Esempio n. 1
0
    def start(self, cfg):
        """ """ """ """ """ """ """ """ """ """ """
        GET CONFIG
        """ """ """ """ """ """ """ """ """ """ """
        FORCE_GPU_COMPATIBLE = cfg['force_gpu_compatible']
        SAVE_TO_FILE = cfg['save_to_file']
        VISUALIZE = cfg['visualize']
        VIS_WORKER = cfg['vis_worker']
        VIS_TEXT = cfg['vis_text']
        MAX_FRAMES = cfg['max_frames']
        WIDTH = cfg['width']
        HEIGHT = cfg['height']
        FPS_INTERVAL = cfg['fps_interval']
        DET_INTERVAL = cfg['det_interval']
        DET_TH = cfg['det_th']
        SPLIT_MODEL = cfg['split_model']
        LOG_DEVICE = cfg['log_device']
        ALLOW_MEMORY_GROWTH = cfg['allow_memory_growth']
        SPLIT_SHAPE = cfg['split_shape']
        DEBUG_MODE = cfg['debug_mode']
        LABEL_PATH = cfg['label_path']
        NUM_CLASSES = cfg['num_classes']
        SRC_FROM = cfg['src_from']
        CAMERA = 0
        MOVIE = 1
        IMAGE = 2
        if SRC_FROM == 'camera':
            SRC_FROM = CAMERA
            VIDEO_INPUT = cfg['camera_input']
        elif SRC_FROM == 'movie':
            SRC_FROM = MOVIE
            VIDEO_INPUT = cfg['movie_input']
        elif SRC_FROM == 'image':
            SRC_FROM = IMAGE
            VIDEO_INPUT = cfg['image_input']
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        LOAD FROZEN_GRAPH
        """ """ """ """ """ """ """ """ """ """ """
        load_frozen_graph = LoadFrozenGraph(cfg)
        graph = load_frozen_graph.load_graph()
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        LOAD LABEL MAP
        """ """ """ """ """ """ """ """ """ """ """
        llm = LoadLabelMap()
        category_index = llm.load_label_map(cfg)
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        PREPARE TF CONFIG OPTION
        """ """ """ """ """ """ """ """ """ """ """
        # Session Config: allow seperate GPU/CPU adressing and limit memory allocation
        config = tf.ConfigProto(allow_soft_placement=True,
                                log_device_placement=LOG_DEVICE)
        config.gpu_options.allow_growth = ALLOW_MEMORY_GROWTH
        config.gpu_options.force_gpu_compatible = FORCE_GPU_COMPATIBLE
        #config.gpu_options.per_process_gpu_memory_fraction = 0.01 # 80MB memory is enough to run on TX2
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        PREPARE GRAPH I/O TO VARIABLE
        """ """ """ """ """ """ """ """ """ """ """
        # Define Input and Ouput tensors
        image_tensor = graph.get_tensor_by_name('image_tensor:0')
        detection_boxes = graph.get_tensor_by_name('detection_boxes:0')
        detection_scores = graph.get_tensor_by_name('detection_scores:0')
        detection_classes = graph.get_tensor_by_name('detection_classes:0')
        num_detections = graph.get_tensor_by_name('num_detections:0')

        if SPLIT_MODEL:
            SPLIT_TARGET_NAME = [
                'SecondStagePostprocessor/stack_1',
                'SecondStagePostprocessor/BatchMultiClassNonMaxSuppression/map/strided_slice',
                'BatchMultiClassNonMaxSuppression/map/TensorArrayStack_4/TensorArrayGatherV3',
                'Squeeze_2',
                'Squeeze_3',
                'SecondStagePostprocessor/Reshape_4',
            ]
            ''' BAD SPLIT POINT
            SPLIT_TARGET_OUT_NAME = ['ExpandDims_4',
                             'Preprocessor/map/TensorArrayStack_1/TensorArrayGatherV3',
                             'Shape',
                             'ExpandDims_1',
                             'Reshape_4',
                             'Reshape_3',
                             'FirstStageFeatureExtractor/InceptionV2/InceptionV2/Mixed_4e/concat',
                                 ]
            SPLIT_TARGET_IN_NAME = ['ExpandDims_4',
                             'Preprocessor/map/TensorArrayStack_1/TensorArrayGatherV3',
                             'Shape_6',
                             'ExpandDims_1',
                             'Reshape_4',
                             'Reshape_3',
                             'FirstStageFeatureExtractor/InceptionV2/InceptionV2/Mixed_4e/concat',
                                 ]

            '''

            split_out = []
            split_in = []
            for stn in SPLIT_TARGET_NAME:
                split_out += [graph.get_tensor_by_name(stn + ':0')]
                split_in += [graph.get_tensor_by_name(stn + '_1:0')]
            ''' BAD SPLIT POINT
            for stn in SPLIT_TARGET_OUT_NAME:
                split_out += [graph.get_tensor_by_name(stn+':0')]
            for stn in SPLIT_TARGET_IN_NAME:
                if stn == 'Shape_6':
                    split_in += [graph.get_tensor_by_name(stn+':0')]
                else:
                    split_in += [graph.get_tensor_by_name(stn+'_1:0')]
            '''
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        START WORKER THREAD
        """ """ """ """ """ """ """ """ """ """ """
        # gpu_worker uses in split_model and non-split_model
        gpu_tag = 'GPU'
        cpu_tag = 'CPU'
        gpu_worker = SessionWorker(gpu_tag, graph, config)
        if SPLIT_MODEL:
            gpu_opts = split_out
            cpu_worker = SessionWorker(cpu_tag, graph, config)
            cpu_opts = [
                detection_boxes, detection_scores, detection_classes,
                num_detections
            ]
        else:
            gpu_opts = [
                detection_boxes, detection_scores, detection_classes,
                num_detections
            ]
        """ """
        """
        START VISUALIZE WORKER
        """
        if VISUALIZE and VIS_WORKER:
            q_out = Queue.Queue()
            vis_worker = MPVisualizeWorker(cfg, MPVariable.vis_in_con)
            """ """ """ """ """ """ """ """ """ """ """
            START SENDER THREAD
            """ """ """ """ """ """ """ """ """ """ """
            start_sender(MPVariable.det_out_con, q_out)
        proc_frame_counter = 0
        vis_proc_time = 0
        """ """ """ """ """ """ """ """ """ """ """
        WAIT UNTIL THE FIRST DUMMY IMAGE DONE
        """ """ """ """ """ """ """ """ """ """ """
        print('Loading...')
        sleep_interval = 0.1
        """
        PUT DUMMY DATA INTO GPU WORKER
        """
        '''
        gpu_feeds = {image_tensor:  [np.zeros((300, 300, 3))]}
        gpu_extras = {}
        gpu_worker.put_sess_queue(gpu_opts, gpu_feeds, gpu_extras)
        if SPLIT_MODEL:
            """
            PUT DUMMY DATA INTO CPU WORKER
            """
            cpu_feeds = {split_in[1]: np.zeros((1))}
            cpu_extras = {}
            cpu_worker.put_sess_queue(cpu_opts, cpu_feeds, cpu_extras)
        """
        WAIT UNTIL JIT-COMPILE DONE
        """
        while True:
            g = gpu_worker.get_result_queue()
            if g is None:
                time.sleep(sleep_interval)
            else:
                break
        if SPLIT_MODEL:
            while True:
                c = cpu_worker.get_result_queue()
                if c is None:
                    time.sleep(sleep_interval)
                else:
                    break
        """ """
        '''
        """ """ """ """ """ """ """ """ """ """ """
        START CAMERA
        """ """ """ """ """ """ """ """ """ """ """
        if SRC_FROM == CAMERA:
            from lib.webcam import WebcamVideoStream as VideoReader
        elif SRC_FROM == MOVIE:
            from lib.video import VideoReader
        elif SRC_FROM == IMAGE:
            from lib.image import ImageReader as VideoReader
        video_reader = VideoReader()

