Coral USB Accelerator

Posst date : 2020-03-07

エッジで推論を行うために設計された Google の専用 ASIC，Google Edge TPU (Tensor Processing Unit)，を用いた USB デバイス

Linux (debian, ubuntu, Raspbian), Mac, Windows で利用可能

Coral USB Accelerator (https://coral.ai/products/accelerator/)
Edge TPU (https://cloud.google.com/edge-tpu/)

主として次のサイトに従って，ubunutu18.04LTS (ThinkaPad X280 i5) にインストールを行った

https://coral.ai/docs/accelerator/get-started/

1. Install the Edge TPU runtime

2. Install the TensorFlow Lite library (https://www.tensorflow.org/lite/guide/python)

3. Download example codes and models (google-coral/tflite)

3-1. Image classification

3-2. Object detection example

4. posenet

注意：

RaspberryPi3B + Coral USB Accelerator

1. Install the Edge TPU runtime

$ wget -qO- https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -

OK

$ sudo apt-get update

ヒット:1 http://jp.archive.ubuntu.com/ubuntu bionic InRelease

ヒット:2 http://jp.archive.ubuntu.com/ubuntu bionic-updates InRelease

ヒット:3 http://jp.archive.ubuntu.com/ubuntu bionic-backports InRelease

ヒット:4 https://repo.skype.com/deb stable InRelease

取得:5 https://packages.cloud.google.com/apt coral-edgetpu-stable InRelease [6,332 B]

取得:6 https://packages.cloud.google.com/apt coral-edgetpu-stable/main amd64 Packages [1,390 B]

取得:7 https://packages.cloud.google.com/apt coral-edgetpu-stable/main i386 Packages [673 B]

取得:8 http://security.ubuntu.com/ubuntu bionic-security InRelease [88.7 kB]

ヒット:9 http://archive.ubuntu.com/ubuntu bionic InRelease

97.1 kB を 6秒 で取得しました (15.1 kB/s)

パッケージリストを読み込んでいます... 完了

$ sudo apt-get install libedgetpu1-std

 v.13.0

$ lsusb

 Bus 003 Device 003: ID 18d1:9302 Google Inc.

怖いので libedgetpu1-max (max freq. 版)は使わない

2. Install the TensorFlow Lite library (https://www.tensorflow.org/lite/guide/python)

python3 のバージョンを確認

$ python3

Python 3.6.9 (default, Nov  7 2019, 10:44:02)

[GCC 8.3.0] on linux

Type "help", "copyright", "credits" or "license" for more information.

>>>

python3 のバージョンに合った runtime をインストール

$ pip3 install https://dl.google.com/coral/python/tflite_runtime-2.1.0.post1-cp36-cp36m-linux_x86_64.whl

Collecting tflite-runtime==2.1.0.post1 from https://dl.google.com/coral/python/tflite_runtime-2.1.0.post1-cp36-cp36m-linux_x86_64.whl

  Downloading https://dl.google.com/coral/python/tflite_runtime-2.1.0.post1-cp36-cp36m-linux_x86_64.whl (1.9MB)

    100% |████████████████████████████████| 1.9MB 505kB/s

Collecting numpy>=1.12.1 (from tflite-runtime==2.1.0.post1)

  Using cached https://files.pythonhosted.org/packages/62/20/4d43e141b5bc426ba38274933ef8e76e85c7adea2c321ecf9ebf7421cedf/numpy-1.18.1-cp36-cp36m-manylinux1_x86_64.whl

Installing collected packages: numpy, tflite-runtime

Successfully installed numpy-1.18.1 tflite-runtime-2.1.0.post1

3. Download example codes and models (google-coral/tflite)

$ git clone https://github.com/google-coral/tflite.git

Cloning into 'tflite'...

remote: Enumerating objects: 65, done.

remote: Counting objects: 100% (65/65), done.

remote: Compressing objects: 100% (34/34), done.

remote: Total 193 (delta 23), reused 59 (delta 19), pack-reused 128

Receiving objects: 100% (193/193), 675.91 KiB | 755.00 KiB/s, done.

Resolving deltas: 100% (69/69), done.

