def __init__(
self,
hparams: OTEClassificationParameters,
label_schema: LabelSchemaEntity,
model_file: Union[str, bytes],
weight_file: Union[str, bytes, None] = None,
device: str = "CPU",
num_requests: int = 1,
):
"""
Inferencer implementation for OTEDetection using OpenVINO backend.
:param model: Path to model to load, `.xml`, `.bin` or `.onnx` file.
:param hparams: Hyper parameters that the model should use.
:param num_requests: Maximum number of requests that the inferencer can make.
Good value is the number of available cores. Defaults to 1.
:param device: Device to run inference on, such as CPU, GPU or MYRIAD. Defaults to "CPU".
"""
multilabel = len(label_schema.get_groups(False)) > 1 and \
len(label_schema.get_groups(False)) == len(label_schema.get_labels(include_empty=False))
self.label_schema = label_schema
model_adapter = OpenvinoAdapter(create_core(),
model_file,
weight_file,
device=device,
max_num_requests=num_requests)
self.configuration = {'multilabel': multilabel}
self.model = Model.create_model("ote_classification",
model_adapter,
self.configuration,
preload=True)
self.converter = ClassificationToAnnotationConverter(self.label_schema)
def create_model(model_path: Path,
config_file: Optional[Path] = None) -> Model:
"""
Create model using ModelAPI factory
"""
model_adapter = OpenvinoAdapter(create_core(), get_model_path(model_path))
parameters = get_parameters(config_file)
try:
importlib.import_module(".model", "demo_package")
except ImportError:
print("Using model wrapper from Open Model Zoo ModelAPI")
# labels for modelAPI wrappers can be empty, because unused in pre- and postprocessing
parameters["model_parameters"]["labels"] = []
model = Model.create_model(
parameters["type_of_model"],
model_adapter,
parameters["model_parameters"],
preload=True,
)
return model
def main():
args = build_argparser().parse_args()
cap = open_images_capture(args.input, args.loop)
next_frame_id = 1
next_frame_id_to_show = 0
metrics = PerformanceMetrics()
render_metrics = PerformanceMetrics()
video_writer = cv2.VideoWriter()
if args.adapter == 'openvino':
plugin_config = get_user_config(args.device, args.num_streams,
args.num_threads)
model_adapter = OpenvinoAdapter(
create_core(),
args.model,
device=args.device,
plugin_config=plugin_config,
max_num_requests=args.num_infer_requests)
elif args.adapter == 'ovms':
model_adapter = OVMSAdapter(args.model)
start_time = perf_counter()
frame = cap.read()
if frame is None:
raise RuntimeError("Can't read an image from the input")
model = Deblurring(model_adapter, preload=False)
model.reshape(frame.shape)
model.log_layers_info()
pipeline = AsyncPipeline(model)
pipeline.submit_data(frame, 0, {'frame': frame, 'start_time': start_time})
presenter = monitors.Presenter(
args.utilization_monitors, 55,
(round(frame.shape[1] / 4), round(frame.shape[0] / 8)))
if args.output and not video_writer.open(
args.output, cv2.VideoWriter_fourcc(*'MJPG'), cap.fps(),
(2 * frame.shape[1], frame.shape[0])):
raise RuntimeError("Can't open video writer")
while True:
if pipeline.is_ready():
# Get new image/frame
start_time = perf_counter()
frame = cap.read()
if frame is None:
break
# Submit for inference
pipeline.submit_data(frame, next_frame_id, {
'frame': frame,
'start_time': start_time
})
next_frame_id += 1
else:
# Wait for empty request
pipeline.await_any()
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
# Process all completed requests
results = pipeline.get_result(next_frame_id_to_show)
if results:
result_frame, frame_meta = results
input_frame = frame_meta['frame']
start_time = frame_meta['start_time']
rendering_start_time = perf_counter()
if input_frame.shape != result_frame.shape:
input_frame = cv2.resize(
input_frame,
(result_frame.shape[1], result_frame.shape[0]))
final_image = cv2.hconcat([input_frame, result_frame])
render_metrics.update(rendering_start_time)
presenter.drawGraphs(final_image)
metrics.update(start_time, final_image)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(final_image)
next_frame_id_to_show += 1
if not args.no_show:
cv2.imshow('Deblurring Results', final_image)
key = cv2.waitKey(1)
if key == 27 or key == 'q' or key == 'Q':
break
presenter.handleKey(key)
pipeline.await_all()
# Process completed requests
for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id):
results = pipeline.get_result(next_frame_id_to_show)
while results is None:
results = pipeline.get_result(next_frame_id_to_show)
result_frame, frame_meta = results
input_frame = frame_meta['frame']
start_time = frame_meta['start_time']
rendering_start_time = perf_counter()
if input_frame.shape != result_frame.shape:
input_frame = cv2.resize(
input_frame, (result_frame.shape[1], result_frame.shape[0]))
final_image = cv2.hconcat([input_frame, result_frame])
render_metrics.update(rendering_start_time)
presenter.drawGraphs(final_image)
metrics.update(start_time, final_image)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(final_image)
if not args.no_show:
cv2.imshow('Deblurring Results', final_image)
key = cv2.waitKey(1)
metrics.log_total()
log_latency_per_stage(cap.reader_metrics.get_latency(),
pipeline.preprocess_metrics.get_latency(),
pipeline.inference_metrics.get_latency(),
pipeline.postprocess_metrics.get_latency(),
render_metrics.get_latency())
for rep in presenter.reportMeans():
log.info(rep)
def main():
args = build_argparser().parse_args()
paragraphs = get_paragraphs(args.input)
preprocessing_start_time = perf_counter()
vocab = load_vocab_file(args.vocab)
log.debug("Loaded vocab file from {}, get {} tokens".format(args.vocab, len(vocab)))
# get context as a string (as we might need it's length for the sequence reshape)
context = '\n'.join(paragraphs)
sentences = re.split(sentence_splitter, context)
preprocessed_sentences = [text_to_tokens(sentence, vocab) for sentence in sentences]