        if SRC_FROM == IMAGE:
            video_reader.start(VIDEO_INPUT, save_to_file=SAVE_TO_FILE)
        else:  # CAMERA, MOVIE
            video_reader.start(VIDEO_INPUT,
                               WIDTH,
                               HEIGHT,
                               save_to_file=SAVE_TO_FILE)
            frame_cols, frame_rows = video_reader.getSize()
            """ STATISTICS FONT """
            fontScale = frame_rows / 1000.0
            if fontScale < 0.4:
                fontScale = 0.4
            fontThickness = 1 + int(fontScale)
        fontFace = cv2.FONT_HERSHEY_SIMPLEX
        if SRC_FROM == MOVIE:
            dir_path, filename = os.path.split(VIDEO_INPUT)
            filepath_prefix = filename
        elif SRC_FROM == CAMERA:
            filepath_prefix = 'frame'
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        DETECTION LOOP
        """ """ """ """ """ """ """ """ """ """ """
        print('Starting Detection')
        sleep_interval = 0.005
        top_in_time = None
        frame_in_processing_counter = 0
        try:
            if not video_reader.running:
                raise IOError(("Input src error."))
            while MPVariable.running.value:
                if top_in_time is None:
                    top_in_time = time.time()
                """
                SPRIT/NON-SPLIT MODEL CAMERA TO WORKER
                """
                if video_reader.running:
                    if gpu_worker.is_sess_empty():  # must need for speed
                        cap_in_time = time.time()
                        if SRC_FROM == IMAGE:
                            frame, filepath = video_reader.read()
                            if frame is not None:
                                frame_in_processing_counter += 1
                        else:
                            frame = video_reader.read()
                            if frame is not None:
                                filepath = filepath_prefix + '_' + str(
                                    proc_frame_counter) + '.png'
                                frame_in_processing_counter += 1
                        if frame is not None:
                            image_expanded = np.expand_dims(
                                cv2.cvtColor(frame, cv2.COLOR_BGR2RGB),
                                axis=0)  # np.expand_dims is faster than []
                            #image_expanded = np.expand_dims(frame, axis=0) # BGR image for input. Of couse, bad accuracy in RGB trained model, but speed up.
                            cap_out_time = time.time()
                            # put new queue
                            gpu_feeds = {image_tensor: image_expanded}
                            gpu_extras = {
                                'image': frame,
                                'top_in_time': top_in_time,
                                'cap_in_time': cap_in_time,
                                'cap_out_time': cap_out_time,
                                'filepath': filepath
                            }  # always image draw.
                            gpu_worker.put_sess_queue(gpu_opts, gpu_feeds,
                                                      gpu_extras)
                elif frame_in_processing_counter <= 0:
                    MPVariable.running.value = False
                    break

                g = gpu_worker.get_result_queue()
                if SPLIT_MODEL:
                    # if g is None: gpu thread has no output queue. ok skip, let's check cpu thread.
                    if g is not None:
                        # gpu thread has output queue.
                        result_slice_out, extras = g['results'], g['extras']
                        if cpu_worker.is_sess_empty():
                            # When cpu thread has no next queue, put new queue.
                            # else, drop gpu queue.
                            cpu_feeds = {}
                            for i in range(len(result_slice_out)):
                                cpu_feeds.update(
                                    {split_in[i]: result_slice_out[i]})
                            cpu_extras = extras
                            cpu_worker.put_sess_queue(cpu_opts, cpu_feeds,
                                                      cpu_extras)
                        else:
                            # else: cpu thread is busy. don't put new queue. let's check cpu result queue.
                            frame_in_processing_counter -= 1
                    # check cpu thread.
                    q = cpu_worker.get_result_queue()
                else:
                    """
                    NON-SPLIT MODEL
                    """
                    q = g
                if q is None:
                    """
                    SPLIT/NON-SPLIT MODEL
                    """
                    # detection is not complete yet. ok nothing to do.
                    time.sleep(sleep_interval)
                    continue

                frame_in_processing_counter -= 1
                boxes, scores, classes, num, extras = q['results'][0], q[
                    'results'][1], q['results'][2], q['results'][3], q[
                        'extras']
                det_out_time = time.time()
                """
                ALWAYS BOX DRAW ON IMAGE
                """
                vis_in_time = time.time()
                image = extras['image']
                if SRC_FROM == IMAGE:
                    filepath = extras['filepath']
                    frame_rows, frame_cols = image.shape[:2]
                    """ STATISTICS FONT """
                    fontScale = frame_rows / 1000.0
                    if fontScale < 0.4:
                        fontScale = 0.4
                    fontThickness = 1 + int(fontScale)
                else:
                    filepath = extras['filepath']
                image = visualization(category_index,
                                      image,
                                      boxes,
                                      scores,
                                      classes,
                                      DEBUG_MODE,
                                      VIS_TEXT,
                                      FPS_INTERVAL,
                                      fontFace=fontFace,
                                      fontScale=fontScale,
                                      fontThickness=fontThickness)
                """
                VISUALIZATION
                """
                if VISUALIZE:
                    if (MPVariable.vis_skip_rate.value
                            == 0) or (proc_frame_counter %
                                      MPVariable.vis_skip_rate.value < 1):
                        if VIS_WORKER:
                            q_out.put({
                                'image': image,
                                'vis_in_time': vis_in_time
                            })
                        else:
                            """
                            SHOW
                            """
                            cv2.imshow("Object Detection", image)
                            # Press q to quit
                            if cv2.waitKey(1) & 0xFF == 113:  #ord('q'):
                                break
                            MPVariable.vis_frame_counter.value += 1
                            vis_out_time = time.time()
                            """
                            PROCESSING TIME
                            """
                            vis_proc_time = vis_out_time - vis_in_time
                            MPVariable.vis_proc_time.value += vis_proc_time
                else:
                    """
                    NO VISUALIZE
                    """
                    for box, score, _class in zip(np.squeeze(boxes),
                                                  np.squeeze(scores),
                                                  np.squeeze(classes)):
                        if proc_frame_counter % DET_INTERVAL == 0 and score > DET_TH:
                            label = category_index[_class]['name']
                            print("label: {}\nscore: {}\nbox: {}".format(
                                label, score, box))

                    vis_out_time = time.time()
                    """
                    PROCESSING TIME
                    """
                    vis_proc_time = vis_out_time - vis_in_time

                if SAVE_TO_FILE:
                    if SRC_FROM == IMAGE:
                        video_reader.save(image, filepath)
                    else:
                        video_reader.save(image)

                proc_frame_counter += 1
                if proc_frame_counter > 100000:
                    proc_frame_counter = 0
                """
                PROCESSING TIME
                """
                top_in_time = extras['top_in_time']
                cap_proc_time = extras['cap_out_time'] - extras['cap_in_time']
                gpu_proc_time = extras[gpu_tag +
                                       '_out_time'] - extras[gpu_tag +
                                                             '_in_time']
                if SPLIT_MODEL:
                    cpu_proc_time = extras[cpu_tag +
                                           '_out_time'] - extras[cpu_tag +
                                                                 '_in_time']
                else:
                    cpu_proc_time = 0
                lost_proc_time = det_out_time - top_in_time - cap_proc_time - gpu_proc_time - cpu_proc_time
                total_proc_time = det_out_time - top_in_time
                MPVariable.cap_proc_time.value += cap_proc_time
                MPVariable.gpu_proc_time.value += gpu_proc_time
                MPVariable.cpu_proc_time.value += cpu_proc_time
                MPVariable.lost_proc_time.value += lost_proc_time
                MPVariable.total_proc_time.value += total_proc_time