3-1. Image classification

$ cd tflite/python/examples/classification

$ bash install_requirements.sh

Collecting numpy

  Using cached https://files.pythonhosted.org/packages/62/20/4d43e141b5bc426ba38274933ef8e76e85c7adea2c321ecf9ebf7421cedf/numpy-1.18.1-cp36-cp36m-manylinux1_x86_64.whl

Collecting Pillow

  Downloading https://files.pythonhosted.org/packages/19/5e/23dcc0ce3cc2abe92efd3cd61d764bee6ccdf1b667a1fb566f45dc249953/Pillow-7.0.0-cp36-cp36m-manylinux1_x86_64.whl (2.1MB)

    100% |████████████████████████████████| 2.1MB 341kB/s

Installing collected packages: numpy, Pillow

Successfully installed Pillow-7.0.0 numpy-1.18.1

  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current

                                 Dload  Upload   Total   Spent    Left  Speed

100   189  100   189    0     0    282      0 --:--:-- --:--:-- --:--:--   282

100 3988k  100 3988k    0     0   280k      0  0:00:14  0:00:14 --:--:--  411k

100   181  100   181    0     0    301      0 --:--:-- --:--:-- --:--:--  176k

100 3448k  100 3448k    0     0   265k      0  0:00:12  0:00:12 --:--:--  308k

100   158  100   158    0     0    528      0 --:--:-- --:--:-- --:--:--   528

100 40895  100 40895    0     0  63304      0 --:--:-- --:--:-- --:--:--  670k

  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current

                                 Dload  Upload   Total   Spent    Left  Speed

100   148  100   148    0     0    266      0 --:--:-- --:--:-- --:--:--   266

100 3068k  100 3068k    0     0   472k      0  0:00:06  0:00:06 --:--:--  609k

$ python3 classify_image.py \

--model models/mobilenet_v2_1.0_224_inat_bird_quant_edgetpu.tflite \

--labels models/inat_bird_labels.txt \

--input images/parrot.jpg

----INFERENCE TIME----

Note: The first inference on Edge TPU is slow because it includes loading the model into Edge TPU memory.

12.6ms

4.3ms

4.1ms

4.1ms

4.3ms

-------RESULTS--------

Ara macao (Scarlet Macaw): 0.76172

3-2. Object detection example

$ cd tflite/python/examples/detection

$ ./install_requirements.sh

Collecting numpy

  Using cached https://files.pythonhosted.org/packages/62/20/4d43e141b5bc426ba38274933ef8e76e85c7adea2c321ecf9ebf7421cedf/numpy-1.18.1-cp36-cp36m-manylinux1_x86_64.whl

Collecting Pillow

  Using cached https://files.pythonhosted.org/packages/19/5e/23dcc0ce3cc2abe92efd3cd61d764bee6ccdf1b667a1fb566f45dc249953/Pillow-7.0.0-cp36-cp36m-manylinux1_x86_64.whl

Installing collected packages: numpy, Pillow

Successfully installed Pillow-7.0.0 numpy-1.18.1

./install_requirements.sh: 行 26: /proc/device-tree/model: そのようなファイルやディレクトリはありません

  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current

                                 Dload  Upload   Total   Spent    Left  Speed

100   192  100   192    0     0   1548      0 --:--:-- --:--:-- --:--:--  1548

100 6785k  100 6785k    0     0   788k      0  0:00:08  0:00:08 --:--:-- 1081k

100   184  100   184    0     0   4717      0 --:--:-- --:--:-- --:--:--  4717

100 6030k  100 6030k    0     0   813k      0  0:00:07  0:00:07 --:--:-- 1079k

100   153  100   153    0     0   3923      0 --:--:-- --:--:-- --:--:--  3923

100   930  100   930    0     0   2224      0 --:--:-- --:--:-- --:--:--     0

  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current

                                 Dload  Upload   Total   Spent    Left  Speed

100   154  100   154    0     0    312      0 --:--:-- --:--:-- --:--:--   313

100  918k  100  918k    0     0   443k      0  0:00:02  0:00:02 --:--:-- 1053k

$ python3 detect_image.py \

  --model models/mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite \

  --labels models/coco_labels.txt \

  --input images/grace_hopper.bmp \

  --output images/grace_hopper_processed.bmp

----INFERENCE TIME----

Note: The first inference is slow because it includes loading the model into Edge TPU memory.

28.14 ms

11.95 ms

12.14 ms

12.87 ms

12.63 ms

-------RESULTS--------

person

  id:     0

  score:  0.83984375

  bbox:   BBox(xmin=1, ymin=2, xmax=516, ymax=599)

tie

  id:     31

  score:  0.83984375

  bbox:   BBox(xmin=228, ymin=421, xmax=293, ymax=545)

##### without TPU #####

$ python3 detect_image.py \

  --model models/mobilenet_ssd_v2_coco_quant_postprocess.tflite \

  --labels models/coco_labels.txt \

  --input images/grace_hopper.bmp

 --labels models/coco_labels.txt   --input images/grace_hopper.bmp

----INFERENCE TIME----

Note: The first inference is slow because it includes loading the model into Edge TPU memory.