max_sentence_length = max([len(tokens) + 2 for tokens, _ in preprocessed_sentences])
preprocessing_total_time = (perf_counter() - preprocessing_start_time) * 1e3
source = tuple(zip(sentences, preprocessed_sentences))
if args.adapter == 'openvino':
plugin_config = get_user_config(args.device, args.num_streams, args.num_threads)
model_adapter = OpenvinoAdapter(create_core(), args.model, device=args.device, plugin_config=plugin_config,
max_num_requests=args.num_infer_requests, model_parameters = {'input_layouts': args.layout})
elif args.adapter == 'ovms':
model_adapter = OVMSAdapter(args.model)
enable_padding = not args.dynamic_shape
model = BertNamedEntityRecognition(model_adapter, {'vocab': vocab, 'input_names': args.input_names, 'enable_padding': enable_padding})
if max_sentence_length > model.max_length:
model.reshape(max_sentence_length if enable_padding else (1, max_sentence_length))
model.log_layers_info()
pipeline = AsyncPipeline(model)
next_sentence_id = 0
next_sentence_id_to_show = 0
start_time = perf_counter()
while True:
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
results = pipeline.get_result(next_sentence_id_to_show)
if results:
(score, filtered_labels_id), meta = results
next_sentence_id_to_show += 1
print_raw_results(score, filtered_labels_id, meta)
continue
if pipeline.is_ready():
if next_sentence_id == len(source):
break
sentence, (c_tokens_id, c_token_s_e) = source[next_sentence_id]
pipeline.submit_data(c_tokens_id, next_sentence_id, {'sentence': sentence, 'c_token_s_e': c_token_s_e})
next_sentence_id += 1
else:
pipeline.await_any()
pipeline.await_all()
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
for sentence_id in range(next_sentence_id_to_show, next_sentence_id):
results = pipeline.get_result(sentence_id)
(score, filtered_labels_id), meta = results
print_raw_results(score, filtered_labels_id, meta)
total_latency = (perf_counter() - start_time) * 1e3 + preprocessing_total_time
log.info("Metrics report:")
log.info("\tLatency: {:.1f} ms".format(total_latency))
def main():
args = build_argparser().parse_args()
cap = open_images_capture(args.input, args.loop)
next_frame_id = 1
next_frame_id_to_show = 0
metrics = PerformanceMetrics()
render_metrics = PerformanceMetrics()
video_writer = cv2.VideoWriter()
plugin_config = get_user_config(args.device, args.num_streams,
args.num_threads)
model_adapter = OpenvinoAdapter(
create_core(),
args.model,
device=args.device,
plugin_config=plugin_config,
max_num_requests=args.num_infer_requests,
model_parameters={'input_layouts': args.layout})
start_time = perf_counter()
frame = cap.read()
if frame is None:
raise RuntimeError("Can't read an image from the input")
config = {
'target_size': args.tsize,
'aspect_ratio': frame.shape[1] / frame.shape[0],
'confidence_threshold': args.prob_threshold,
'padding_mode': 'center' if args.architecture_type == 'higherhrnet'
else None, # the 'higherhrnet' and 'ae' specific
'delta': 0.5 if args.architecture_type == 'higherhrnet' else
None, # the 'higherhrnet' and 'ae' specific
}
model = ImageModel.create_model(ARCHITECTURES[args.architecture_type],
model_adapter, config)
model.log_layers_info()
hpe_pipeline = AsyncPipeline(model)
hpe_pipeline.submit_data(frame, 0, {
'frame': frame,
'start_time': start_time
})
output_transform = OutputTransform(frame.shape[:2], args.output_resolution)
if args.output_resolution:
output_resolution = output_transform.new_resolution
else:
output_resolution = (frame.shape[1], frame.shape[0])
presenter = monitors.Presenter(
args.utilization_monitors, 55,
(round(output_resolution[0] / 4), round(output_resolution[1] / 8)))
if args.output and not video_writer.open(args.output,
cv2.VideoWriter_fourcc(*'MJPG'),
cap.fps(), output_resolution):
raise RuntimeError("Can't open video writer")
while True:
if hpe_pipeline.callback_exceptions:
raise hpe_pipeline.callback_exceptions[0]
# Process all completed requests
results = hpe_pipeline.get_result(next_frame_id_to_show)
if results:
(poses, scores), frame_meta = results
frame = frame_meta['frame']
start_time = frame_meta['start_time']
if len(poses) and args.raw_output_message:
print_raw_results(poses, scores, next_frame_id_to_show)
presenter.drawGraphs(frame)
rendering_start_time = perf_counter()
frame = draw_poses(frame, poses, args.prob_threshold,
output_transform)
render_metrics.update(rendering_start_time)
metrics.update(start_time, frame)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
next_frame_id_to_show += 1
if not args.no_show:
cv2.imshow('Pose estimation results', frame)
key = cv2.waitKey(1)
ESC_KEY = 27
# Quit.
if key in {ord('q'), ord('Q'), ESC_KEY}:
break
presenter.handleKey(key)
continue
if hpe_pipeline.is_ready():
# Get new image/frame
start_time = perf_counter()
frame = cap.read()
if frame is None:
break
# Submit for inference
hpe_pipeline.submit_data(frame, next_frame_id, {
'frame': frame,
'start_time': start_time
})
next_frame_id += 1
else:
# Wait for empty request
hpe_pipeline.await_any()
hpe_pipeline.await_all()
if hpe_pipeline.callback_exceptions:
raise hpe_pipeline.callback_exceptions[0]
# Process completed requests
for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id):
results = hpe_pipeline.get_result(next_frame_id_to_show)
(poses, scores), frame_meta = results
frame = frame_meta['frame']
start_time = frame_meta['start_time']
if len(poses) and args.raw_output_message:
print_raw_results(poses, scores, next_frame_id_to_show)
presenter.drawGraphs(frame)
rendering_start_time = perf_counter()
frame = draw_poses(frame, poses, args.prob_threshold, output_transform)
render_metrics.update(rendering_start_time)
metrics.update(start_time, frame)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
if not args.no_show:
cv2.imshow('Pose estimation results', frame)
key = cv2.waitKey(1)