                if DEBUG_MODE:
                    if SPLIT_MODEL:
                        sys.stdout.write(
                            'snapshot FPS:{: ^5.1f} total:{: ^10.5f} cap:{: ^10.5f} gpu:{: ^10.5f} cpu:{: ^10.5f} lost:{: ^10.5f} | vis:{: ^10.5f}\n'
                            .format(MPVariable.fps.value, total_proc_time,
                                    cap_proc_time, gpu_proc_time,
                                    cpu_proc_time, lost_proc_time,
                                    vis_proc_time))
                    else:
                        sys.stdout.write(
                            'snapshot FPS:{: ^5.1f} total:{: ^10.5f} cap:{: ^10.5f} gpu:{: ^10.5f} lost:{: ^10.5f} | vis:{: ^10.5f}\n'
                            .format(MPVariable.fps.value, total_proc_time,
                                    cap_proc_time, gpu_proc_time,
                                    lost_proc_time, vis_proc_time))
                """
                EXIT WITHOUT GUI
                """
                if not VISUALIZE and MAX_FRAMES > 0:
                    if proc_frame_counter >= MAX_FRAMES:
                        MPVariable.running.value = False
                        break
                """
                CHANGE SLEEP INTERVAL
                """
                if MPVariable.frame_counter.value == 0 and MPVariable.fps.value > 0:
                    sleep_interval = 0.1 / MPVariable.fps.value
                    MPVariable.sleep_interval.value = sleep_interval
                MPVariable.frame_counter.value += 1
                top_in_time = None
            """
            END while
            """
        except:
            import traceback
            traceback.print_exc()
        finally:
            """ """ """ """ """ """ """ """ """ """ """
            CLOSE
            """ """ """ """ """ """ """ """ """ """ """
            if VISUALIZE and VIS_WORKER:
                q_out.put(None)
            MPVariable.running.value = False
            gpu_worker.stop()
            if SPLIT_MODEL:
                cpu_worker.stop()
            video_reader.stop()

            if VISUALIZE:
                cv2.destroyAllWindows()
            """ """

        return
Esempio n. 2
0
    def start(self, cfg):
        """ """ """ """ """ """ """ """ """ """ """
        GET CONFIG
        """ """ """ """ """ """ """ """ """ """ """
        FORCE_GPU_COMPATIBLE = cfg['force_gpu_compatible']
        SAVE_TO_FILE         = cfg['save_to_file']
        VISUALIZE            = cfg['visualize']
        VIS_WORKER           = cfg['vis_worker']
        VIS_TEXT             = cfg['vis_text']
        MAX_FRAMES           = cfg['max_frames']
        WIDTH                = cfg['width']
        HEIGHT               = cfg['height']
        FPS_INTERVAL         = cfg['fps_interval']
        DET_INTERVAL         = cfg['det_interval']
        DET_TH               = cfg['det_th']
        LOG_DEVICE           = cfg['log_device']
        ALLOW_MEMORY_GROWTH  = cfg['allow_memory_growth']
        SPLIT_SHAPE          = cfg['split_shape']
        DEBUG_MODE           = cfg['debug_mode']
        LABEL_PATH           = cfg['label_path']
        NUM_CLASSES          = cfg['num_classes']
        SRC_FROM             = cfg['src_from']
        DEVICE               = eval(cfg['device'])
    
        MOVIE  = 1

        SRC_FROM = MOVIE
        VIDEO_INPUT = cfg['movie_input']
        """ """

        """ """ """ """ """ """ """ """ """ """ """
        LOAD FROZEN_GRAPH
        """ """ """ """ """ """ """ """ """ """ """
        load_frozen_graph = LoadFrozenGraph(cfg)
        graph = load_frozen_graph.load_graph()
        """ """

        """ """ """ """ """ """ """ """ """ """ """
        LOAD LABEL MAP
        """ """ """ """ """ """ """ """ """ """ """
        llm = LoadLabelMap()
        category_index = llm.load_label_map(cfg)
        """ """

        """ """ """ """ """ """ """ """ """ """ """
        PREPARE TF CONFIG OPTION
        """ """ """ """ """ """ """ """ """ """ """
        # Session Config: allow seperate GPU/CPU adressing and limit memory allocation
        config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True)
        config.gpu_options.allow_growth = ALLOW_MEMORY_GROWTH
        config.gpu_options.force_gpu_compatible = FORCE_GPU_COMPATIBLE
        #config.gpu_options.visible_device_list = '0,1,2,3' # For GPU SERVER
        #config.gpu_options.per_process_gpu_memory_fraction = 0.01 # 80MB memory is enough to run on TX2
        """ """
        
        """ """ """ """ """ """ """ """ """ """ """
        PREPARE GRAPH I/O TO VARIABLE
        """ """ """ """ """ """ """ """ """ """ """
        # Define Input and Ouput tensors
        image_tensor = graph.get_tensor_by_name('image_tensor:0')
        detection_boxes = graph.get_tensor_by_name('detection_boxes:0')
        detection_scores = graph.get_tensor_by_name('detection_scores:0')
        detection_classes = graph.get_tensor_by_name('detection_classes:0')
        

        """ """ """ """ """ """ """ """ """ """ """
        START WORKER THREAD
        """ """ """ """ """ """ """ """ """ """ """
        # gpu_worker uses in split_model and non-split_model
        gpu_tag = 'GPU'
        cpu_tag = 'CPU'
        gpu_worker = SessionWorker(gpu_tag, graph, config)
        gpu_opts = [detection_boxes, detection_scores, detection_classes]
        """ """

        """
        START VISUALIZE WORKER
        """
        if VISUALIZE and VIS_WORKER:
            q_out = Queue.Queue()
            vis_worker = MPVisualizeWorker(cfg, MPVariable.vis_in_con)
            """ """ """ """ """ """ """ """ """ """ """
            START SENDER THREAD
            """ """ """ """ """ """ """ """ """ """ """
            start_sender(MPVariable.det_out_con, q_out)
        proc_frame_counter = 0
        vis_proc_time = 0

        if VISUALIZE:
            cv2.namedWindow("Object Detection", 0)#The 0 represents the window type
            cv2.resizeWindow("Object Detection", WIDTH, HEIGHT)
        
        """ """ """ """ """ """ """ """ """ """ """
        START FRAME ACQUISITION
        """ """ """ """ """ """ """ """ """ """ """

        from lib.video import VideoReader
        video_reader = VideoReader()

        video_reader.start(VIDEO_INPUT, WIDTH, HEIGHT, save_to_file=SAVE_TO_FILE)
        frame_cols, frame_rows = video_reader.getSize()
        """ STATISTICS FONT """
        fontScale = frame_rows/1000.0
        if fontScale < 0.4:
            fontScale = 0.4
        fontThickness = 1 + int(fontScale)
        fontFace = cv2.FONT_HERSHEY_SIMPLEX

        dir_path, filename = os.path.split(VIDEO_INPUT)
        filepath_prefix = filename

        """ """


        """ """ """ """ """ """ """ """ """ """ """
        DETECTION LOOP
        """ """ """ """ """ """ """ """ """ """ """
        print('Starting Detection')
        pr = PowerRead(device = DEVICE.value)
        sleep_interval = 0.005
        top_in_time = None
        frame_in_processing_counter = 0
        start_time = time.time()
        try:
            if not video_reader.running:
                raise IOError(("Input src error."))
            while MPVariable.running.value:
                if top_in_time is None:
                    top_in_time = time.time()
                """
                SPRIT/NON-SPLIT MODEL CAMERA TO WORKER
                """
                if video_reader.running:
                    if gpu_worker.is_sess_empty(): # must need for speed
                        cap_in_time = time.time()
                        if SRC_FROM == IMAGE:
                            frame, filepath = video_reader.read()
                            if frame is not None:
                                frame_in_processing_counter += 1
                        else:
                            frame = video_reader.read()
                            if frame is not None:
                                filepath = filepath_prefix+'_'+str(proc_frame_counter)+'.png'
                                frame_in_processing_counter += 1
                        if frame is not None:
                            image_expanded = np.expand_dims(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB), axis=0) # np.expand_dims is faster than []
                            #image_expanded = np.expand_dims(frame, axis=0) # BGR image for input. Of couse, bad accuracy in RGB trained model, but speed up.
                            cap_out_time = time.time()
                            # put new queue
                            gpu_feeds = {image_tensor: image_expanded}
                            gpu_extras = {'image':frame, 'top_in_time':top_in_time, 'cap_in_time':cap_in_time, 'cap_out_time':cap_out_time, 'filepath': filepath} # always image draw.
                            gpu_worker.put_sess_queue(gpu_opts, gpu_feeds, gpu_extras)
                elif frame_in_processing_counter <= 0:
                    MPVariable.running.value = False
                    break

