Skip to content

Latest commit

 

History

History
executable file
·
98 lines (72 loc) · 3.79 KB

MobileDetEdgeTPU.md

File metadata and controls

executable file
·
98 lines (72 loc) · 3.79 KB
Raw

MobileDet-EdgeTPU

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 2.4 i.e. please set release_tag="1.22.2" and AIMET_VARIANT="tf_gpu" in the above instructions.

Experiment setup

Additional Dependencies

Package
tensorflow/models
pycocotools
tf_slim

Install dependent pip packages:

pip install pycocotools
pip install --upgrade tf_slim

Download and install protoc

protoc is a standalone binary for the Google protobuf compiler.

  • we use protoc in the version 3.14.0, download protoc-3.14.0-linux-x86_64.zip to ROOT_PATH
  • install
cd ROOT_PATH
unzip protoc-3.14.0-linux-x86_64.zip
cd protoc-3.14.0-linux-x86_64/bin
chmod +x protoc
export PATH=/ROOT_PATH/protoc-3.14.0-linux-x86_64/bin:$PATH

Also, you can take this reference for automatically download and installation.

Clone TensorFlow model zoo as the FP32 source

git clone https://github.com/tensorflow/models.git
git checkout master
cd models/research
protoc object_detection/protos/*.proto --python_out=.

Download model checkpoint for AIMET optimization

MobileDet-EdgeTPU FP32 pretrained checkpoint used for AIMET quantization can be downloaded from the Releases page.

Dataset: MSCOCO in the tfrecord format

TFRecord format of COCO dataset is needed. There are two options for download and process MSCOCO dataset:

cd models/research/object_detection/dataset_tools
./download_and_preprocess_mscoco.sh <mscoco_dir>
  • Option 2: If COCO dataset is already available or you want to download COCO dataset separately
    • COCO dataset can be download here: COCO
      • Please download the 2017 Version
    • create_coco_tf_record.py can be used to convert dataset into TFRecord
python object_detection/dataset_tools/create_coco_tf_record.py --logtostderr --include_masks --train_image_dir=./MSCOCO_PATH/images/train2017/ --val_image_dir=./MSCOCO_PATH/images/val2017/ --test_image_dir=./MSCOCO_PATH/images/test2017/ --train_annotations_file=./MSCOCO_PATH/annotations/instances_train2017.json --val_annotations_file=./MSCOCO_PATH/annotations/instances_val2017.json --testdev_annotations_file=./MSCOCO_PATH/annotations/image_info_test2017.json --output_dir=./OUTPUT_DIR/

Note: The --include_masks option must be used.

Usage

  • mobiledet_edgetpu_quanteval.py has two required arguments, an example usage is shown below
python mobiledet_edgetpu_quanteval.py --dataset-path <path to tfrecord dataset> --annotation-json-file <path to instances json file>/instances_val2017.json

Quantization configuration

In the evaluation script included, we have manually configured the quantizer ops with the following assumptions:

  • Weight quantization: 8 bits, per-tensor symmetric quantization
  • Bias parameters are not quantized
  • Activation quantization: 8 bits, asymmetric quantization
  • Model inputs are not quantized
  • TF was used for weight quantization scheme
  • TF was used for activation quantization scheme
  • Weights are optimzied by per-tensor Adaround in TF_enhanced scheme