ESC_KEY = 27
# Quit.
if key in {ord('q'), ord('Q'), ESC_KEY}:
break
presenter.handleKey(key)
metrics.log_total()
log_latency_per_stage(cap.reader_metrics.get_latency(),
hpe_pipeline.preprocess_metrics.get_latency(),
hpe_pipeline.inference_metrics.get_latency(),
hpe_pipeline.postprocess_metrics.get_latency(),
render_metrics.get_latency())
for rep in presenter.reportMeans():
log.info(rep)
def main():
args = build_argparser().parse_args()
cap = open_images_capture(args.input, args.loop)
if args.adapter == 'openvino':
plugin_config = get_user_config(args.device, args.num_streams,
args.num_threads)
model_adapter = OpenvinoAdapter(
create_core(),
args.model,
device=args.device,
plugin_config=plugin_config,
max_num_requests=args.num_infer_requests,
model_parameters={'input_layouts': args.layout})
elif args.adapter == 'ovms':
model_adapter = OVMSAdapter(args.model)
model = SegmentationModel.create_model(args.architecture_type,
model_adapter,
{'path_to_labels': args.labels})
if args.architecture_type == 'segmentation':
visualizer = SegmentationVisualizer(args.colors)
if args.architecture_type == 'salient_object_detection':
visualizer = SaliencyMapVisualizer()
model.log_layers_info()
pipeline = AsyncPipeline(model)
next_frame_id = 0
next_frame_id_to_show = 0
metrics = PerformanceMetrics()
render_metrics = PerformanceMetrics()
presenter = None
output_transform = None
video_writer = cv2.VideoWriter()
only_masks = args.only_masks
while True:
if pipeline.is_ready():
# Get new image/frame
start_time = perf_counter()
frame = cap.read()
if frame is None:
if next_frame_id == 0:
raise ValueError("Can't read an image from the input")
break
if next_frame_id == 0:
output_transform = OutputTransform(frame.shape[:2],
args.output_resolution)
if args.output_resolution:
output_resolution = output_transform.new_resolution
else:
output_resolution = (frame.shape[1], frame.shape[0])
presenter = monitors.Presenter(
args.utilization_monitors, 55,
(round(output_resolution[0] / 4),
round(output_resolution[1] / 8)))
if args.output and not video_writer.open(
args.output, cv2.VideoWriter_fourcc(*'MJPG'),
cap.fps(), output_resolution):
raise RuntimeError("Can't open video writer")
# Submit for inference
pipeline.submit_data(frame, next_frame_id, {
'frame': frame,
'start_time': start_time
})
next_frame_id += 1
else:
# Wait for empty request
pipeline.await_any()
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
# Process all completed requests
results = pipeline.get_result(next_frame_id_to_show)
if results:
objects, frame_meta = results
if args.raw_output_message:
print_raw_results(objects, next_frame_id_to_show, model.labels)
frame = frame_meta['frame']
start_time = frame_meta['start_time']
rendering_start_time = perf_counter()
frame = render_segmentation(frame, objects, visualizer,
output_transform, only_masks)
render_metrics.update(rendering_start_time)
presenter.drawGraphs(frame)
metrics.update(start_time, frame)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
next_frame_id_to_show += 1
if not args.no_show:
cv2.imshow('Segmentation Results', frame)
key = cv2.waitKey(1)
if key == 27 or key == 'q' or key == 'Q':
break
if key == 9:
only_masks = not only_masks
presenter.handleKey(key)
pipeline.await_all()
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
# Process completed requests
for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id):
results = pipeline.get_result(next_frame_id_to_show)
objects, frame_meta = results
if args.raw_output_message:
print_raw_results(objects, next_frame_id_to_show, model.labels)
frame = frame_meta['frame']
start_time = frame_meta['start_time']
rendering_start_time = perf_counter()
frame = render_segmentation(frame, objects, visualizer,
output_transform, only_masks)
render_metrics.update(rendering_start_time)
presenter.drawGraphs(frame)
metrics.update(start_time, frame)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
if not args.no_show:
cv2.imshow('Segmentation Results', frame)
key = cv2.waitKey(1)
metrics.log_total()
log_latency_per_stage(cap.reader_metrics.get_latency(),
pipeline.preprocess_metrics.get_latency(),
pipeline.inference_metrics.get_latency(),
pipeline.postprocess_metrics.get_latency(),
render_metrics.get_latency())
for rep in presenter.reportMeans():
log.info(rep)
def main():
args = build_argparser().parse_args()
cap = open_images_capture(args.input, args.loop)
target_bgr = open_images_capture(args.target_bgr, loop=True) if args.target_bgr else None
if args.adapter == 'openvino':
plugin_config = get_user_config(args.device, args.num_streams, args.num_threads)
model_adapter = OpenvinoAdapter(create_core(), args.model, device=args.device, plugin_config=plugin_config,
max_num_requests=args.num_infer_requests)
elif args.adapter == 'ovms':
model_adapter = OVMSAdapter(args.model)
labels = ['__background__', 'person'] if args.labels is None else load_labels(args.labels)
assert len(labels), 'The file with class labels is empty'
configuration = {
'confidence_threshold': args.prob_threshold,
'resize_type': args.resize_type
}
model, need_bgr_input = get_model(model_adapter, configuration, args)
input_bgr = open_images_capture(args.background, False).read() if need_bgr_input else None
person_id = -1
for i, label in enumerate(labels):
if label == 'person':
person_id = i
break
assert person_id >= 0, 'Person class did not find in labels list.'