                g = gpu_worker.get_result_queue()
                if SPLIT_MODEL:
                    # if g is None: gpu thread has no output queue. ok skip, let's check cpu thread.
                    if g is not None:
                        # gpu thread has output queue.
                        result_slice_out, extras = g['results'], g['extras']
                        if cpu_worker.is_sess_empty():
                            # When cpu thread has no next queue, put new queue.
                            # else, drop gpu queue.
                            cpu_feeds = {}
                            for i in range(len(result_slice_out)):
                                cpu_feeds.update({split_in[i]:result_slice_out[i]})
                            cpu_extras = extras
                            cpu_worker.put_sess_queue(cpu_opts, cpu_feeds, cpu_extras)
                        else:
                            # else: cpu thread is busy. don't put new queue. let's check cpu result queue.
                            frame_in_processing_counter -= 1
                    # check cpu thread.
                    q = cpu_worker.get_result_queue()
                else:
                    """
                    NON-SPLIT MODEL
                    """
                    q = g
                if q is None:
                    """
                    SPLIT/NON-SPLIT MODEL
                    """
                    # detection is not complete yet. ok nothing to do.
                    time.sleep(sleep_interval)
                    continue

                frame_in_processing_counter -= 1
                boxes, scores, classes, extras = q['results'][0], q['results'][1], q['results'][2], q['extras']
                boxes, scores, classes = np.squeeze(boxes), np.squeeze(scores), np.squeeze(classes)
                det_out_time = time.time()

                """
                ALWAYS BOX DRAW ON IMAGE
                """
                vis_in_time = time.time()
                image = extras['image']
                if SRC_FROM == IMAGE:
                    filepath = extras['filepath']
                    frame_rows, frame_cols = image.shape[:2]
                    """ STATISTICS FONT """
                    fontScale = frame_rows/1000.0
                    if fontScale < 0.4:
                        fontScale = 0.4
                    fontThickness = 1 + int(fontScale)
                else:
                    filepath = extras['filepath']
                image = visualization(category_index, image, boxes, scores, classes, DEBUG_MODE, VIS_TEXT, FPS_INTERVAL,
                                      fontFace=fontFace, fontScale=fontScale, fontThickness=fontThickness)

                """
                VISUALIZATION
                """
                if VISUALIZE:
                    if (MPVariable.vis_skip_rate.value == 0) or (proc_frame_counter % MPVariable.vis_skip_rate.value < 1):
                        if VIS_WORKER:
                            q_out.put({'image':image, 'vis_in_time':vis_in_time})
                        else:
                            """
                            SHOW
                            """
                            cv2.imshow("Object Detection", image)
                            # Press q to quit
                            if cv2.waitKey(1) & 0xFF == 113: #ord('q'):
                                break
                            MPVariable.vis_frame_counter.value += 1
                            vis_out_time = time.time()
                            """
                            PROCESSING TIME
                            """
                            vis_proc_time = vis_out_time - vis_in_time
                            MPVariable.vis_proc_time.value += vis_proc_time
                else:
                    """
                    NO VISUALIZE
                    """
                    for box, score, _class in zip(boxes, scores, classes):
                        if proc_frame_counter % DET_INTERVAL == 0 and score > DET_TH:
                            label = category_index[_class]['name']
                            print("label: {}\nscore: {}\nbox: {}".format(label, score, box))

                    vis_out_time = time.time()
                    """
                    PROCESSING TIME
                    """
                    vis_proc_time = vis_out_time - vis_in_time

                if SAVE_TO_FILE:
                    if SRC_FROM == IMAGE:
                        video_reader.save(image, filepath)
                    else:
                        video_reader.save(image)

                proc_frame_counter += 1
                if proc_frame_counter > 100000:
                    proc_frame_counter = 0
                """
                PROCESSING TIME
                """
                top_in_time = extras['top_in_time']
                cap_proc_time = extras['cap_out_time'] - extras['cap_in_time']
                gpu_proc_time = extras[gpu_tag+'_out_time'] - extras[gpu_tag+'_in_time']
                if SPLIT_MODEL:
                    cpu_proc_time = extras[cpu_tag+'_out_time'] - extras[cpu_tag+'_in_time']
                else:
                    cpu_proc_time = 0
                lost_proc_time = det_out_time - top_in_time - cap_proc_time - gpu_proc_time - cpu_proc_time
                total_proc_time = det_out_time - top_in_time
                MPVariable.cap_proc_time.value += cap_proc_time
                MPVariable.gpu_proc_time.value += gpu_proc_time
                MPVariable.cpu_proc_time.value += cpu_proc_time
                MPVariable.lost_proc_time.value += lost_proc_time
                MPVariable.total_proc_time.value += total_proc_time
                
                with open("Processing_Times.txt", "a") as f:
                    f.write("{:f} {:f} {:f}\n".format(gpu_proc_time, vis_proc_time, total_proc_time))
                    f.close()
                if DEBUG_MODE:
                    if SPLIT_MODEL:
                        sys.stdout.write('snapshot FPS:{: ^5.1f} total:{: ^10.5f} cap:{: ^10.5f} gpu:{: ^10.5f} cpu:{: ^10.5f} lost:{: ^10.5f} | vis:{: ^10.5f}\n'.format(
                            MPVariable.fps.value, total_proc_time, cap_proc_time, gpu_proc_time, cpu_proc_time, lost_proc_time, vis_proc_time))
                    else:
                        sys.stdout.write('snapshot FPS:{: ^5.1f} total:{: ^10.5f} cap:{: ^10.5f} gpu:{: ^10.5f} lost:{: ^10.5f} | vis:{: ^10.5f}\n'.format(
                            MPVariable.fps.value, total_proc_time, cap_proc_time, gpu_proc_time, lost_proc_time, vis_proc_time))
                """
                EXIT WITHOUT GUI
                """
                if not VISUALIZE and MAX_FRAMES > 0:
                    if proc_frame_counter >= MAX_FRAMES:
                        MPVariable.running.value = False
                        break

                """
                CHANGE SLEEP INTERVAL
                """
                if MPVariable.frame_counter.value == 0 and MPVariable.fps.value > 0:
                    sleep_interval = 0.1 / MPVariable.fps.value
                    MPVariable.sleep_interval.value = sleep_interval
                MPVariable.frame_counter.value += 1
                top_in_time = None
            """
            END while
            """
        except KeyboardInterrupt:
            pass
        except:
            import traceback
            traceback.print_exc()
        finally:
            end_time = time.time() - start_time
            print("Total Time to Inference and Process Image: {:.3f} s".format(end_time))
            """ """ """ """ """ """ """ """ """ """ """
            CLOSE
            """ """ """ """ """ """ """ """ """ """ """
            if VISUALIZE and VIS_WORKER:
                q_out.put(None)
            MPVariable.running.value = False
            gpu_worker.stop()
            pr.stop()
            if SPLIT_MODEL:
                cpu_worker.stop()
            video_reader.stop()

            if VISUALIZE:
                cv2.destroyAllWindows()
            num_gpu = 1
            pw = []
            for n in range(num_gpu):
                with open("Power_Measurements_{}.txt".format(n), "r") as f:
                    pw += [[float(i) / 1000 for i in f]]
                    f.close()
                
            with open("Processing_Times.txt") as f:
                next(f)
                tm = [[float(x)*1000 for x in i.split()] for i in f]
                f.close()
            tm = np.array(tm)
            pw = np.array(pw)
            import matplotlib.pyplot as plt
            import matplotlib as mpl
#            from math import sqrt
            mpl.use('tkagg')
#            x = np.arange(0,len(pw))
            if DEVICE.value in (0, 1):
                pwr = []
                for n in range(num_gpu):
                    pwr += [[ x for x in pw[n] if 65 <= x <= 180]]
            elif DEVICE.value == 1:
                pwr = pw
            
            for n in range(num_gpu):
                #Power Measurements Plot
                fig, ax = plt.subplots()
    #            plt.stem(x, pw, use_line_collection = True)
                plt.ylabel('Frequency')
                plt.xlabel('GPU Power (W)')
                _, b, _ = plt.hist(pwr[n], bins = 'auto', edgecolor = 'black', 
                                   linewidth = 0.75)
    #            plt.xlabel('GPU Power (W)')
    #            plt.ylabel('Number of Times')
                plt.title('Power Consumed by GPU_{} During Inferencing'.format(n))
                ax.set_axisbelow(True)
                
