fp32
sh run_benchmark.sh --dataset_location=/path/to/imagenet --mode=performance --input_model=weights/peleenet_acc7208.pth.tar