Estimation of Google Edge TPU Accelerator

Estimates Edge TPU performace on some linux system such as RaspberryPi, Rock64@Pine64 etc.

As of now, Raspbian OS bases on Ubuntu Xenial(16.04 LTS) build as 32bit OS.

Rock64@Pine64 has compatible form factor of RaspberryPi 3 Model B(+), but Rock64@Pine64 supports many variant Destribution of Linux such as Ambian, Ubuntu, Debian etc. android too, and supports 32bit and 64bit OS.

RaspberryPi-3 has 4 USB-2.0 ports.
Rock64@Pine64 has 2 USB-2.0 ports and 1 USB-3.0 port.
Edge TPU expects using via USB-3.0 and on 64bit aarch64 OS, therefore Rock64@Pine64 is suitable for Edge TPU.

Getting start of Google Edge TPU Accelerator Coral.

According to Getting start,

Requirements
Any Linux computer with a USB port
Debian 6.0 or higher, or any derivative thereof (such as Ubuntu 10.0+)
System architecture of either x86_64 or ARM64 with ARMv8 instruction set
And yes, this means Raspberry Pi is supported. However, it must be Raspberry Pi 2/3 Model B/B+ running Raspbian (or another Debian derivative).

It's a very easy to install Edge TPU on Debian linux system but Windows OS is not supported.
Notice!: Needs Internet access from host system.

# apt update
$ cd ~/
$ wget https://dl.google.com/coral/edgetpu_api/edgetpu_api_latest.tar.gz -O edgetpu_api.tar.gz --trust-server-names --no-check-certificate
$ tar xzf edgetpu_api.tar.gz
$ cd edgetpu_api
$ bash ./install.sh

Investigating install.sh script we can found out that script supports bellow CPU architectures,

• armv7l
  When uname -m returns armv7l and system file /proc/device-tree/model denotes any words such as RaspberryPi-3, 32bit library is installed on somewhere of /usr/lib...

• x86_64 or aarch64
  When uname -m returns x86_64 or aarch64 64bit library is installed on somewhere of /usr/lib...
  But if system was 32bit OS on 64bit Hardware then 64bit library will be installed and may be system fall into crash!

Python API

According to Python API two key api are provided as ClassificationEngine and DetectionEngine.

Edge TPU Accelerator on RaspberryPi-3 Model B+

• Install OS image and boot
  Download and flash SDCard bellow,
  2018-10-09-raspbian-stretch.img

• Update system
  $ apt update
  $ apt upgrade

• Check python3 --version
  Python 3.5.3

• Install edgetpu_api according to geting start

• Run ClassificationEngine demo with parrot.jpg

  $ cd ~/Downloads/
  $ wget https://storage.googleapis.com/cloud-iot-edge-pretrained-models/canned_models/mobilenet_v2_1.0_224_inat_bird_quant_edgetpu.tflite \
    http://storage.googleapis.com/cloud-iot-edge-pretrained-models/canned_models/inat_bird_labels.txt \
    https://coral.withgoogle.com/static/images/parrot.jpg  --no-check-certificate
    
  $ cd /usr/local/lib/python3.5/dist-packages/edgetpu/demo
  $ python3 classify_image.py \
    --model ~/Downloads/mobilenet_v2_1.0_224_inat_bird_quant_edgetpu.tflite \
    --label ~/Downloads/inat_bird_labels.txt \
    --image ~/Downloads/parrot.jpg
    
     W0208 14:27:30.933504    1573 package_registry.cc:65] Minimum runtime
     version required by package (5) is lower than expected(10).
     ---------------------------
     Ara macao (Scarlet Macaw)
     Score :  0.761719 
  

This demo model can recognize 964 kinds of birds.
...
960 Anser anser domesticus (Domestic Goose)
961 Mitrephanes phaeocercus (Tufted Flycatcher)
962 Ardenna creatopus (Pink-footed Shearwater)
963 Ardenna gravis (Great Shearwater)
...
Ara macao means KONGO-INKO in japanese. It's a kind of parrot(bird).

