Skip to content

Latest commit

 

History

History
53 lines (42 loc) · 2.99 KB

PoseEstimation.md

File metadata and controls

53 lines (42 loc) · 2.99 KB

Tensorflow Pose Estimation

Setup AI Model Efficiency Toolkit (AIMET)

Please install and setup AIMET before proceeding further. This evaluation was run using AIMET 1.22.2 for TensorFlow 1.15 i.e. please set release_tag="1.22.2" and AIMET_VARIANT="tf_gpu_tf115" in the above instructions.

NOTE: This model is expected not to work with GPUs at or after NVIDIA 30-series (e.g. RTX 3050), as those bring a new architecture not fully compatible with TF 1.X.

Additional Dependencies

Package Version
pycocotools 2.0.2
scipy 1.1.0

Adding dependencies within Docker Image

  • 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

Model weights and configuration

  • Downloading model checkpoints and configuration file for quantization is handled by evalution script.
  • The pose estimation model can be downloaded from here:
  • This model has been compressed and its weights are optimized by applying DFQ (Data Free Quantization).

Dataset

  • This evaluation script is built to evaluate on COCO2014 validation images with person keypoints.
  • coco dataset can be downloaded from here:
  • 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.

Usage

  • To run evaluation with QuantSim in AIMET, use the following
python pose_estimation_quanteval.py \
  --dataset_path < Path to COCO 2014 dataset> \
  --num-imgs < number of images to evaluate of COCO 2014 validation dataset> \
  --model-to-eval < which model to evaluate, two options are available: 'fp32' for evaluating original fp32 model, 'int8' for evaluating int8 quantized model >
  • We only support evaluation on COCO 2014 val images with person keypoints.
  • The results reported was evaluation on the whole dataset, which contains over 40k images and takes 15+ hours on a single RTX 2080Ti GPU. To run partial evaluation on MSCOCO validation dataset, specify --num-imgs argument.

Quantization configuration

  • Weight quantization: 8 bits per tensor asymmetric quantization
  • Bias parameters are not quantized
  • Activation quantization: 8 bits, asymmetric quantization
  • Model inputs are quantized
  • 2K Images from COCO validation dataset are used as calibration dataset