• Automation framework: CK
  
• Development repository: ck-tensorflow-codereef
  
• Source: GitHub
  
• Available command lines:
  
  • ck run program:object-detection-tflite-benchmark --cmd_key=default (META)
    
• Support for host OS: any
  
• Support for target OS: android, linux
  
• Tags: tensorflow-lite,object-detection,tflite,lang-cpp
  
• Template: Object Detection via TFLite (C++)
  
• How to get the stable version via the client:
  
  pip install cbench
  cb download program:object-detection-tflite-benchmark --version=1.0.0 --all
  ck run program:object-detection-tflite-benchmark
• How to get the development version:
  
  pip install ck
  ck pull repo --url=https://github.com/code-reef/ck-tensorflow-codereef
  ck run program:object-detection-tflite-benchmark
  

TensorFlow Lite (C++) object detection program

This program uses a statically linked TensorFlow Lite (C++) library and TensorFlow Lite SSD MobileNet models.

Note: It uses TensorFlow Lite realisation of Non Max Suppresion for custom operator.

Prerequisites

Repositories

$ ck pull repo:ck-tensorflow
$ ck pull repo:ck-mlperf

TensorFlow Lite library

$ ck install package:lib-tflite-1.13.1-src-static

Flatbuffers library

$ ck install package --tags=lib,flatbuffers

xOpenme library

$ ck install package --tags=lib,xopenme

TensorFlow models API

$ ck install ck-tensorflow:package:tensorflowmodel-api

or

$ ck install package --tags=tensorflowmodel,api

Python libraries

Numpy

$ ck install package --tags=lib,python-package,numpy

Pillow

$ ck install package --tags=lib,python-package,pillow

Matplotlib

$ ck install package --tags=lib,python-package,matplotlib

Python API for COCO-dataset

$ ck install package --tags=tool,coco

TensorFlow models

Install a TensorFlow model SSD MobileNet via:

$ ck install ck-tensorflow:package:-object-detection-ssd-mobilenet-v1-coco

or:

$ ck install ck-mlperf:package:model-tflite-mlperf-ssd-mobilenet

Datasets

$ ck install package --tags=dataset,object-detection,coco

NB: If you have previously installed the coco datasets, you should probably renew them:

$ ck refresh env:{dataset-env-uoa}

where dataset-env-uoa is one of the env identifiers returned by:

$ ck show env --tags=dataset,coco

Compiling

$ ck compile ck-tensorflow:program:object-detection-tflite

Running

$ ck run ck-tensorflow:program:object-detection-tflite

Program parameters

CK_BATCH_COUNT

The number of images to be processed.

Default: 1

$ ck run ck-tensorflow:program:object-detection-tflite --env.CK_BATCH_COUNT=100

CK_HOST_CPU_NUMBER_OF_PROCESSORS

The number of threads used by TF Lite library for job.

Default: 1

$ ck run ck-tensorflow:program:object-detection-tflite --env.CK_HOST_CPU_NUMBER_OF_PROCESSORS=2

FULL_REPORT or VERBOSE

Show additional info (FULL_REPORT) or complete additional info (VERBOSE)

Default: no

$ ck run ck-tensorflow:program:object-detection-tflite --env.FULL_REPORT=yes

or:

$ ck run ck-tensorflow:program:object-detection-tflite --env.VERBOSE=yes

USE_CUSTOM_NMS_SETTINGS

Enable tuning model settings (just for custom operator use).

Note: Are you sure you know to do?

Default: no

Available settings are:

MAX_CLASSES_PER_DETECTION (integer, >0)
MAX_DETECTIONS (integer, >0)
DETECTIONS_PER_CLASS (integer, >0)

NMS_SCORE_THRESHOLD (float, >=0.0)
NMS_IOU_THRESHOLD (float, >=0.0)
SCALE_H (float, >0.0)
SCALE_W (float, >0.0)
SCALE_X (float, >0.0)
SCALE_Y (float, >0.0)

Usage example:

$ ck run ck-tensorflow:program:object-detection-tflite \
    --env.CK_BATCH_COUNT=10 \
    --env.FULL_REPORT=yes \
    --env.USE_CUSTOM_NMS_SETTINGS=yes \
    --env.MAX_DETECTIONS=100 \
    --env.NMS_SCORE_THRESHOLD=0.25 \
    --env.NMS_IOU_THRESHOLD=0.7

• 1.0.0 (2019-12-30)

• .cm/desc.json (3 bytes)
• .cm/info.json (562 bytes)
• .cm/meta.json (5674 bytes)
• README.md (3427 bytes)
• detect.cpp (7027 bytes)
• includes/benchmark.h (18853 bytes)
• includes/detect.hpp (6671 bytes)
• includes/detection_postprocess.hpp (30988 bytes)
• includes/settings.h (17215 bytes)