RaspberryPi3(USB2.0)とLaptopPC(USB3.1)でGoogle Edge TPU Acceleratorを使用してMobileNet-SSD v2の動作スピードを検証してみました（MS-COCO）

TPU-MobilenetSSD　

１．はじめに

Edge TPU Accelerator を手に入れましたので、早速 NCS2 との性能差を比較するため、 MobileNet-SSD v2 (MS-COCO) を同期処理と非同期処理で実装してみました。

２．導入手順

とても簡単で一瞬で終わります。

Install_the_Edge_TPU_Python_library

$ wget http://storage.googleapis.com/cloud-iot-edge-pretrained-models/edgetpu_api.tar.gz
$ tar xzf edgetpu_api.tar.gz
$ cd python-tflite-source
$ bash ./install.sh

３．テスト用プログラムで味見

Run_test

# From the python-tflite-source directory
cd edgetpu/

python3 demo/classify_image.py \
--model test_data/mobilenet_v2_1.0_224_inat_bird_quant_edgetpu.tflite \
--label test_data/inat_bird_labels.txt \
--image test_data/parrot.jpg

ふーん、という感じ。

Results

---------------------------
Ara macao (Scarlet Macaw)
Score :  0.61328125
---------------------------
Platycercus elegans (Crimson Rosella)
Score :  0.15234375

参考URL
Built for the Edge TPU - Models - URL
Edge TPU API overview & demos - URL

４．検証結果

４−１．MobileNet-SSD v2 + RaspberryPi3 + USB 2.0 + TPU + Sync mode (同期処理)

データ・セットは 90クラス の MS-COCO です。

MobileNet-SSD-TPU-sync.py

import argparse
import platform
import numpy as np
import cv2
import time
from PIL import Image
from edgetpu.detection.engine import DetectionEngine


# Function to read labels from text files.
def ReadLabelFile(file_path):
  with open(file_path, 'r') as f:
    lines = f.readlines()
  ret = {}
  for line in lines:
    pair = line.strip().split(maxsplit=1)
    ret[int(pair[0])] = pair[1].strip()
  return ret


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", default="mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite", help="Path of the detection model.")
    parser.add_argument("--label", default="coco_labels.txt", help="Path of the labels file.")
    parser.add_argument("--usbcamno", type=int, default=0, help="USB Camera number.")
    args = parser.parse_args()

    fps = ""
    detectfps = ""
    framecount = 0
    detectframecount = 0
    time1 = 0
    time2 = 0
    box_color = (255, 128, 0)
    box_thickness = 1
    label_background_color = (125, 175, 75)
    label_text_color = (255, 255, 255)
    percentage = 0.0

    camera_width = 320
    camera_height = 240

    cap = cv2.VideoCapture(args.usbcamno)
    cap.set(cv2.CAP_PROP_FPS, 150)
    cap.set(cv2.CAP_PROP_FRAME_WIDTH, camera_width)
    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, camera_height)

    # Initialize engine.
    engine = DetectionEngine(args.model)
    labels = ReadLabelFile(args.label) if args.label else None

    while True:
        t1 = time.perf_counter()

        ret, color_image = cap.read()
        if not ret:
            break

        # Run inference.
        prepimg = color_image[:, :, ::-1].copy()
        prepimg = Image.fromarray(prepimg)

        tinf = time.perf_counter()
        ans = engine.DetectWithImage(prepimg, threshold=0.5, keep_aspect_ratio=True, relative_coord=False, top_k=10)
        print(time.perf_counter() - tinf, "sec")


        # Display result.
        if ans:
            detectframecount += 1
            for obj in ans:
                box = obj.bounding_box.flatten().tolist()
                box_left = int(box[0])
                box_top = int(box[1])
                box_right = int(box[2])
                box_bottom = int(box[3])
                cv2.rectangle(color_image, (box_left, box_top), (box_right, box_bottom), box_color, box_thickness)

                percentage = int(obj.score * 100)
                label_text = labels[obj.label_id] + " (" + str(percentage) + "%)" 

                label_size = cv2.getTextSize(label_text, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)[0]
                label_left = box_left
                label_top = box_top - label_size[1]
                if (label_top < 1):
                    label_top = 1
                label_right = label_left + label_size[0]
                label_bottom = label_top + label_size[1]
                cv2.rectangle(color_image, (label_left - 1, label_top - 1), (label_right + 1, label_bottom + 1), label_background_color, -1)
                cv2.putText(color_image, label_text, (label_left, label_bottom), cv2.FONT_HERSHEY_SIMPLEX, 0.5, label_text_color, 1)