• Run DetectionEngine demo with face.jpg

  $ cd ~/Downloads
  $ wget https://storage.googleapis.com/cloud-iot-edge-pretrained-models/canned_models/mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite https://coral.withgoogle.com/static/images/face.jpg  --no-check-certificate
  # apt install feh
  $ cd /usr/local/lib/python3.5/dist-packages/edgetpu/demo
  $ python3 object_detection.py \
    --model ~/Downloads/mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite \
    --input ~/Downloads/face.jpg --output ~/Downloads/detection_results.jpg
    W0208 15:14:29.736634    2141 package_registry.cc:65] Minimum runtime version required by package (5)
    is lower than expected (10).
-----------------------------------------
score =  0.996094
box =  [474.22854804992676, 38.03488787482766, 738.8013491630554, 353.5309683683231]
-----------------------------------------
score =  0.992188
box =  [205.4297697544098, 110.28378465056959, 487.75309658050537, 439.73802454331343]
-----------------------------------------
score =  0.832031
box =  [6.2277887016534805, 182.35811898071842, 127.13575917482376, 326.5376813379348]
-----------------------------------------
score =  0.5
box =  [859.8422718048096, 213.5472493581642, 1008.978108882904, 383.9367261515483]

feh : feh/stable,now 2.18-2 armhf imlib2 based image viewer
detection_result is,

Demo via csi-camera

• Setup MIPI-CSI Camera on RaspberryPI-3

• Test COCO Object Detection Demo script like bellow,

  $ cd RaspberryPi
  $ python3 demo_csi_ssd_mobilenet.py

• How to start Demo at immediately after starting X-Window session

Edge TPU Accelerator on Rock64@Pine64

Demo for Rock64

• Software download
  There are latest OS images supported by Rock64.

• Software github
  There are archives of OS historical images.

Rock64@Pine64 is ARM 64bit Hardware therefore with 64bit Debian Linux OS should be selected.

We tried some OS bellows,

• Armbian_5.75_Rock64_Debian_stretch_default_4.4.174_desktop
  Debian 9 stretch 64bit build with desktop environment.
  Default root/pass = root/1234

• stretch-minimal-rock64-0.7.8-1061-arm64
  Debian 9 stretch minimal build without desktop environment.
  Default root/pass = rock64/rock64

• bionic-lxde-rock64-0.8.0rc9-1120-arm64.img
  Ubuntu 18.06 64bit build with desktop environment.
  Default root/pass = rock64/rock64

Summary of installation of Edge TPU Accelerator

OS	File	Python	edgetpu_api	Edge TPU Status
bionic 18.04.2 LTS	bionic-lxde-rock64-0.8.0rc9-1120-arm64	3.6.12	19.2-py3	cannot open shared object file
Debian 9(stretch)	stretch-minimal-rock64-0.7.8-1061-arm64	3.5.3	19.2-py3	Work file 😄
Debian 9(stretch)	Armbian_5.75_Rock64_Debian_stretch_default_4.4.174_desktop	3.5.3	19.2-py3	Work file 😄

Using stretch-minimal-rock64-0.7.8-1061-arm64

OS image stretch-minimal bases on Debian 9 stretch without desktop environment.

To enable dhcp, jp106 keyboard etc.

• edit /etc/network/intefaces for DHCP if need.

 auto eth0
 allow-hotplug eth0
 iface eth0 inet dhcp

• Edit /etc/default/keyboard for jp.

 XKBMODEL="jp106"
 XKBLAYOUT="jp"

• How to check OS bit,

 $ file /bin/ls
 /bin/ls: ELF 64-bit LSB executable, x86-64, version 1 (SYSV), dynamically linked, interpreter /lib64/l,
 for GNU/Linux 2.6.32, BuildID[sha1]=d0bc0fb9b3f60f72bbad3c5a1d24c9e2a1fde775, stripped
    or 
 $ objdump -p /bin/ls
 /bin/ls:     file format elf64-x86-64

• Modify Destribution Environment
  • Official support OS is Debian 6 or later but not work on Ubuntu bionic. I gess that Edge TPU Accelerator needs Python 3.5.xx version reason why provided _edge_cpp_wrapper.so only is 35m.
    Look at installed edgetpu library /usr/local/lib/python3.5/dist-packages/edgetpu/swig directory,

-rw-r--r--  1 root staff 2232960 Jul 21 21:36 _edgetpu_cpp_wrapper.cpython-35m-aarch64-linux-gnu.so
-rw-r--r--  1 root staff 2089516 Jul 21 21:36 _edgetpu_cpp_wrapper.cpython-35m-arm-linux-gnueabihf.so
-rw-r--r--  1 root staff 1967728 Jul 21 21:36 _edgetpu_cpp_wrapper.cpython-35m-x86_64-linux-gnu.so
-rw-r--r--  1 root staff 1992232 Jul 21 21:36 _edgetpu_cpp_wrapper.cpython-36m-x86_64-linux-gnu.so
-rw-r--r--  1 root staff   16554 Jul 21 21:36 edgetpu_cpp_wrapper.py

It seems only Python3.5 or 3.6 are supported and on aarch64 supports 3.5 only now.