model.log_layers_info()
pipeline = AsyncPipeline(model)
next_frame_id = 0
next_frame_id_to_show = 0
metrics = PerformanceMetrics()
render_metrics = PerformanceMetrics()
presenter = None
output_transform = None
video_writer = cv2.VideoWriter()
while True:
if pipeline.is_ready():
# Get new image/frame
start_time = perf_counter()
frame = cap.read()
bgr = target_bgr.read() if target_bgr is not None else None
if frame is None:
if next_frame_id == 0:
raise ValueError("Can't read an image from the input")
break
if next_frame_id == 0:
output_transform = OutputTransform(frame.shape[:2], args.output_resolution)
if args.output_resolution:
output_resolution = output_transform.new_resolution
else:
output_resolution = (frame.shape[1], frame.shape[0])
presenter = monitors.Presenter(args.utilization_monitors, 55,
(round(output_resolution[0] / 4), round(output_resolution[1] / 8)))
if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'),
cap.fps(), tuple(output_resolution)):
raise RuntimeError("Can't open video writer")
# Submit for inference
data = {'src': frame, 'bgr': input_bgr} if input_bgr is not None else frame
pipeline.submit_data(data, next_frame_id, {'frame': frame, 'start_time': start_time})
next_frame_id += 1
else:
# Wait for empty request
pipeline.await_any()
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
# Process all completed requests
results = pipeline.get_result(next_frame_id_to_show)
if results:
objects, frame_meta = results
if args.raw_output_message:
print_raw_results(objects, next_frame_id_to_show)
frame = frame_meta['frame']
start_time = frame_meta['start_time']
rendering_start_time = perf_counter()
frame = render_results(frame, objects, output_resolution, bgr, person_id,
args.blur_bgr, args.show_with_original_frame)
render_metrics.update(rendering_start_time)
presenter.drawGraphs(frame)
metrics.update(start_time, frame)
if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit-1):
video_writer.write(frame)
next_frame_id_to_show += 1
if not args.no_show:
cv2.imshow('Background subtraction results', frame)
key = cv2.waitKey(1)
if key == 27 or key == 'q' or key == 'Q':
break
presenter.handleKey(key)
pipeline.await_all()
# Process completed requests
for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id):
results = pipeline.get_result(next_frame_id_to_show)
while results is None:
results = pipeline.get_result(next_frame_id_to_show)
objects, frame_meta = results
if args.raw_output_message:
print_raw_results(objects, next_frame_id_to_show, model.labels)
frame = frame_meta['frame']
start_time = frame_meta['start_time']
rendering_start_time = perf_counter()
frame = render_results(frame, objects, output_resolution, bgr, person_id,
args.blur_bgr, args.show_with_original_frame)
render_metrics.update(rendering_start_time)
presenter.drawGraphs(frame)
metrics.update(start_time, frame)
if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit-1):
video_writer.write(frame)
if not args.no_show:
cv2.imshow('Background subtraction results', frame)
cv2.waitKey(1)
metrics.log_total()
log_latency_per_stage(cap.reader_metrics.get_latency(),
pipeline.preprocess_metrics.get_latency(),
pipeline.inference_metrics.get_latency(),
pipeline.postprocess_metrics.get_latency(),
render_metrics.get_latency())
for rep in presenter.reportMeans():
log.info(rep)
def main():
args = build_argparser().parse_args()
cap = open_images_capture(args.input, args.loop)
if args.adapter == 'openvino':
plugin_config = get_user_config(args.device, args.num_streams, args.num_threads)
model_adapter = OpenvinoAdapter(create_core(), args.model, device=args.device, plugin_config=plugin_config,
max_num_requests=args.num_infer_requests)
elif args.adapter == 'ovms':
model_adapter = OVMSAdapter(args.model)
model = MonoDepthModel(model_adapter)
model.log_layers_info()
pipeline = AsyncPipeline(model)
next_frame_id = 0
next_frame_id_to_show = 0
metrics = PerformanceMetrics()
presenter = None
output_transform = None
video_writer = cv2.VideoWriter()
while True:
if pipeline.is_ready():
# Get new image/frame
start_time = perf_counter()
frame = cap.read()
if frame is None:
if next_frame_id == 0:
raise ValueError("Can't read an image from the input")
break
if next_frame_id == 0:
output_transform = OutputTransform(frame.shape[:2], args.output_resolution)
if args.output_resolution:
output_resolution = output_transform.new_resolution
else:
output_resolution = (frame.shape[1], frame.shape[0])
presenter = monitors.Presenter(args.utilization_monitors, 55,
(round(output_resolution[0] / 4), round(output_resolution[1] / 8)))
if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'),
cap.fps(), output_resolution):
raise RuntimeError("Can't open video writer")
# Submit for inference
pipeline.submit_data(frame, next_frame_id, {'start_time': start_time})
next_frame_id += 1
else:
# Wait for empty request
pipeline.await_any()
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
# Process all completed requests
results = pipeline.get_result(next_frame_id_to_show)
if results:
depth_map, frame_meta = results
depth_map = apply_color_map(depth_map, output_transform)
start_time = frame_meta['start_time']
presenter.drawGraphs(depth_map)
metrics.update(start_time, depth_map)
if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit-1):
video_writer.write(depth_map)
next_frame_id_to_show += 1