                ax.minorticks_on()
                ax.grid(which = 'major', linestyle = '-',
                         linewidth = 0.5, color = 'black')
                ax.grid(which = 'minor', linestyle = ':',
                         linewidth = 0.5, color = 'black')
                print("Mean GPU_{} Consumption: {:.3f} W".format(n,np.mean(pwr)))
                print("Standard GPU_{} Consumption Deviation: {:.3f} W".format(n,np.std(pwr)))
            #Worker Processing time plot
            fig, ax = plt.subplots()
            ax.set_axisbelow(True)
            
            ax.minorticks_on()
            ax.grid(which = 'major', linestyle = '-',
                     linewidth = 0.5, color = 'black')
            ax.grid(which = 'minor', linestyle = ':',
                     linewidth = 0.5, color = 'black')
#            plt.stem(np.arange(0, len(tm[:,0])), tm[:,0], 
#                     use_line_collection = True, 
#                     label = 'Worker Processing Time')
#            plt.stem(np.arange(0, len(tm[:,0])), tm[:,1], 
#                     use_line_collection = True, 
#                     label = 'Visualization Processing Time',
#                     linefmt = 'C1-', markerfmt = 'C1o')
#            plt.stem(np.arange(0, len(tm[:,0])), tm[:,2], 
#                     use_line_collection = True, 
#                     label = 'Total Processing Time',
#                     linefmt = 'C2-', markerfmt = 'C2o')
            
            plt.ylabel('Frequency')
            plt.xlabel('Worker Time (ms)')
            plt.hist(tm[:,0], bins = 'auto', label = 'Worker Processing Time', 
                     edgecolor = 'black', linewidth = 0.75)
            plt.legend(loc = 'best')
            plt.title('Time taken to inference image')
            plt.grid(True, which = 'both', axis = 'both', zorder = 0)
            print("Mean Worker Time: {:.3f} ms".format(np.mean(tm[1:,0])))
            print("Standard Worker Deviation: {:.3f} ms".format(np.std(tm[1:,0])))
            
            plt.show()

            from numba import cuda
            cuda.select_device(0)
            cuda.close()
            """ """

        return
Esempio n. 3
0
    def start(self, cfg):
        """ """ """ """ """ """ """ """ """ """ """
        GET CONFIG
        """ """ """ """ """ """ """ """ """ """ """
        FORCE_GPU_COMPATIBLE = cfg['force_gpu_compatible']
        SAVE_TO_MOVIE        = cfg['save_to_movie']
        VISUALIZE            = cfg['visualize']
        VIS_WORKER           = cfg['vis_worker']
        VIS_TEXT             = cfg['vis_text']
        MAX_FRAMES           = cfg['max_frames']
        WIDTH                = cfg['width']
        HEIGHT               = cfg['height']
        FPS_INTERVAL         = cfg['fps_interval']
        DET_INTERVAL         = cfg['det_interval']
        DET_TH               = cfg['det_th']
        SPLIT_MODEL          = cfg['split_model']
        LOG_DEVICE           = cfg['log_device']
        ALLOW_MEMORY_GROWTH  = cfg['allow_memory_growth']
        SPLIT_SHAPE          = cfg['split_shape']
        DEBUG_MODE           = cfg['debug_mode']
        LABEL_PATH           = cfg['label_path']
        NUM_CLASSES          = cfg['num_classes']
        FROM_CAMERA          = cfg['from_camera']
        if FROM_CAMERA:
            VIDEO_INPUT = cfg['camera_input']
        else:
            VIDEO_INPUT = cfg['movie_input']
        """ """

        """ """ """ """ """ """ """ """ """ """ """
        LOAD FROZEN_GRAPH
        """ """ """ """ """ """ """ """ """ """ """
        load_frozen_graph = LoadFrozenGraph(cfg)
        graph = load_frozen_graph.load_graph()
        """ """

        """ """ """ """ """ """ """ """ """ """ """
        LOAD LABEL MAP
        """ """ """ """ """ """ """ """ """ """ """
        llm = LoadLabelMap()
        category_index = llm.load_label_map(cfg)
        """ """

        """ """ """ """ """ """ """ """ """ """ """
        PREPARE TF CONFIG OPTION
        """ """ """ """ """ """ """ """ """ """ """
        # Session Config: allow seperate GPU/CPU adressing and limit memory allocation
        config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=LOG_DEVICE)
        config.gpu_options.allow_growth = ALLOW_MEMORY_GROWTH
        config.gpu_options.force_gpu_compatible = FORCE_GPU_COMPATIBLE
        #config.gpu_options.per_process_gpu_memory_fraction = 0.01 # 80MB memory is enough to run on TX2
        """ """

        """ """ """ """ """ """ """ """ """ """ """
        PREPARE GRAPH I/O TO VARIABLE
        """ """ """ """ """ """ """ """ """ """ """
        # Define Input and Ouput tensors
        image_tensor = graph.get_tensor_by_name('image_tensor:0')
        detection_boxes = graph.get_tensor_by_name('detection_boxes:0')
        detection_scores = graph.get_tensor_by_name('detection_scores:0')
        detection_classes = graph.get_tensor_by_name('detection_classes:0')
        num_detections = graph.get_tensor_by_name('num_detections:0')
        detection_masks = graph.get_tensor_by_name('detection_masks:0')


        """ """ """ """ """ """ """ """ """ """ """
        START CAMERA
        """ """ """ """ """ """ """ """ """ """ """
        if FROM_CAMERA:
            from lib.webcam import WebcamVideoStream as VideoReader
        else:
            from lib.video import VideoReader
        video_reader = VideoReader()
        video_reader.start(VIDEO_INPUT, WIDTH, HEIGHT, save_to_movie=SAVE_TO_MOVIE)
        frame_cols, frame_rows = video_reader.getSize()
        """ """


        """ """ """ """ """ """ """ """ """ """ """
        FONT
        """ """ """ """ """ """ """ """ """ """ """
        """ STATISTICS FONT """
        fontFace = cv2.FONT_HERSHEY_SIMPLEX
        fontScale = frame_rows/1000.0
        if fontScale < 0.4:
            fontScale = 0.4
        fontThickness = 1 + int(fontScale)


        """ """ """ """ """ """ """ """ """ """ """
        PREAPRE GRAPH MASK OUTPUT
        """ """ """ """ """ """ """ """ """ """ """
        # The following processing is only for single image
        _detection_boxes = tf.squeeze(detection_boxes, [0])
        _detection_masks = tf.squeeze(detection_masks, [0])
        # Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.
        _real_num_detection = tf.cast(num_detections[0], tf.int32)
        _detection_boxes = tf.slice(_detection_boxes, [0, 0], [_real_num_detection, -1])
        _detection_masks = tf.slice(_detection_masks, [0, 0, 0], [_real_num_detection, -1, -1])
        _detection_masks_reframed = utils_ops.reframe_box_masks_to_image_masks(
            _detection_masks, _detection_boxes, frame_rows, frame_cols)
        _detection_masks_reframed = tf.cast(
            tf.greater(_detection_masks_reframed, 0.5), tf.uint8)
        # Follow the convention by adding back the batch dimension
        detection_masks = tf.expand_dims(
            _detection_masks_reframed, 0)


        if SPLIT_MODEL:
            SPLIT_TARGET_NAME = ['Gather',
                                 'BatchMultiClassNonMaxSuppression_1/map/TensorArrayStack_1/TensorArrayGatherV3',
                                 'BatchMultiClassNonMaxSuppression_1/map/TensorArrayStack_4/TensorArrayGatherV3',
                                 'Shape_11',
                                 'Shape_12',
                                 'add_1',
                                 'detection_boxes',
            ]
            split_out = []
            split_in = []
            for stn in SPLIT_TARGET_NAME:
                split_out += [graph.get_tensor_by_name(stn+':0')]
                split_in += [graph.get_tensor_by_name(stn+'_1:0')]
        """ """