        cv2.putText(color_image, fps,       (camera_width-170,15), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (38,0,255), 1, cv2.LINE_AA)
        cv2.putText(color_image, detectfps, (camera_width-170,30), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (38,0,255), 1, cv2.LINE_AA)

        cv2.namedWindow('USB Camera', cv2.WINDOW_AUTOSIZE)
        cv2.imshow('USB Camera', color_image)

        if cv2.waitKey(1)&0xFF == ord('q'):
            break

        # FPS calculation
        framecount += 1
        if framecount >= 15:
            fps       = "(Playback) {:.1f} FPS".format(time1/15)
            detectfps = "(Detection) {:.1f} FPS".format(detectframecount/time2)
            framecount = 0
            detectframecount = 0
            time1 = 0
            time2 = 0
        t2 = time.perf_counter()
        elapsedTime = t2-t1
        time1 += 1/elapsedTime
        time2 += elapsedTime

if __name__ == '__main__':
    main()

４−２．MobileNet-SSD v2 + RaspberryPi3 + USB 2.0 + TPU + Async mode (非同期マルチプロセス処理)

こちらも、データ・セットは 90クラス の MS-COCO です。

MobileNet-SSD-TPU-async.py

import argparse
import platform
import numpy as np
import cv2
import time
from PIL import Image
from time import sleep
import multiprocessing as mp
from edgetpu.detection.engine import DetectionEngine

lastresults = None
processes = []
frameBuffer = None
results = None
fps = ""
detectfps = ""
framecount = 0
detectframecount = 0
time1 = 0
time2 = 0
box_color = (255, 128, 0)
box_thickness = 1
label_background_color = (125, 175, 75)
label_text_color = (255, 255, 255)
percentage = 0.0

# Function to read labels from text files.
def ReadLabelFile(file_path):
  with open(file_path, 'r') as f:
    lines = f.readlines()
  ret = {}
  for line in lines:
    pair = line.strip().split(maxsplit=1)
    ret[int(pair[0])] = pair[1].strip()
  return ret


def camThread(label, results, frameBuffer, camera_width, camera_height, vidfps, usbcamno):

    global fps
    global detectfps
    global framecount
    global detectframecount
    global time1
    global time2
    global lastresults
    global cam
    global window_name

    cam = cv2.VideoCapture(usbcamno)
    cam.set(cv2.CAP_PROP_FPS, vidfps)
    cam.set(cv2.CAP_PROP_FRAME_WIDTH, camera_width)
    cam.set(cv2.CAP_PROP_FRAME_HEIGHT, camera_height)
    window_name = "USB Camera"
    cv2.namedWindow(window_name, cv2.WINDOW_AUTOSIZE)

    while True:
        t1 = time.perf_counter()

        ret, color_image = cam.read()
        if not ret:
            continue
        if frameBuffer.full():
            frameBuffer.get()
        frames = color_image
        frameBuffer.put(color_image.copy())
        res = None

        if not results.empty():
            res = results.get(False)
            detectframecount += 1
            imdraw = overlay_on_image(frames, res, label, camera_width, camera_height)
            lastresults = res
        else:
            imdraw = overlay_on_image(frames, lastresults, label, camera_width, camera_height)

        cv2.imshow('USB Camera', imdraw)

        if cv2.waitKey(1)&0xFF == ord('q'):
            break

        # FPS calculation
        framecount += 1
        if framecount >= 15:
            fps       = "(Playback) {:.1f} FPS".format(time1/15)
            detectfps = "(Detection) {:.1f} FPS".format(detectframecount/time2)
            framecount = 0
            detectframecount = 0
            time1 = 0
            time2 = 0
        t2 = time.perf_counter()
        elapsedTime = t2-t1
        time1 += 1/elapsedTime
        time2 += elapsedTime



def inferencer(results, frameBuffer, model, camera_width, camera_height):

    engine = DetectionEngine(model)

    while True:

        if frameBuffer.empty():
            continue

        # Run inference.
        color_image = frameBuffer.get()
        prepimg = color_image[:, :, ::-1].copy()
        prepimg = Image.fromarray(prepimg)