550.27 ms

557.99 ms

544.21 ms

542.66 ms

542.42 ms

-------RESULTS--------

person

  id:     0

  score:  0.83984375

  bbox:   BBox(xmin=2, ymin=5, xmax=513, ymax=596)

tie

  id:     31

  score:  0.83984375

  bbox:   BBox(xmin=228, ymin=418, xmax=293, ymax=544)

性能比 : 543.42/12.63 = 42.95

(https://coral.ai/examples/detect-image/)

4. posenet

(https://github.com/google-coral/project-posenet)

Pose Net Model : Pose detection

Pose detection in the browser (https://github.com/tensorflow/tfjs-models/tree/master/posenet)
TensorFlowLite : Pose estimation (https://www.tensorflow.org/lite/models/pose_estimation/overview)

$ sudo apt-get install python3-edgetpu

$ sh ./install_requirements.sh

パッケージリストを読み込んでいます... 完了

依存関係ツリーを作成しています

状態情報を読み取っています... 完了

python3-numpy はすでに最新バージョン (1:1.13.3-2ubuntu1) です。

gstreamer1.0-plugins-good はすでに最新バージョン (1.14.5-0ubuntu1~18.04.1) です。

gstreamer1.0-plugins-good は手動でインストールしたと設定されました。

python3-gi はすでに最新バージョン (3.26.1-2ubuntu1) です。

python3-gi は手動でインストールしたと設定されました。

gstreamer1.0-plugins-bad はすでに最新バージョン (1.14.5-0ubuntu1~18.04.1) です。

以下のパッケージが新たにインストールされます:

  python3-gst-1.0

アップグレード: 0 個、新規インストール: 1 個、削除: 0 個、保留: 0 個。

14.3 kB のアーカイブを取得する必要があります。

この操作後に追加で 64.5 kB のディスク容量が消費されます。

取得:1 http://jp.archive.ubuntu.com/ubuntu bionic-updates/universe amd64 python3-gst-1.0 amd64 1.14.5-0ubuntu1~18.04.1 [14.3 kB]

14.3 kB を 0秒 で取得しました (45.9 kB/s)

以前に未選択のパッケージ python3-gst-1.0 を選択しています。

(データベースを読み込んでいます ... 現在 305932 個のファイルとディレクトリがインストールされています。)

.../python3-gst-1.0_1.14.5-0ubuntu1~18.04.1_amd64.deb を展開する準備をしています ...

python3-gst-1.0 (1.14.5-0ubuntu1~18.04.1) を展開しています...

python3-gst-1.0 (1.14.5-0ubuntu1~18.04.1) を設定しています ...

libc-bin (2.27-3ubuntu1) のトリガを処理しています ...

Collecting svgwrite

  Downloading https://files.pythonhosted.org/packages/4f/2e/f36cfec1da6162055b884e6366074cff18475a9538559ceae0c0bc58e186/svgwrite-1.3.1-py2.py3-none-any.whl (67kB)

    100% |████████████████████████████████| 71kB 919kB/s

Collecting pyparsing>=2.0.1 (from svgwrite)

  Downloading https://files.pythonhosted.org/packages/5d/bc/1e58593167fade7b544bfe9502a26dc860940a79ab306e651e7f13be68c2/pyparsing-2.4.6-py2.py3-none-any.whl (67kB)

    100% |████████████████████████████████| 71kB 1.2MB/s

Installing collected packages: pyparsing, svgwrite

Successfully installed pyparsing-2.4.6 svgwrite-1.3.1

[ERROR]

$ python3 simple_pose.py

Traceback (most recent call last):

  File "simple_pose.py", line 18, in <module>

    from pose_engine import PoseEngine

  File "/home/ttks/src/Coral/tflite/python/examples/project-posenet/pose_engine.py", line 20, in <module>

    from edgetpu import __version__ as edgetpu_version

ModuleNotFoundError: No module named 'edgetpu'

[FIX] python3-edgetpu をインストール

$ apt-get install python3-edgetpu

python3-edgetpu_13.0_amd64.deb

$ python3 simple_pose.py

--2020-03-07 17:07:35--  https://upload.wikimedia.org/wikipedia/commons/thumb/3/38/Hindu_marriage_ceremony_offering.jpg/640px-Hindu_marriage_ceremony_offering.jpg

upload.wikimedia.org (upload.wikimedia.org) をDNSに問いあわせています... 103.102.166.240

upload.wikimedia.org (upload.wikimedia.org)|103.102.166.240|:443 に接続しています... 接続しました。

HTTP による接続要求を送信しました、応答を待っています... 200 OK

長さ: 83328 (81K) [image/jpeg]