        """ """ """ """ """ """ """ """ """ """ """
        START WORKER THREAD
        """ """ """ """ """ """ """ """ """ """ """
        # gpu_worker uses in split_model and non-split_model
        gpu_tag = 'GPU'
        cpu_tag = 'CPU'
        gpu_worker = SessionWorker(gpu_tag, graph, config)
        if SPLIT_MODEL:
            gpu_opts = split_out
            cpu_worker = SessionWorker(cpu_tag, graph, config)
            cpu_opts = [detection_boxes, detection_scores, detection_classes, num_detections, detection_masks]
        else:
            gpu_opts = [detection_boxes, detection_scores, detection_classes, num_detections, detection_masks]
        """ """

        """
        START VISUALIZE WORKER
        """
        if VISUALIZE and VIS_WORKER:
            q_out = Queue.Queue()
            vis_worker = MPVisualizeWorker(cfg, MPVariable.vis_in_con)
            """ """ """ """ """ """ """ """ """ """ """
            START SENDER THREAD
            """ """ """ """ """ """ """ """ """ """ """
            start_sender(MPVariable.det_out_con, q_out)
        proc_frame_counter = 0
        vis_proc_time = 0

        """ """ """ """ """ """ """ """ """ """ """
        LOAD LABEL MAP
        """ """ """ """ """ """ """ """ """ """ """
        llm = LoadLabelMap()
        category_index = llm.load_label_map(cfg)
        """ """


        """ """ """ """ """ """ """ """ """ """ """
        WAIT UNTIL THE FIRST DUMMY IMAGE DONE
        """ """ """ """ """ """ """ """ """ """ """
        print('Loading...')
        sleep_interval = 0.1
        '''
        """
        PUT DUMMY DATA INTO GPU WORKER
        """
        gpu_feeds = {image_tensor:  [np.zeros((300, 300, 3))]}
        gpu_extras = {}
        gpu_worker.put_sess_queue(gpu_opts, gpu_feeds, gpu_extras)
        if SPLIT_MODEL:
            """
            PUT DUMMY DATA INTO CPU WORKER
            """
            slice1 = np.zeros((1, SPLIT_SHAPE, NUM_CLASSES))
            expand = np.zeros((1, SPLIT_SHAPE, 1, 4))
            stack = [[0., 0., 1., 1.]]

            cpu_feeds = {slice1_in: slice1, expand_in: expand, stack_in: stack}
            cpu_extras = {}
            cpu_worker.put_sess_queue(cpu_opts, cpu_feeds, cpu_extras)
        """
        WAIT UNTIL JIT-COMPILE DONE
        """
        while True:
            g = gpu_worker.get_result_queue()
            if g is None:
                time.sleep(sleep_interval)
            else:
                break
        if SPLIT_MODEL:
            while True:
                c = cpu_worker.get_result_queue()
                if c is None:
                    time.sleep(sleep_interval)
                else:
                    break
        """ """
        '''

        """ """ """ """ """ """ """ """ """ """ """
        DETECTION LOOP
        """ """ """ """ """ """ """ """ """ """ """
        print('Starting Detection')
        sleep_interval = 0.005
        top_in_time = None
        try:
            while video_reader.running and MPVariable.running.value:
                if top_in_time is None:
                    top_in_time = time.time()
                """
                SPRIT/NON-SPLIT MODEL CAMERA TO WORKER
                """
                if gpu_worker.is_sess_empty(): # must need for speed
                    cap_in_time = time.time()
                    frame = video_reader.read()
                    if frame is None:
                        MPVariable.running.value = False
                        break
                    image_expanded = np.expand_dims(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB), axis=0) # np.expand_dims is faster than []
                    #image_expanded = np.expand_dims(frame, axis=0) # BGR image for input. Of couse, bad accuracy in RGB trained model, but speed up.
                    cap_out_time = time.time()
                    # put new queue
                    gpu_feeds = {image_tensor: image_expanded}
                    gpu_extras = {'image':frame, 'top_in_time':top_in_time, 'cap_in_time':cap_in_time, 'cap_out_time':cap_out_time} # always image draw.
                    gpu_worker.put_sess_queue(gpu_opts, gpu_feeds, gpu_extras)

                g = gpu_worker.get_result_queue()
                if SPLIT_MODEL:
                    # if g is None: gpu thread has no output queue. ok skip, let's check cpu thread.
                    if g is not None:
                        # gpu thread has output queue.
                        result_slice_out, extras = g['results'], g['extras']
                        if cpu_worker.is_sess_empty():
                            # When cpu thread has no next queue, put new queue.
                            # else, drop gpu queue.
                            cpu_feeds = {}
                            for i in range(len(result_slice_out)):
                                cpu_feeds.update({split_in[i]:result_slice_out[i]})
                            # add image_tensor
                            cpu_feeds.update({image_tensor: image_expanded})
                            cpu_extras = extras
                            cpu_worker.put_sess_queue(cpu_opts, cpu_feeds, cpu_extras)
                        # else: cpu thread is busy. don't put new queue. let's check cpu result queue.
                    # check cpu thread.
                    q = cpu_worker.get_result_queue()
                else:
                    """
                    NON-SPLIT MODEL
                    """
                    q = g
                if q is None:
                    """
                    SPLIT/NON-SPLIT MODEL
                    """
                    # detection is not complete yet. ok nothing to do.
                    time.sleep(sleep_interval)
                    continue

                boxes, scores, classes, num, masks, extras = q['results'][0], q['results'][1], q['results'][2], q['results'][3], q['results'][4], q['extras']
                det_out_time = time.time()

                """
                ALWAYS BOX DRAW ON IMAGE
                """
                vis_in_time = time.time()
                image = extras['image']
                image = visualization(category_index, image, boxes, scores, classes, DEBUG_MODE, VIS_TEXT, FPS_INTERVAL,
                                      fontFace=fontFace, fontScale=fontScale, fontThickness=fontThickness, masks=masks)

                """
                VISUALIZATION
                """
                if VISUALIZE:
                    if (MPVariable.vis_skip_rate.value == 0) or (proc_frame_counter % MPVariable.vis_skip_rate.value < 1):
                        if VIS_WORKER:
                            q_out.put({'image':image, 'vis_in_time':vis_in_time})
                        else:
                            #np.set_printoptions(precision=5, suppress=True, threshold=np.inf)  # suppress scientific float notation
                            """
                            SHOW
                            """
                            cv2.imshow("Object Detection", image)
                            # Press q to quit
                            if cv2.waitKey(1) & 0xFF == 113: #ord('q'):
                                break
                            MPVariable.vis_frame_counter.value += 1
                            vis_out_time = time.time()
                            """
                            PROCESSING TIME
                            """
                            vis_proc_time = vis_out_time - vis_in_time
                            MPVariable.vis_proc_time.value += vis_proc_time
                else:
                    """
                    NO VISUALIZE
                    """
                    for box, score, _class in zip(np.squeeze(boxes), np.squeeze(scores), np.squeeze(classes)):
                        if proc_frame_counter % DET_INTERVAL == 0 and score > DET_TH:
                            label = category_index[_class]['name']
                            print("label: {}\nscore: {}\nbox: {}".format(label, score, box))

