Skip to content

Latest commit

 

History

History
75 lines (66 loc) · 3.35 KB

HRNet-w48.md

File metadata and controls

75 lines (66 loc) · 3.35 KB

Pytorch HRNet-W48

Setup AI Model Efficiency Toolkit

Please install and setup AIMET before proceeding further. This model was tested with the torch_gpu variant of AIMET 1.21.0.

Experiment setup

  • Get HRNet pytorch-v1.1 branch: 
    git clone https://github.com/HRNet/HRNet-Semantic-Segmentation.git -b pytorch-v1.1
  • Makes sure additional dependencies pyyaml, yacs are installed: 
    pip install pyyaml
pip install 'yacs>=0.1.5'
  • Loading AIMET model zoo libraries
export PYTHONPATH=$PYTHONPATH:<aimet_model_zoo_path>

Model checkpoints and configuration

  • The original HRNet-W48 checkpoint can be downloaded from links provided at HRNet pytorch-v1.1.
  • Optimized HRNet checkpoint can be downloaded from the Releases page.
  • The Quantization Simulation (Quantsim) Configuration file can be downloaded from here: default_config_per_channel.json (Please see this page for more information on this file).
  • Downloading optimized checkpoints and quantization configuration file are also handled through evaluation script.

Dataset

  • This evaluation was designed for Cityscapes dataset, which can be downloaded through registration on https://www.cityscapes-dataset.com/.
  • After registration, go to https://www.cityscapes-dataset.com/downloads/ to download gtFine_trainvaltest.zip and leftImg8bit_trainvaltest.zip
  • Copy leftImg8bit_trainvaltest.zip and gtFine_trainvaltest.zip to $SEG_ROOT/data/cityscapes/ ($SEG_ROOT denotes path to HRNet-Semantic-Segmentation)

NOTE! Data has to be organized in the following way: 

$SEG_ROOT/data
├── cityscapes
│   ├── gtFine
│   │   ├── test
│   │   ├── train
│   │   └── val
│   └── leftImg8bit
│       ├── test
│       ├── train
│       └── val
├── list
│   ├── cityscapes
│   │   ├── test.lst
│   │   ├── trainval.lst
│   │   └── val.lst
│   ├── lip
│   │   ├── testvalList.txt
│   │   ├── trainList.txt
│   │   └── valList.txt

Usage

To run evaluation with QuantSim in AIMET, use the following

python hrnet-w48_quanteval.py \
	--default-param-bw <8|4> \
	--hrnet-path <Direct path way to HRnet github repo locally> \
	--use-cuda <Use GPU for evaluation>

Quantization Configuration

  • Weight quantization: 8 or 4 bits, per channel symmetric quantization
  • Bias parameters are not quantized
  • Activation quantization: 8 bits asymmetric quantization
  • Model inputs are quantized
  • TF_enhanced was used for weight quantization scheme
  • TF_enhanced was used for activation quantization scheme
  • Cross layer equalization and Adaround have been applied on optimized checkpoint
  • 2K Images from Cityscapes test dataset are used as calibration dataset