if not args.no_show:
cv2.imshow(DEMO_NAME, depth_map)
key = cv2.waitKey(1)
if key == 27 or key == 'q' or key == 'Q':
break
presenter.handleKey(key)
pipeline.await_all()
# Process completed requests
for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id):
results = pipeline.get_result(next_frame_id_to_show)
while results is None:
results = pipeline.get_result(next_frame_id_to_show)
depth_map, frame_meta = results
depth_map = apply_color_map(depth_map, output_transform)
start_time = frame_meta['start_time']
presenter.drawGraphs(depth_map)
metrics.update(start_time, depth_map)
if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit-1):
video_writer.write(depth_map)
if not args.no_show:
cv2.imshow(DEMO_NAME, depth_map)
key = cv2.waitKey(1)
if key == 27 or key == 'q' or key == 'Q':
break
presenter.handleKey(key)
metrics.log_total()
for rep in presenter.reportMeans():
log.info(rep)
def main():
args = build_argparser().parse_args()
if args.architecture_type != 'yolov4' and args.anchors:
log.warning(
'The "--anchors" option works only for "-at==yolov4". Option will be omitted'
)
if args.architecture_type != 'yolov4' and args.masks:
log.warning(
'The "--masks" option works only for "-at==yolov4". Option will be omitted'
)
if args.architecture_type not in ['nanodet', 'nanodet-plus'
] and args.num_classes:
log.warning(
'The "--num_classes" option works only for "-at==nanodet" and "-at==nanodet-plus". Option will be omitted'
)
cap = open_images_capture(args.input, args.loop)
if args.adapter == 'openvino':
plugin_config = get_user_config(args.device, args.num_streams,
args.num_threads)
model_adapter = OpenvinoAdapter(
create_core(),
args.model,
device=args.device,
plugin_config=plugin_config,
max_num_requests=args.num_infer_requests,
model_parameters={'input_layouts': args.layout})
elif args.adapter == 'ovms':
model_adapter = OVMSAdapter(args.model)
configuration = {
'resize_type': args.resize_type,
'mean_values': args.mean_values,
'scale_values': args.scale_values,
'reverse_input_channels': args.reverse_input_channels,
'path_to_labels': args.labels,
'confidence_threshold': args.prob_threshold,
'input_size': args.input_size, # The CTPN specific
'num_classes':
args.num_classes, # The NanoDet and NanoDetPlus specific
}
model = DetectionModel.create_model(args.architecture_type, model_adapter,
configuration)
model.log_layers_info()
detector_pipeline = AsyncPipeline(model)
next_frame_id = 0
next_frame_id_to_show = 0
palette = ColorPalette(len(model.labels) if model.labels else 100)
metrics = PerformanceMetrics()
render_metrics = PerformanceMetrics()
presenter = None
output_transform = None
video_writer = cv2.VideoWriter()
while True:
if detector_pipeline.callback_exceptions:
raise detector_pipeline.callback_exceptions[0]
# Process all completed requests
results = detector_pipeline.get_result(next_frame_id_to_show)
if results:
objects, frame_meta = results
frame = frame_meta['frame']
start_time = frame_meta['start_time']
if len(objects) and args.raw_output_message:
print_raw_results(objects, model.labels, next_frame_id_to_show)
presenter.drawGraphs(frame)
rendering_start_time = perf_counter()
frame = draw_detections(frame, objects, palette, model.labels,
output_transform)
render_metrics.update(rendering_start_time)
metrics.update(start_time, frame)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
next_frame_id_to_show += 1
if not args.no_show:
cv2.imshow('Detection Results', frame)
key = cv2.waitKey(1)
ESC_KEY = 27
# Quit.
if key in {ord('q'), ord('Q'), ESC_KEY}:
break
presenter.handleKey(key)
continue
if detector_pipeline.is_ready():
# Get new image/frame
start_time = perf_counter()
frame = cap.read()
if frame is None:
if next_frame_id == 0:
raise ValueError("Can't read an image from the input")
break
if next_frame_id == 0:
output_transform = OutputTransform(frame.shape[:2],
args.output_resolution)
if args.output_resolution:
output_resolution = output_transform.new_resolution
else:
output_resolution = (frame.shape[1], frame.shape[0])
presenter = monitors.Presenter(
args.utilization_monitors, 55,
(round(output_resolution[0] / 4),
round(output_resolution[1] / 8)))
if args.output and not video_writer.open(
args.output, cv2.VideoWriter_fourcc(*'MJPG'),
cap.fps(), output_resolution):
raise RuntimeError("Can't open video writer")
# Submit for inference
detector_pipeline.submit_data(frame, next_frame_id, {
'frame': frame,
'start_time': start_time
})
next_frame_id += 1
else:
# Wait for empty request
detector_pipeline.await_any()
detector_pipeline.await_all()
if detector_pipeline.callback_exceptions:
raise detector_pipeline.callback_exceptions[0]
# Process completed requests
for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id):
results = detector_pipeline.get_result(next_frame_id_to_show)
objects, frame_meta = results
frame = frame_meta['frame']
start_time = frame_meta['start_time']
if len(objects) and args.raw_output_message:
print_raw_results(objects, model.labels, next_frame_id_to_show)
presenter.drawGraphs(frame)
rendering_start_time = perf_counter()
frame = draw_detections(frame, objects, palette, model.labels,
output_transform)
render_metrics.update(rendering_start_time)
metrics.update(start_time, frame)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
if not args.no_show:
cv2.imshow('Detection Results', frame)
key = cv2.waitKey(1)