`couple.jpg' に保存中

couple.jpg          100%[===================>]  81.38K   377KB/s    時間 0.2s

2020-03-07 17:07:36 (377 KB/s) - `couple.jpg' へ保存完了 [83328/83328]

Inference time: 20ms

Pose Score:  0.61885977

 nose                 x=210  y=152  score=1.0

 left eye             x=224  y=138  score=1.0

 right eye            x=199  y=136  score=1.0

 left ear             x=245  y=135  score=1.0

 right ear            x=183  y=129  score=0.8

 left shoulder        x=268  y=168  score=0.8

 right shoulder       x=161  y=172  score=1.0

 left elbow           x=282  y=255  score=0.6

 right elbow          x=154  y=254  score=0.9

 left wrist           x=236  y=333  score=0.7

 right wrist          x=163  y=301  score=0.6

 left hip             x=323  y=181  score=0.2

 right hip            x=191  y=251  score=0.0

 left knee            x=343  y=84   score=0.8

 right knee           x=162  y=295  score=0.0

 left ankle           x=318  y=174  score=0.1

 right ankle          x=167  y=309  score=0.0

Pose Score:  0.5783217

 nose                 x=398  y=145  score=1.0

 left eye             x=416  y=129  score=1.0

 right eye            x=382  y=127  score=1.0

 left ear             x=457  y=110  score=0.9

 right ear            x=370  y=121  score=0.2

 left shoulder        x=490  y=168  score=0.9

 right shoulder       x=362  y=150  score=0.8

 left elbow           x=466  y=291  score=0.8

 right elbow          x=329  y=246  score=0.7

 left wrist           x=339  y=303  score=0.9

 right wrist          x=234  y=331  score=0.5

 left hip             x=525  y=106  score=0.0

 right hip            x=361  y=163  score=0.2

 left knee            x=501  y=90   score=0.6

 right knee           x=353  y=183  score=0.1

 left ankle           x=503  y=177  score=0.0

 right ankle          x=308  y=177  score=0.1

$ python3 pose_camera.py

OK

注意：

posenet には python3-edgetpu が必要

References :

ganganの技術備忘録 Google-Edge TPU USB アクセラレータ
Qiita : [150 FPS ++] Coral Edge TPU Accelerator を３本突き刺して並列推論し超速のPosenet性能を獲得するー無駄な高性能の極みへー (@PINTO)
Qiita : Jetson Nano + EdgeTPU で爆速PoseNet (ラズパイとのパフォーマンス比較) (@rhene)
新しいRaspberry Pi 4、モデルBでのTensorFlow Liteのベンチマーク (Hanalili)， TensorFlowLite での Mobilenet v.1, 2 の性能比較)
Deep Learningアプリケーション開発 (7) TensorFlow Lite with Python on Raspberry Pi and Edge TPU

RaspberryPi3B + Coral USB Accelerator

インストール方法は Linux 用を参照 (ubuntu と同じ)

