Pytorch Pose Estimation

Setup AI Model Efficiency Toolkit (AIMET)

Please install and setup AIMET before proceeding further. This model was tested with the torch_gpu variant of AIMET 1.22.2.

Package	Version
pycocotools	2.0.2
scipy	1.1.0

If you are using a docker image, e.g. AIMET development docker, please add the following lines to the Dockerfile and rebuild the Docker image

RUN pip install git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI
RUN pip install scipy==1.1.0

export PYTHONPATH=$PYTHONPATH:<aimet_model_zoo_path>

Downloading model checkpoints is handled by evaluation script pose_estimation_quanteval.py. Configuration is set to default by evaluation script pose_estimation_quanteval.py.
The pose estimation model checkpoints can be downloaded here:
- Pose Estimation pytorch model
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).

This evaluation script is built to evaluate on COCO2014 validation images with person keypoints.
COCO dataset can be downloaded from here:
- COCO 2014 Val images
- COCO 2014 Train/Val annotations
The COCO dataset path should include coco images and annotations. It assumes a folder structure containing two subdirectories: images/val2014 and annotations. Corresponding images and annotations should be put into the two subdirectories.

To run evaluation with QuantSim in AIMET, use the following

python pose_estimation_quanteval.py \
  --dataset-path < Path to COCO 2014 dataset> \
  --use-cuda < Run evaluation using GPU >

The results reported was evaluation on the whole dataset, which contains over 40k images and takes ~5 hours on a single RTX 2080Ti GPU. So in case you want to run a faster evaluation, specifiy num_imgs argument to the second call with a small number to evaluate_session so that you run evaluation only on a partial dataset.