                    MPVariable.vis_frame_counter.value += 1
                    vis_out_time = time.time()
                    """
                    PROCESSING TIME
                    """
                    vis_proc_time = vis_out_time - vis_in_time
                            
                if SAVE_TO_MOVIE:
                    video_reader.save(image)

                proc_frame_counter += 1
                if proc_frame_counter > 100000:
                    proc_frame_counter = 0
                """
                PROCESSING TIME
                """
                top_in_time = extras['top_in_time']
                cap_proc_time = extras['cap_out_time'] - extras['cap_in_time']
                gpu_proc_time = extras[gpu_tag+'_out_time'] - extras[gpu_tag+'_in_time']
                if SPLIT_MODEL:
                    cpu_proc_time = extras[cpu_tag+'_out_time'] - extras[cpu_tag+'_in_time']
                else:
                    cpu_proc_time = 0
                lost_proc_time = det_out_time - top_in_time - cap_proc_time - gpu_proc_time - cpu_proc_time
                total_proc_time = det_out_time - top_in_time
                MPVariable.cap_proc_time.value += cap_proc_time
                MPVariable.gpu_proc_time.value += gpu_proc_time
                MPVariable.cpu_proc_time.value += cpu_proc_time
                MPVariable.lost_proc_time.value += lost_proc_time
                MPVariable.total_proc_time.value += total_proc_time

                if DEBUG_MODE:
                    if SPLIT_MODEL:
                        sys.stdout.write('snapshot FPS:{: ^5.1f} total:{: ^10.5f} cap:{: ^10.5f} gpu:{: ^10.5f} cpu:{: ^10.5f} lost:{: ^10.5f} | vis:{: ^10.5f}\n'.format(
                            MPVariable.fps.value, total_proc_time, cap_proc_time, gpu_proc_time, cpu_proc_time, lost_proc_time, vis_proc_time))
                    else:
                        sys.stdout.write('snapshot FPS:{: ^5.1f} total:{: ^10.5f} cap:{: ^10.5f} gpu:{: ^10.5f} lost:{: ^10.5f} | vis:{: ^10.5f}\n'.format(
                            MPVariable.fps.value, total_proc_time, cap_proc_time, gpu_proc_time, lost_proc_time, vis_proc_time))
                """
                EXIT WITHOUT GUI
                """
                if not VISUALIZE and MAX_FRAMES > 0:
                    if proc_frame_counter >= MAX_FRAMES:
                        MPVariable.running.value = False
                        break

                """
                CHANGE SLEEP INTERVAL
                """
                if MPVariable.frame_counter.value == 0 and MPVariable.fps.value > 0:
                    sleep_interval = 0.1 / MPVariable.fps.value
                    MPVariable.sleep_interval.value = sleep_interval
                MPVariable.frame_counter.value += 1
                top_in_time = None
            """
            END while
            """
        except:
            import traceback
            traceback.print_exc()
        finally:
            """ """ """ """ """ """ """ """ """ """ """
            CLOSE
            """ """ """ """ """ """ """ """ """ """ """
            if VISUALIZE and VIS_WORKER:
                q_out.put(None)
            MPVariable.running.value = False
            gpu_worker.stop()
            if SPLIT_MODEL:
                cpu_worker.stop()
            video_reader.stop()

            if VISUALIZE:
                cv2.destroyAllWindows()
            """ """

        return
    def start(self, cfg):
        """ """ """ """ """ """ """ """ """ """ """
        GET CONFIG
        """ """ """ """ """ """ """ """ """ """ """
        FORCE_GPU_COMPATIBLE = cfg['force_gpu_compatible']
        SAVE_TO_FILE = cfg['save_to_file']
        VISUALIZE = cfg['visualize']
        VIS_WORKER = cfg['vis_worker']
        VIS_TEXT = cfg['vis_text']
        MAX_FRAMES = cfg['max_frames']
        WIDTH = cfg['width']
        HEIGHT = cfg['height']
        FPS_INTERVAL = cfg['fps_interval']
        DET_INTERVAL = cfg['det_interval']
        DET_TH = cfg['det_th']
        LOG_DEVICE = cfg['log_device']
        ALLOW_MEMORY_GROWTH = cfg['allow_memory_growth']
        DEBUG_MODE = cfg['debug_mode']
        LABEL_PATH = cfg['label_path']
        NUM_CLASSES = cfg['num_classes']
        MIN_AREA = 500
        SRC_FROM = cfg['src_from']
        CAMERA = 0
        MOVIE = 1
        IMAGE = 2
        if SRC_FROM == 'camera':
            SRC_FROM = CAMERA
            VIDEO_INPUT = cfg['camera_input']
        elif SRC_FROM == 'movie':
            SRC_FROM = MOVIE
            VIDEO_INPUT = cfg['movie_input']
        elif SRC_FROM == 'image':
            SRC_FROM = IMAGE
            VIDEO_INPUT = cfg['image_input']
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        LOAD FROZEN_GRAPH
        """ """ """ """ """ """ """ """ """ """ """
        load_frozen_graph = LoadFrozenGraph(cfg)
        graph = load_frozen_graph.load_graph()
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        LOAD LABEL MAP
        """ """ """ """ """ """ """ """ """ """ """
        llm = LoadLabelMap()
        category_index = llm.load_label_map(cfg)
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        PREPARE TF CONFIG OPTION
        """ """ """ """ """ """ """ """ """ """ """
        # Session Config: allow seperate GPU/CPU adressing and limit memory allocation
        config = tf.ConfigProto(allow_soft_placement=True,
                                log_device_placement=LOG_DEVICE)
        config.gpu_options.allow_growth = ALLOW_MEMORY_GROWTH
        config.gpu_options.force_gpu_compatible = FORCE_GPU_COMPATIBLE
        #config.gpu_options.per_process_gpu_memory_fraction = 0.01 # 80MB memory is enough to run on TX2
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        PREPARE GRAPH I/O TO VARIABLE
        """ """ """ """ """ """ """ """ """ """ """
        # Define Input and Ouput tensors
        image_tensor = graph.get_tensor_by_name('ImageTensor:0')
        semantic_predictions = graph.get_tensor_by_name(
            'SemanticPredictions:0')
        """ """ """ """ """ """ """ """ """ """ """
        START CAMERA
        """ """ """ """ """ """ """ """ """ """ """
        if SRC_FROM == CAMERA:
            from lib.webcam import WebcamVideoStream as VideoReader
        elif SRC_FROM == MOVIE:
            from lib.video import VideoReader
        elif SRC_FROM == IMAGE:
            from lib.image import ImageReader as VideoReader
        video_reader = VideoReader()

        if SRC_FROM == IMAGE:
            video_reader.start(VIDEO_INPUT, save_to_file=SAVE_TO_FILE)
            frame_cols, frame_rows = HEIGHT, WIDTH
        else:  # CAMERA, MOVIE
            video_reader.start(VIDEO_INPUT,
                               WIDTH,
                               HEIGHT,
                               save_to_file=SAVE_TO_FILE)
            frame_cols, frame_rows = video_reader.getSize()
            """ STATISTICS FONT """
            fontScale = frame_rows / 1000.0
            if fontScale < 0.4:
                fontScale = 0.4
            fontThickness = 1 + int(fontScale)
        fontFace = cv2.FONT_HERSHEY_SIMPLEX
        if SRC_FROM == MOVIE:
            dir_path, filename = os.path.split(VIDEO_INPUT)
            filepath_prefix = filename
        elif SRC_FROM == CAMERA:
            filepath_prefix = 'frame'
        """ """