        tinf = time.perf_counter()
        ans = engine.DetectWithImage(prepimg, threshold=0.5, keep_aspect_ratio=True, relative_coord=False, top_k=10)
        print(time.perf_counter() - tinf, "sec")
        results.put(ans)



def overlay_on_image(frames, object_infos, label, camera_width, camera_height):

    color_image = frames

    if isinstance(object_infos, type(None)):
        return color_image
    img_cp = color_image.copy()

    for obj in object_infos:
        box = obj.bounding_box.flatten().tolist()
        box_left = int(box[0])
        box_top = int(box[1])
        box_right = int(box[2])
        box_bottom = int(box[3])
        cv2.rectangle(img_cp, (box_left, box_top), (box_right, box_bottom), box_color, box_thickness)

        percentage = int(obj.score * 100)
        label_text = label[obj.label_id] + " (" + str(percentage) + "%)" 

        label_size = cv2.getTextSize(label_text, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)[0]
        label_left = box_left
        label_top = box_top - label_size[1]
        if (label_top < 1):
            label_top = 1
        label_right = label_left + label_size[0]
        label_bottom = label_top + label_size[1]
        cv2.rectangle(img_cp, (label_left - 1, label_top - 1), (label_right + 1, label_bottom + 1), label_background_color, -1)
        cv2.putText(img_cp, label_text, (label_left, label_bottom), cv2.FONT_HERSHEY_SIMPLEX, 0.5, label_text_color, 1)

    cv2.putText(img_cp, fps,       (camera_width-170,15), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (38,0,255), 1, cv2.LINE_AA)
    cv2.putText(img_cp, detectfps, (camera_width-170,30), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (38,0,255), 1, cv2.LINE_AA)

    return img_cp

if __name__ == '__main__':

    parser = argparse.ArgumentParser()
    parser.add_argument("--model", default="mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite", help="Path of the detection model.")
    parser.add_argument("--label", default="coco_labels.txt", help="Path of the labels file.")
    parser.add_argument("--usbcamno", type=int, default=0, help="USB Camera number.")
    args = parser.parse_args()

    model    = args.model
    label    = ReadLabelFile(args.label)
    usbcamno = args.usbcamno

    camera_width = 320
    camera_height = 240
    vidfps = 30

    try:
        mp.set_start_method('forkserver')
        frameBuffer = mp.Queue(10)
        results = mp.Queue()

        # Start streaming
        p = mp.Process(target=camThread,
                       args=(label, results, frameBuffer, camera_width, camera_height, vidfps, usbcamno),
                       daemon=True)
        p.start()
        processes.append(p)

        # Activation of inferencer
        p = mp.Process(target=inferencer,
                       args=(results, frameBuffer, model, camera_width, camera_height),
                       daemon=True)
        p.start()
        processes.append(p)

        while True:
            sleep(1)

    finally:
        for p in range(len(processes)):
            processes[p].terminate()

４−３．MobileNet-SSD v2 + Core i7 + USB 3.1 + TPU + Async mode (非同期マルチプロセス処理)

同じモデルとデータセットですが、めちゃくちゃ速いです。。。
60 FPS - 80 FPS の間で揺らいでいますが、転送レートが上がるだけでココまで差がでるとは。。。
速すぎてフレームレートが人力では読めません。。。
12 ms から 15 ms で１フレームを推論できるようです。

５．おわりに

実装が悪いのか、RaspberryPi3 では、画像の転送部分で USB 2.0 がボトルネックになりすぎてかなり遅いです。
NCS2 の ４リクエスト並列処理のほうが２倍のパフォーマンスが出ていました。
もう少しプログラムのブラッシュアップが必要かもしれませんが、USB 3.0 で本来の性能が出ていそうです。
なお、 USB 3.1 の同期処理では 48 FPS 出ました。
USB 3.1 の非同期処理では上述のとおり 60 FPS - 80 FPS 出ました。

＜2019.03.12 追記＞
LattePanda Alpha Core m3 + USB 3.0 + Google Edge TPU Accelerator +
MobileNet-SSD v2 + Async mode 640x480
about 60 - 80 FPS
https://youtu.be/OFEQHCQ5MsM
https://twitter.com/PINTO03091/status/1105120574316044290

LattePanda Alpha Core m3 + USB 3.0 + Google Edge TPU Accelerator +
MobileNet-SSD v2 + Async mode 320x240
about 80 - 90 FPS
https://youtu.be/LERXuDXn0kY
https://twitter.com/PINTO03091/status/1105120075156152320