Install the edgeTPU
1. $ lsusb
  1. Bus 001 Device 007: ID 1a6e:089a Global Unichip Corp.
  2. Bus 001 Device 005: ID 04ca:004f Lite-On Technology Corp. SK-9020 keyboard
  3. Bus 001 Device 004: ID 1bcf:0007 Sunplus Innovation Technology Inc. Optical Mouse
  4. Bus 001 Device 006: ID 0424:7800 Standard Microsystems Corp.
  5. Bus 001 Device 003: ID 0424:2514 Standard Microsystems Corp. USB 2.0 Hub
  6. Bus 001 Device 002: ID 0424:2514 Standard Microsystems Corp. USB 2.0 Hub
  7. Bus 001 Device 001: ID 1d6b:0002 Linux Foundation 2.0 root hub
Install TensorFlow Lite
1. $ pip3 install https://dl.google.com/coral/python/tflite_runtime-2.1.0.post1-cp37-cp37m-linux_armv7l.whl
Examples
1. Image classification
  1. $ python3 classify_image.py --model models/mobilenet_v2_1.0_224_inat_bird_quant_edgetpu.tflite --labels models/inat_bird_labels.txt --input images/parrot.jpg
  2. ----INFERENCE TIME----
  3. Note: The first inference on Edge TPU is slow because it includes loading the model into Edge TPU memory.
  4. 117.3ms
  5. 11.9ms
  6. 11.9ms
  7. 11.7ms
  8. 12.1ms
  9. -------RESULTS--------
  10. Ara macao (Scarlet Macaw): 0.76562
2. Object detection
  1. ### with TPU ###
    1. $ python3 detect_image.py \
    2. > --model models/mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite \
    3. > --labels models/coco_labels.txt \
    4. > --input images/grace_hopper.bmp \
    5. > --output images/grace_hopper_processed.bmp
    6. ----INFERENCE TIME----
    7. Note: The first inference is slow because it includes loading the model into Edge TPU memory.
    8. 231.92 ms
    9. 55.53 ms
    10. 55.38 ms
    11. 52.67 ms
    12. 54.86 ms
    13. -------RESULTS--------
    14. person
    15. id: 0
    16. score: 0.83984375
    17. bbox: BBox(xmin=1, ymin=2, xmax=516, ymax=599)
    18. tie
    19. id: 31
    20. score: 0.83984375
    21. bbox: BBox(xmin=228, ymin=421, xmax=293, ymax=545)
  2. ### without TPU ###
    1. $ python3 detect_image.py --model models/mobilenet_ssd_v2_coco_quant_postprocess.tflite --labels models/coco_labels.txt --input images/grace_hopper.bmp
    2. ----INFERENCE TIME----
    3. Note: The first inference is slow because it includes loading the model into Edge TPU memory.
    4. 520.56 ms
    5. 508.28 ms
    6. 508.13 ms
    7. 508.34 ms
    8. 508.36 ms
    9. -------RESULTS--------
    10. person
    11. id: 0
    12. score: 0.83984375
    13. bbox: BBox(xmin=2, ymin=5, xmax=513, ymax=596)
    14. tie
    15. id: 31
    16. score: 0.83984375
    17. bbox: BBox(xmin=228, ymin=418, xmax=293, ymax=544)
posenet
1. simple_pose.py
  1. $ python3 simple_pose.py
  2. --2020-03-11 15:10:58-- https://upload.wikimedia.org/wikipedia/commons/thumb/3/38/Hindu_marriage_ceremony_offering.jpg/640px-Hindu_marriage_ceremony_offering.jpg
  3. upload.wikimedia.org (upload.wikimedia.org) をDNSに問いあわせています... 103.102.166.240
  4. upload.wikimedia.org (upload.wikimedia.org)|103.102.166.240|:443 に接続しています... 接続しました。
  5. HTTP による接続要求を送信しました、応答を待っています... 200 OK
  6. 長さ: 83328 (81K) [image/jpeg]
  7. `couple.jpg' に保存中
  9. couple.jpg 100%[===================>] 81.38K 366KB/s 時間 0.2s
  11. 2020-03-11 15:10:59 (366 KB/s) - `couple.jpg' へ保存完了 [83328/83328]
  13. Inference time: 119ms
  15. Pose Score: 0.61855656
  16. nose x=210 y=152 score=1.0
  17. left eye x=224 y=138 score=1.0
  18. right eye x=199 y=137 score=1.0
  19. left ear x=244 y=135 score=1.0
  20. right ear x=182 y=129 score=0.8
  21. left shoulder x=268 y=168 score=0.8
  22. right shoulder x=160 y=173 score=1.0
  23. left elbow x=282 y=255 score=0.6
  24. right elbow x=154 y=256 score=0.9
  25. left wrist x=236 y=333 score=0.8
  26. right wrist x=162 y=299 score=0.6
  27. left hip x=324 y=179 score=0.2
  28. right hip x=203 y=233 score=0.0
  29. left knee x=344 y=86 score=0.8
  30. right knee x=170 y=294 score=0.0
  31. left ankle x=316 y=175 score=0.1
  32. right ankle x=166 y=307 score=0.0
  34. Pose Score: 0.5867945
  35. nose x=398 y=145 score=1.0
  36. left eye x=416 y=128 score=1.0
  37. right eye x=382 y=127 score=1.0
  38. left ear x=457 y=110 score=0.9
  39. right ear x=370 y=120 score=0.2
  40. left shoulder x=491 y=169 score=0.9
  41. right shoulder x=362 y=151 score=0.8
  42. left elbow x=465 y=290 score=0.9
  43. right elbow x=328 y=246 score=0.8
  44. left wrist x=340 y=303 score=0.9
  45. right wrist x=233 y=331 score=0.5
  46. left hip x=525 y=106 score=0.0
  47. right hip x=361 y=161 score=0.3
  48. left knee x=501 y=90 score=0.6
  49. right knee x=350 y=167 score=0.1
  50. left ankle x=504 y=178 score=0.0
  51. right ankle x=306 y=176 score=0.1
2. pose_camera.py
  1. $ python3 pose_camera.py
    1. PoseNet 102ms (9.8fps) TrueFPS:8fps

Google Sites

Report abuse