[Class02 HW#3] OTX를 이용한 model training

[J-WBaek]

• 자기 이름의 issue ticket을 생성
  • e.g) Class02_hw3_OTX를_이용한_ModelTraining<본인이름>
  • Issue ticket을 자기에게 할당 assign yourself.
  • Milestone을 [Class01 HW#3] OTX를 이용한 model training로 설정
  • 제출
  • /class02/homework/hw03_template.md를 내 과제 디렉토리에 `README.md`로 복사 ``` mkdir -p intel-02/class02/homework/<내이름>/hw3 && cd $_ cp ../../hw03_template.md ./hw3/README.md ```
  • 새로운 workspace를 만들고 tool로 수집한 dataset으로 불량 분류 모델을 otx로 raining
  • 결과물 IR(xml, bin)을 찾아서 hw3/<model_name>/안에 복사
  • splitted_dataset의 ds_count ./splitted_dataset 2결과를 Dataset 구조에 작성
  • 4개의 classification model들을 training하고 정확도, FPS, training 소요시간, 사용한 batch 크기, learning rate, 추가로 조절한 parameter들을 테이블로 작성.
  • FPS를 어떻게 측정했는지를 기술
  • 생성한 PR의 Development 항목에 생성한 issue를 link
  • reviewer를 J-WBaek로 할당

[kgkorchamhrd]

#!/usr/bin/env python3

import os
import threading
from argparse import ArgumentParser
from queue import Empty, Queue
from time import sleep

import cv2
import numpy as np
import openvino as ov
#from openvino.inference_engine import IECore

from iotdemo import FactoryController, ColorDetector, MotionDetector

FORCE_STOP = False

def thread_cam1(qMotion:Queue)->None:
    # TODO: MotionDetector
    Mdetection = MotionDetector()
    Mdetection.load_preset('resources/motion.cfg')

    # TODO: Load and initialize OpenVINO
    core = ov.Core()
    model = core.read_model(model='resources/model.xml')
    compiled_model = core.compile_model(model=model, device_name='CPU')
    output_layer = compiled_model.output(0)

    # TODO: HW2 Open video clip resources/conveyor.mp4 instead of camera device.
    cap = cv2.VideoCapture('resources/conveyor.mp4')
    while not FORCE_STOP:
        sleep(0.03)
        _, frame = cap.read()
        if frame is None:
            break

        # TODO: HW2 Enqueue "VIDEO:Cam1 live", frame info
        qMotion.put( ("VIDEO:Cam1 live", frame) )
        # TODO: Motion detect
        detect = Mdetection.detect(frame)

        if detect is None:
            continue

        # TODO: Enqueue "VIDEO:Cam1 detected", detected info.
        qMotion.put( ("VIDEO:Cam2 detected", detect) )

        # pre-process
        detect = cv2.cvtColor(detect, cv2.COLOR_BGR2RGB)

        input_image = np.expand_dims(detect.transpose(2, 0, 1), 0)

        #detect = np.expand_dims(detect, 0)
        #detect = np.transpose(detect, [0, 3, 1, 2])

        # abnormal detect
        # frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        # reshaped = detected[:, :, [2, 1, 0]]
        # np_data = np.moveaxis(reshaped, -1, 0)
        # preprocessed_numpy = [((np_data / 255.0) - 0.5) * 2]
        # batch_tensor = np.stack(preprocessed_numpy, axis=0)
# 
        # TODO: Inference OpenVINO
        predict = compiled_model(input_image)[output_layer]
        x_ratio, circle_ratio = predict[0]

        # TODO: Calculate ratios
        print(f"X = {x_ratio*100:.2f}%, Circle = {circle_ratio*100:.2f}%")

        # TODO: in queue for moving the actuator 1
        if x_ratio > circle_ratio:
            qMotion.put( ("PUSH", 1) )

    cap.release()
    qMotion.put(('DONE', None))
    exit()

def thread_cam2(qColor:Queue)->None:
    # TODO: MotionDetector
    Mdetection = MotionDetector()
    Mdetection.load_preset('resources/motion.cfg')
    # TODO: ColorDetector
    Cdetection = ColorDetector()
    Cdetection.load_preset('resources/color.cfg')
    # TODO: HW2 Open "resources/conveyor.mp4" video clip
    cap = cv2.VideoCapture('resources/conveyor.mp4')
    while not FORCE_STOP:
        sleep(0.03)
        _, frame = cap.read()
        if frame is None:
            break
        # TODO: HW2 Enqueue "VIDEO:Cam2 live", frame info
        qColor.put( ("VIDEO:Cam2 live", frame) )
        # TODO: Detect motion
        detect = Mdetection.detect(frame)

        if detect is None:
            continue

        # TODO: Enqueue "VIDEO:Cam2 detected", detected info.
        qColor.put( ("VIDEO:Cam2 detected", detect) )
        # TODO: Detect color
        predict = Cdetection.detect(detect)
        name, ratio = predict[0]
        # TODO: Compute ratio
        print(f"{name}: {ratio*100:.2f}%")

        # TODO: Enqueue to handle actuator 2
        if name == 'blue':
            qColor.put( ("PUSH", 2) )

    cap.release()
    qColor.put(('DONE', None))
    exit()

def imshow(title, frame, pos=None):
    cv2.namedWindow(title)
    if pos:
        cv2.moveWindow(title, pos[0], pos[1])
    cv2.imshow(title, frame)

def main():
    global FORCE_STOP

    parser = ArgumentParser(prog='python3 factory.py',
                            description="Factory tool")

    parser.add_argument("-d",
                        "--device",
                        default=None,
                        type=str,
                        help="Arduino port")
    args = parser.parse_args()

    # TODO: HW2 Create a Queue
    queue = Queue()
    # TODO: HW2 Create thread_cam1 and thread_cam2 threads and start them.
    Mdetection = threading.Thread(target = thread_cam1, args = (queue,))
    Cdetection = threading.Thread(target = thread_cam2, args = (queue,))

    Mdetection.start()
    Cdetection.start()

    with FactoryController(args.device) as ctrl:
        while not FORCE_STOP:
            if cv2.waitKey(10) & 0xff == ord('q'):
                break

            # TODO: HW2 get an item from the queue. You might need to properly handle exceptions.
            try:
               event = queue.get_nowait()
            except Empty:
                continue
            # de-queue name and data
            name, data = event

            # TODO: HW2 show videos with titles of 'Cam1 live' 
            # and 'Cam2 live' respectively. "VIDEO:Cam2 live"
            if name[11:] == "live":
                imshow(name[6:], data)
            # TODO: Control actuator, name == 'PUSH'
            elif name == "PUSH":
                ctrl.push_actuator(data)
            elif name == 'DONE':
                FORCE_STOP = True

            queue.task_done()

    Mdetection.join()
    Cdetection.join()
    cv2.destroyAllWindows()

if __name__ == '__main__':
    try:
        main()
    except Exception:
        os._exit()