        LABEL_NAMES = np.asarray([
            'background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle',
            'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse',
            'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train',
            'tv'
        ])
        FULL_LABEL_MAP = np.arange(len(LABEL_NAMES)).reshape(
            len(LABEL_NAMES), 1)
        FULL_COLOR_MAP = label_to_color_image(FULL_LABEL_MAP)
        """ """
        """ """ """ """ """ """ """ """ """ """ """
        START WORKER THREAD
        """ """ """ """ """ """ """ """ """ """ """
        # gpu_worker uses in split_model and non-split_model
        gpu_tag = 'GPU'
        cpu_tag = 'CPU'
        gpu_worker = SessionWorker(gpu_tag, graph, config)
        gpu_opts = [semantic_predictions]
        """ """
        """
        START VISUALIZE WORKER
        """
        if VISUALIZE and VIS_WORKER:
            q_out = Queue.Queue()
            vis_worker = MPVisualizeWorker(cfg, MPVariable.vis_in_con)
            """ """ """ """ """ """ """ """ """ """ """
            START SENDER THREAD
            """ """ """ """ """ """ """ """ """ """ """
            start_sender(MPVariable.det_out_con, q_out)
        proc_frame_counter = 0
        vis_proc_time = 0
        """ """ """ """ """ """ """ """ """ """ """
        WAIT UNTIL THE FIRST DUMMY IMAGE DONE
        """ """ """ """ """ """ """ """ """ """ """
        print('Loading...')
        sleep_interval = 0.1
        """ """ """ """ """ """ """ """ """ """ """
        DETECTION LOOP
        """ """ """ """ """ """ """ """ """ """ """
        print('Starting Detection')
        sleep_interval = 0.005
        top_in_time = None
        frame_in_processing_counter = 0
        resize_ratio = 1.0 * 513 / max(frame_cols, frame_rows)
        target_size = (int(resize_ratio * frame_cols),
                       int(resize_ratio * frame_rows))
        try:
            if not video_reader.running:
                raise IOError(("Input src error."))
            while MPVariable.running.value:
                if top_in_time is None:
                    top_in_time = time.time()
                """
                NON-SPLIT MODEL CAMERA TO WORKER
                """
                if video_reader.running:
                    if gpu_worker.is_sess_empty():  # must need for speed
                        cap_in_time = time.time()
                        if SRC_FROM == IMAGE:
                            frame, filepath = video_reader.read()
                            if frame is not None:
                                frame_in_processing_counter += 1
                        else:
                            frame = video_reader.read()
                            if frame is not None:
                                filepath = filepath_prefix + '_' + str(
                                    proc_frame_counter) + '.png'
                                frame_in_processing_counter += 1
                        if frame is not None:
                            frame = cv2.resize(frame, target_size)
                            image_expanded = np.expand_dims(
                                cv2.cvtColor(frame, cv2.COLOR_BGR2RGB),
                                axis=0)  # np.expand_dims is faster than []
                            #image_expanded = np.expand_dims(frame, axis=0) # BGR image for input. Of couse, bad accuracy in RGB trained model, but speed up.
                            cap_out_time = time.time()
                            # put new queue
                            gpu_feeds = {image_tensor: image_expanded}
                            gpu_extras = {
                                'image': frame,
                                'top_in_time': top_in_time,
                                'cap_in_time': cap_in_time,
                                'cap_out_time': cap_out_time,
                                'filepath': filepath
                            }  # always image draw.
                            gpu_worker.put_sess_queue(gpu_opts, gpu_feeds,
                                                      gpu_extras)
                elif frame_in_processing_counter <= 0:
                    MPVariable.running.value = False
                    break

                g = gpu_worker.get_result_queue()
                if g is None:
                    # detection is not complete yet. ok nothing to do.
                    time.sleep(sleep_interval)
                    continue

                frame_in_processing_counter -= 1
                batch_seg_map, extras = g['results'][0], g['extras']
                seg_map = batch_seg_map[0]

                det_out_time = time.time()
                """
                ALWAYS BOX DRAW ON IMAGE
                """
                vis_in_time = time.time()
                image = extras['image']
                if SRC_FROM == IMAGE:
                    filepath = extras['filepath']
                    frame_rows, frame_cols = image.shape[:2]
                    """ STATISTICS FONT """
                    fontScale = frame_rows / 1000.0
                    if fontScale < 0.4:
                        fontScale = 0.4
                    fontThickness = 1 + int(fontScale)
                else:
                    filepath = extras['filepath']

                seg_image = STANDARD_COLORS_ARRAY[seg_map]
                #seg_image = label_to_color_image(seg_map).astype(np.uint8)
                #unique_labels = np.unique(seg_map)
                #rgb_seg = full_color_map[unique_labels].astype(np.uint8)
                ### TODO: to bgr
                #image = to_layer(image, seg_image, background_alpha=1.0, foreground_alpha=1.0, gamma=0)
                b_channel, g_channel, r_channel = cv2.split(seg_image)
                # Make a single channel mask if background: 0 else: 1
                mask = seg_map > 0
                alpha_channel = np.ones(
                    b_channel.shape, dtype=b_channel.dtype
                ) * 0  #creating a dummy alpha channel image.
                seg_image_bgra = cv2.merge(
                    (b_channel, g_channel, r_channel, alpha_channel))
                seg_image = cv2.merge(cv2.split(seg_image_bgra)[:3])
                #image = overdraw(image, seg_image, mask)
                image = blending(image, seg_image)
                image = deeplab_visualization(LABEL_NAMES,
                                              image,
                                              seg_map,
                                              DEBUG_MODE,
                                              VIS_TEXT,
                                              FPS_INTERVAL,
                                              fontFace=fontFace,
                                              fontScale=fontScale,
                                              fontThickness=fontThickness)
                """
                VISUALIZATION
                """
                if VISUALIZE:
                    if (MPVariable.vis_skip_rate.value
                            == 0) or (proc_frame_counter %
                                      MPVariable.vis_skip_rate.value < 1):
                        if VIS_WORKER:
                            q_out.put({
                                'image': image,
                                'vis_in_time': vis_in_time
                            })
                        else:
                            """
                            SHOW
                            """
                            cv2.imshow("Object Detection", image)
                            # Press q to quit
                            if cv2.waitKey(1) & 0xFF == 113:  #ord('q'):
                                break
                            MPVariable.vis_frame_counter.value += 1
                            vis_out_time = time.time()
                            """
                            PROCESSING TIME
                            """
                            vis_proc_time = vis_out_time - vis_in_time
                            MPVariable.vis_proc_time.value += vis_proc_time
                else:
                    """
                    NO VISUALIZE
                    """
                    for box, score, _class in zip(boxes, scores, classes):
                        if proc_frame_counter % DET_INTERVAL == 0 and score > DET_TH:
                            label = category_index[_class]['name']
                            print("label: {}\nscore: {}\nbox: {}".format(
                                label, score, box))

                    vis_out_time = time.time()
                    """
                    PROCESSING TIME
                    """
                    vis_proc_time = vis_out_time - vis_in_time

                if SAVE_TO_FILE:
                    if SRC_FROM == IMAGE:
                        video_reader.save(image, filepath)
                    else:
                        video_reader.save(image)

                proc_frame_counter += 1
                if proc_frame_counter > 100000:
                    proc_frame_counter = 0
                """
                PROCESSING TIME
                """
                top_in_time = extras['top_in_time']
                cap_proc_time = extras['cap_out_time'] - extras['cap_in_time']
                gpu_proc_time = extras[gpu_tag +
                                       '_out_time'] - extras[gpu_tag +
                                                             '_in_time']
                cpu_proc_time = 0
                lost_proc_time = det_out_time - top_in_time - cap_proc_time - gpu_proc_time - cpu_proc_time
                total_proc_time = det_out_time - top_in_time
                MPVariable.cap_proc_time.value += cap_proc_time
                MPVariable.gpu_proc_time.value += gpu_proc_time
                MPVariable.cpu_proc_time.value += cpu_proc_time
                MPVariable.lost_proc_time.value += lost_proc_time
                MPVariable.total_proc_time.value += total_proc_time

                if DEBUG_MODE:
                    sys.stdout.write(
                        'snapshot FPS:{: ^5.1f} total:{: ^10.5f} cap:{: ^10.5f} gpu:{: ^10.5f} lost:{: ^10.5f} | vis:{: ^10.5f}\n'
                        .format(MPVariable.fps.value, total_proc_time,
                                cap_proc_time, gpu_proc_time, lost_proc_time,
                                vis_proc_time))
                """
                EXIT WITHOUT GUI
                """
                if not VISUALIZE and MAX_FRAMES > 0:
                    if proc_frame_counter >= MAX_FRAMES:
                        MPVariable.running.value = False
                        break
                """
                CHANGE SLEEP INTERVAL
                """
                if MPVariable.frame_counter.value == 0 and MPVariable.fps.value > 0:
                    sleep_interval = 0.1 / MPVariable.fps.value
                    MPVariable.sleep_interval.value = sleep_interval
                MPVariable.frame_counter.value += 1
                top_in_time = None
            """
            END while
            """
        except:
            import traceback
            traceback.print_exc()
        finally:
            """ """ """ """ """ """ """ """ """ """ """
            CLOSE
            """ """ """ """ """ """ """ """ """ """ """
            if VISUALIZE and VIS_WORKER:
                q_out.put(None)
            MPVariable.running.value = False
            gpu_worker.stop()
            video_reader.stop()

            if VISUALIZE:
                cv2.destroyAllWindows()
            """ """

        return