• lsb_release information

$ cat /etc/*release*
PRETTY_NAME="Debian GNU/Linux 9 (stretch)"
NAME="Debian GNU/Linux"
VERSION_ID="9"
VERSION="9 (stretch)"
ID=debian
HOME_URL="https://www.debian.org/"
SUPPORT_URL="https://www.debian.org/support"
BUG_REPORT_URL="https://bugs.debian.org/"

• To avoid SSL:Error of pip or pip3
  edit ~/.pip/pip.conf like bellow,

[global]
trusted-host = pypi.python.org
               pypi.org
               files.pythonhosted.org

• Install edgetpu_api according to geting start

• Run ClassificationEngine demo with parrot.jpg

 $ cd ~/Downloads/
 $ wget https://storage.googleapis.com/cloud-iot-edge-pretrained-models/canned_models/mobilenet_v2_1.0_224_inat_bird_quant_edgetpu.tflite \
    http://storage.googleapis.com/cloud-iot-edge-pretrained-models/canned_models/inat_bird_labels.txt \
    https://coral.withgoogle.com/static/images/parrot.jpg  --no-check-certificate
 $ cd /usr/local/lib/python3.5/dist-packages/edgetpu/demo
 $ python3 classify_image.py --model ~/Downloads/mobilenet_v2_1.0_224_inat_bird_quant_edgetpu.tflite \
--label ~/Downloads/inat_bird_labels.txt --image ~/Downloads/parrot.jpg
W0721 22:20:01.232883    3945 package_registry.cc:65] Minimum runtime version required by package (5)
is lower than expected (10).
---------------------------
Ara macao (Scarlet Macaw)
Score :  0.761719

• Run DetectionEngine demo with face.jpg

$ cd ~/Downloads
$ wget https://storage.googleapis.com/cloud-iot-edge-pretrained-models/canned_models/mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite https://coral.withgoogle.com/static/images/face.jpg  --no-check-certificate
# apt install feh
$ python3 object_detection.py \
    --model ~/Downloads/mobilenet_ssd_v2_face_quant_postprocess_edgetpu.tflite \
    --input ~/Downloads/face.jpg --output ~/Downloads/detection_results.jpg
W0721 22:26:39.865575    4098 package_registry.cc:65] Minimum runtime version required by package (5) is lower than expected (10).
-----------------------------------------
score =  0.996094
box =  [474.22854804992676, 38.03488787482766, 738.8013491630554, 353.5309683683231]
-----------------------------------------
score =  0.992188
box =  [205.4297697544098, 110.28378465056959, 487.75309658050537, 439.73802454331343]
-----------------------------------------
score =  0.832031
box =  [6.2277887016534805, 182.35811898071842, 127.13575917482376, 326.5376813379348]
-----------------------------------------
score =  0.5
box =  [859.8422718048096, 213.5472493581642, 1008.978108882904, 383.9367261515483]
Please check  /home/rock64/Downloads/detection_results.jpg
$ ls detection_result.jpg
  detection_results.jpg

• Install eog command.

    # apt install eog \
    xsever-xorg-video-all \
    x11-xserver-utils \
    xinit \
    xterm\
    twm \
    fonts-vlgothic \
    x11-apps
    $ eog detection_result.jpg

Good:smile:

I need to install OpenCV from source code because opencv-python package is not supported for python3 on stretch-minimal, can not use USB Camera via /dev/video0. Refer to Demo for Rock64

Using Armbian_5.75_Rock64_Debian_stretch_default_4.4.174_desktop

For Camera Demo i use Armbian stretch because it bases on Debian 9, has python3.5.3 with Desktop environment.

Performance Report

Comparison of Edge TPU via USB-2.0 with RaspberryPi ModelB+ versus via USB-3.0 with Rock64.

SBC	OS	TPU via USB	CameraIF	FPS
RaspberryPi ModelB+	32bit Raspbian Stretch	2.0	CSI	10
Rock64	64bit Armbian Stretch	3.0	USB	22