ESC_KEY = 27
# Quit.
if key in {ord('q'), ord('Q'), ESC_KEY}:
break
presenter.handleKey(key)
metrics.log_total()
log_latency_per_stage(cap.reader_metrics.get_latency(),
detector_pipeline.preprocess_metrics.get_latency(),
detector_pipeline.inference_metrics.get_latency(),
detector_pipeline.postprocess_metrics.get_latency(),
render_metrics.get_latency())
for rep in presenter.reportMeans():
log.info(rep)
def main():
args = build_argparser().parse_args()
paragraphs = get_paragraphs(args.input)
preprocessing_start_time = perf_counter()
vocab = load_vocab_file(args.vocab)
log.debug("Loaded vocab file from {}, get {} tokens".format(
args.vocab, len(vocab)))
# get context as a string (as we might need it's length for the sequence reshape)
context = '\n'.join(paragraphs)
visualizer = Visualizer(context, args.colors)
# encode context into token ids list
c_tokens = text_to_tokens(context.lower(), vocab)
total_latency = (perf_counter() - preprocessing_start_time) * 1e3
if args.adapter == 'openvino':
plugin_config = get_user_config(args.device, args.num_streams,
args.num_threads)
model_adapter = OpenvinoAdapter(
create_core(),
args.model,
device=args.device,
plugin_config=plugin_config,
max_num_requests=args.num_infer_requests)
elif args.adapter == 'ovms':
model_adapter = OVMSAdapter(args.model)
config = {
'vocab': vocab,
'input_names': args.input_names,
'output_names': args.output_names,
'max_answer_token_num': args.max_answer_token_num,
'squad_ver': args.model_squad_ver
}
model = BertQuestionAnswering(model_adapter, config)
if args.reshape:
# find the closest multiple of 64, if it is smaller than current network's sequence length, do reshape
new_length = min(
model.max_length,
int(
np.ceil(
(len(c_tokens[0]) + args.max_question_token_num) / 64) *
64))
if new_length < model.max_length:
try:
model.reshape(new_length)
except RuntimeError:
log.error(
"Failed to reshape the network, please retry the demo without '-r' option"
)
sys.exit(-1)
else:
log.debug(
"\tSkipping network reshaping,"
" as (context length + max question length) exceeds the current (input) network sequence length"
)
model.log_layers_info()
pipeline = AsyncPipeline(model)
if args.questions:
def questions():
for question in args.questions:
log.info("\n\tQuestion: {}".format(question))
yield question
else:
def questions():
while True:
yield input('\n\tType a question (empty string to exit): ')
for question in questions():
if not question.strip():
break
answers = []
next_window_id = 0
next_window_id_to_show = 0
start_time = perf_counter()
q_tokens_id, _ = text_to_tokens(question.lower(), vocab)
source = ContextSource(q_tokens_id, c_tokens, model.max_length)
while True:
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
results = pipeline.get_result(next_window_id_to_show)
if results:
next_window_id_to_show += 1
update_answers_list(answers, results[0])
continue
if pipeline.is_ready():
if source.is_over():
break
pipeline.submit_data(source.get_data(), next_window_id, None)
next_window_id += 1
else:
pipeline.await_any()
pipeline.await_all()
for window_id in range(next_window_id_to_show, next_window_id):
results = pipeline.get_result(window_id)
while results is None:
results = pipeline.get_result(window_id)
update_answers_list(answers, results[0])
visualizer.show_answers(answers)
total_latency += (perf_counter() - start_time) * 1e3
log.info("Metrics report:")
log.info("\tLatency: {:.1f} ms".format(total_latency))
def main():
args = build_argparser().parse_args()
cap = open_images_capture(args.input, args.loop)
delay = int(cap.get_type() in {'VIDEO', 'CAMERA'})
if args.adapter == 'openvino':
plugin_config = get_user_config(args.device, args.num_streams,
args.num_threads)
model_adapter = OpenvinoAdapter(
create_core(),
args.model,
device=args.device,
plugin_config=plugin_config,
max_num_requests=args.num_infer_requests)
elif args.adapter == 'ovms':
model_adapter = OVMSAdapter(args.model)
config = {
'mean_values': args.mean_values,
'scale_values': args.scale_values,
'reverse_input_channels': args.reverse_input_channels,
'topk': args.topk,
'path_to_labels': args.labels
}
model = Classification(model_adapter, config)
model.log_layers_info()
async_pipeline = AsyncPipeline(model)
next_frame_id = 0
next_frame_id_to_show = 0
metrics = PerformanceMetrics()
render_metrics = PerformanceMetrics()
presenter = None
output_transform = None
video_writer = cv2.VideoWriter()
ESC_KEY = 27
key = -1
while True:
if async_pipeline.callback_exceptions:
raise async_pipeline.callback_exceptions[0]
# Process all completed requests
results = async_pipeline.get_result(next_frame_id_to_show)
if results:
classifications, frame_meta = results
frame = frame_meta['frame']
start_time = frame_meta['start_time']
if args.raw_output_message:
print_raw_results(classifications, next_frame_id_to_show)
presenter.drawGraphs(frame)
rendering_start_time = perf_counter()
frame = draw_labels(frame, classifications, output_transform)
if delay or args.no_show:
render_metrics.update(rendering_start_time)
metrics.update(start_time, frame)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
next_frame_id_to_show += 1
if not args.no_show:
cv2.imshow('Classification Results', frame)
key = cv2.waitKey(delay)
# Quit.
if key in {ord('q'), ord('Q'), ESC_KEY}:
break
presenter.handleKey(key)
continue
if async_pipeline.is_ready():
# Get new image/frame
start_time = perf_counter()
frame = cap.read()
if frame is None:
if next_frame_id == 0:
raise ValueError("Can't read an image from the input")
break
if next_frame_id == 0:
output_transform = OutputTransform(frame.shape[:2],
args.output_resolution)
if args.output_resolution:
output_resolution = output_transform.new_resolution
else:
output_resolution = (frame.shape[1], frame.shape[0])
presenter = monitors.Presenter(
args.utilization_monitors, 55,
(round(output_resolution[0] / 4),
round(output_resolution[1] / 8)))
if args.output and not video_writer.open(
args.output, cv2.VideoWriter_fourcc(*'MJPG'),
cap.fps(), output_resolution):
raise RuntimeError("Can't open video writer")
# Submit for inference
async_pipeline.submit_data(frame, next_frame_id, {
'frame': frame,
'start_time': start_time
})
next_frame_id += 1
else:
# Wait for empty request
async_pipeline.await_any()
async_pipeline.await_all()
if key not in {ord('q'), ord('Q'), ESC_KEY}:
# Process completed requests
for next_frame_id_to_show in range(next_frame_id_to_show,
next_frame_id):
results = async_pipeline.get_result(next_frame_id_to_show)
while results is None:
results = async_pipeline.get_result(next_frame_id_to_show)
classifications, frame_meta = results
frame = frame_meta['frame']
start_time = frame_meta['start_time']
if args.raw_output_message:
print_raw_results(classifications, next_frame_id_to_show)
presenter.drawGraphs(frame)
rendering_start_time = perf_counter()
frame = draw_labels(frame, classifications, output_transform)
if delay or args.no_show:
render_metrics.update(rendering_start_time)
metrics.update(start_time, frame)
if video_writer.isOpened() and (
args.output_limit <= 0
or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
if not args.no_show:
cv2.imshow('Classification Results', frame)
key = cv2.waitKey(delay)
# Quit.
if key in {ord('q'), ord('Q'), ESC_KEY}:
break
presenter.handleKey(key)
if delay or args.no_show:
metrics.log_total()
log_latency_per_stage(cap.reader_metrics.get_latency(),
async_pipeline.preprocess_metrics.get_latency(),
async_pipeline.inference_metrics.get_latency(),
async_pipeline.postprocess_metrics.get_latency(),
render_metrics.get_latency())
for rep in presenter.reportMeans():
log.info(rep)
def main():
args = build_argparser().parse_args()
cap = open_images_capture(args.input, args.loop)
if args.adapter == 'openvino':
plugin_config = get_user_config(args.device, args.num_streams, args.num_threads)
model_adapter = OpenvinoAdapter(create_core(), args.model, device=args.device, plugin_config=plugin_config,
max_num_requests=args.num_infer_requests,
model_parameters={'input_layouts': args.layout})
elif args.adapter == 'ovms':
model_adapter = OVMSAdapter(args.model)
configuration = {
'confidence_threshold': args.prob_threshold,
'path_to_labels': args.labels,
}
model = get_model(model_adapter, configuration)
model.log_layers_info()
pipeline = AsyncPipeline(model)
next_frame_id = 0
next_frame_id_to_show = 0
tracker = None
if not args.no_track and cap.get_type() in {'VIDEO', 'CAMERA'}:
tracker = StaticIOUTracker()
visualizer = InstanceSegmentationVisualizer(model.labels, args.show_boxes, args.show_scores)
metrics = PerformanceMetrics()
render_metrics = PerformanceMetrics()
presenter = None
output_transform = None
video_writer = cv2.VideoWriter()
while True:
if pipeline.is_ready():
# Get new image/frame
start_time = perf_counter()
frame = cap.read()
if frame is None:
if next_frame_id == 0:
raise ValueError("Can't read an image from the input")
break
if next_frame_id == 0:
output_transform = OutputTransform(frame.shape[:2], args.output_resolution)
if args.output_resolution:
output_resolution = output_transform.new_resolution
else:
output_resolution = (frame.shape[1], frame.shape[0])
presenter = monitors.Presenter(args.utilization_monitors, 55,
(round(output_resolution[0] / 4), round(output_resolution[1] / 8)))
if args.output and not video_writer.open(args.output, cv2.VideoWriter_fourcc(*'MJPG'),
cap.fps(), tuple(output_resolution)):
raise RuntimeError("Can't open video writer")
# Submit for inference
pipeline.submit_data(frame, next_frame_id, {'frame': frame, 'start_time': start_time})
next_frame_id += 1
else:
# Wait for empty request
pipeline.await_any()
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
# Process all completed requests
results = pipeline.get_result(next_frame_id_to_show)
if results:
(scores, classes, boxes, masks), frame_meta = results
frame = frame_meta['frame']
start_time = frame_meta['start_time']
if args.raw_output_message:
print_raw_results(boxes, classes, scores, next_frame_id_to_show)
rendering_start_time = perf_counter()
masks_tracks_ids = tracker(masks, classes) if tracker else None
frame = visualizer(frame, boxes, classes, scores, masks, masks_tracks_ids)
render_metrics.update(rendering_start_time)
presenter.drawGraphs(frame)
metrics.update(start_time, frame)
if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
next_frame_id_to_show += 1
if not args.no_show:
cv2.imshow('Instance Segmentation results', frame)
key = cv2.waitKey(1)
if key == 27 or key == 'q' or key == 'Q':
break
presenter.handleKey(key)
pipeline.await_all()
if pipeline.callback_exceptions:
raise pipeline.callback_exceptions[0]
# Process completed requests
for next_frame_id_to_show in range(next_frame_id_to_show, next_frame_id):
results = pipeline.get_result(next_frame_id_to_show)
(scores, classes, boxes, masks), frame_meta = results
frame = frame_meta['frame']
start_time = frame_meta['start_time']
if args.raw_output_message:
print_raw_results(boxes, classes, scores, next_frame_id_to_show)
rendering_start_time = perf_counter()
masks_tracks_ids = tracker(masks, classes) if tracker else None
frame = visualizer(frame, boxes, classes, scores, masks, masks_tracks_ids)
render_metrics.update(rendering_start_time)
presenter.drawGraphs(frame)
metrics.update(start_time, frame)
if video_writer.isOpened() and (args.output_limit <= 0 or next_frame_id_to_show <= args.output_limit - 1):
video_writer.write(frame)
if not args.no_show:
cv2.imshow('Instance Segmentation results', frame)
cv2.waitKey(1)
metrics.log_total()
log_latency_per_stage(cap.reader_metrics.get_latency(),
pipeline.preprocess_metrics.get_latency(),
pipeline.inference_metrics.get_latency(),
pipeline.postprocess_metrics.get_latency(),
render_metrics.get_latency())
for rep in presenter.reportMeans():
log.info(rep)
def main():
args = build_argparser().parse_args()
paragraphs = get_paragraphs(args.input)
vocab_start_time = perf_counter()
vocab = load_vocab_file(args.vocab)
log.debug("Loaded vocab file from {}, get {} tokens".format(
args.vocab, len(vocab)))
visualizer = Visualizer(args.colors)
total_latency = (perf_counter() - vocab_start_time) * 1e3
ie = create_core()
plugin_config = get_user_config(args.device, args.num_streams,
args.num_threads)
model_emb_adapter = OpenvinoAdapter(
ie,
args.model_emb,
device=args.device,
plugin_config=plugin_config,
max_num_requests=args.num_infer_requests)
model_emb = BertEmbedding(model_emb_adapter, {
'vocab': vocab,
'input_names': args.input_names_emb
})
model_emb.log_layers_info()
# reshape BertEmbedding model to infer short questions and long contexts
max_len_context = 384
max_len_question = 32
for new_length in [max_len_question, max_len_context]:
model_emb.reshape(new_length)
if new_length == max_len_question:
emb_exec_net = ie.load_network(model_emb_adapter.net, args.device)
else:
emb_pipeline = AsyncPipeline(model_emb)
if args.model_qa:
model_qa_adapter = OpenvinoAdapter(
ie,
args.model_qa,
device=args.device,
plugin_config=plugin_config,
max_num_requests=args.num_infer_requests)
config = {
'vocab': vocab,
'input_names': args.input_names_qa,
'output_names': args.output_names_qa,
'max_answer_token_num': args.max_answer_token_num,
'squad_ver': args.model_qa_squad_ver
}
model_qa = BertQuestionAnswering(model_qa_adapter, config)
model_qa.log_layers_info()
qa_pipeline = AsyncPipeline(model_qa)
log.info("\t\tStage 1 (Calc embeddings for the context)")
contexts_all = []
start_time = perf_counter()
# get context as string and then encode it into token id list
# calculate number of tokens for context in each request.
# reserve 3 positions for special tokens [CLS] q_tokens [SEP] c_tokens [SEP]
if args.model_qa:
# to make context be able to pass model_qa together with question
c_window_len = model_qa.max_length - (max_len_question + 3)
else:
# to make context be able to pass model_emb without question
c_window_len = max_len_context - 2
def calc_question_embedding(tokens_id):
num = min(max_len_question - 2, len(tokens_id))
inputs, _ = model_emb.preprocess((tokens_id[:num], max_len_question))
raw_result = emb_exec_net.infer(inputs)
return model_emb.postprocess(raw_result, None)
source = ContextSource(paragraphs, vocab, c_window_len)
next_window_id = 0
next_window_id_to_show = 0
contexts_all = []
while True:
if emb_pipeline.callback_exceptions:
raise emb_pipeline.callback_exceptions[0]
results = emb_pipeline.get_result(next_window_id_to_show)
if results:
embedding, meta = results
meta['c_data'].emb = embedding
contexts_all.append(meta['c_data'])
next_window_id_to_show += 1
continue
if emb_pipeline.is_ready():
if source.is_over():
break
c_data = source.get_data()
num = min(max_len_context - 2, len(c_data.c_tokens_id))
emb_pipeline.submit_data(
(c_data.c_tokens_id[:num], max_len_context), next_window_id,
{'c_data': c_data})
next_window_id += 1
else:
emb_pipeline.await_any()
emb_pipeline.await_all()
for window_id in range(next_window_id_to_show, next_window_id):
results = emb_pipeline.get_result(window_id)
while results is None:
results = emb_pipeline.get_result(window_id)
embedding, meta = results
meta['c_data'].emb = embedding
contexts_all.append(meta['c_data'])
next_window_id_to_show += 1
total_latency += (perf_counter() - start_time) * 1e3
context_embeddings_time = total_latency
if args.questions:
def questions():
for question in args.questions:
log.info("\n\tQuestion: {}".format(question))
yield question
else:
def questions():
while True:
yield input('\n\tType a question (empty string to exit): ')
for question in questions():
if not question.strip():
break
start_time = perf_counter()
log.info(
"\t\tStage 2 (Calc question embedding and compare with {} context embeddings)"
.format(len(contexts_all)))
q_tokens_id, _ = text_to_tokens(question.lower(), vocab)
q_emb = calc_question_embedding(q_tokens_id)
distances = [(np.linalg.norm(context.emb - q_emb, 2), context)
for context in contexts_all]
distances.sort(key=lambda x: x[0])
keep_num = min(args.best_n, len(distances))
distances_filtered = distances[:keep_num]
log.info(
"The closest {} contexts to question filtered from {} context embeddings:"
.format(keep_num, len(distances)))
visualizer.show_closest_contexts(distances_filtered)
if args.model_qa:
answers = []
next_context_id = 0
next_context_id_to_show = 0
while True:
if qa_pipeline.callback_exceptions:
raise qa_pipeline.callback_exceptions[0]
results = qa_pipeline.get_result(next_context_id_to_show)
if results:
next_context_id_to_show += 1
output, meta = results
update_answers_list(answers, output, meta['c_data'])
continue
if qa_pipeline.is_ready():
if next_context_id == len(distances_filtered):
break
_, c_data = distances_filtered[next_context_id]
qa_pipeline.submit_data((c_data, q_tokens_id),
next_context_id,
{'c_data': c_data})
next_context_id += 1
else:
qa_pipeline.await_any()
qa_pipeline.await_all()
for context_id in range(next_context_id_to_show, next_context_id):
results = qa_pipeline.get_result(context_id)
while results is None:
results = qa_pipeline.get_result(context_id)
output, meta = results
update_answers_list(answers, output, meta['c_data'])
log.info(
"\t\tStage 3 (Show top 3 answers from {} closest contexts of Stage 1)"
.format(len(answers)))
answers = sorted(answers, key=lambda x: -x[0])[:3]
visualizer.show_answers(answers)
total_latency += (perf_counter() - start_time) * 1e3
log.info("Metrics report:")
log.info("\tContext embeddings latency (stage 1): {:.1f} ms".format(
context_embeddings_time))
log.info("\tLatency (all stages): {:.1f} ms".format(total_latency))