def run(self, image):
if image.shape[1:2] != (self.input_height, self.input_width):
img = cv2.resize(image, (self.input_width, self.input_height))
img = preprocess(img)
self.interpreter.set_tensor(self.tensor_index, img)
self.interpreter.invoke()
return self.get_output_tensors()
def main(args):
video, video_writer, frame_count = init_video_file_capture(
args.video_file_path, args.output_video_file_path)
executor = ArmnnNetworkExecutor(args.model_file_path,
args.preferred_backends)
process_output, resize_factor = get_model_processing(
args.model_name, video, executor.input_binding_info)
labels = dict_labels(args.label_path, include_rgb=True)
for _ in tqdm(frame_count, desc='Processing frames'):
frame_present, frame = video.read()
if not frame_present:
continue
input_tensors = preprocess(frame, executor.input_binding_info)
output_result = executor.run(input_tensors)
detections = process_output(output_result)
draw_bounding_boxes(frame, detections, resize_factor, labels)
video_writer.write(frame)
print('Finished processing frames')
video.release(), video_writer.release()
def run_style_predict(self, style_image):
"""
Creates bottleneck tensor for a given style image.
Args:
style_image: an image to create the style bottleneck
Returns:
style bottleneck tensor
"""
# The style image has to be preprocessed to (1, 256, 256, 3)
preprocessed_style_image = cv_utils.preprocess(
style_image,
self.style_predict_executor.get_data_type(),
self.style_predict_executor.get_shape(),
True,
keep_aspect_ratio=False)
# output[0] is the style bottleneck tensor
style_bottleneck = self.style_predict_executor.run(
[preprocessed_style_image])[0]
return style_bottleneck
def run_style_transfer(self, content_image):
"""
Runs inference for given content_image and style bottleneck to create a stylized image.
Args:
content_image:a content image to stylize
"""
# The content image has to be preprocessed to (1, 384, 384, 3)
preprocessed_style_image = cv_utils.preprocess(
content_image,
np.float32,
self.style_transfer_executor.get_shape(),
True,
keep_aspect_ratio=False)
# Transform content image. output[0] is the stylized image
stylized_image = self.style_transfer_executor.run(
[preprocessed_style_image, self.style_bottleneck])[0]
post_stylized_image = style_transfer_postprocess(
stylized_image, content_image.shape)
return post_stylized_image
def main(args):
video = init_video_stream_capture(args.video_source)
executor = ArmnnNetworkExecutor(args.model_file_path,
args.preferred_backends)
process_output, resize_factor = get_model_processing(
args.model_name, video, executor.input_binding_info)
labels = dict_labels(args.label_path, include_rgb=True)
while True:
frame_present, frame = video.read()
frame = cv2.flip(frame, 1) # Horizontally flip the frame
if not frame_present:
raise RuntimeError('Error reading frame from video stream')
input_tensors = preprocess(frame, executor.input_binding_info)
print("Running inference...")
output_result = executor.run(input_tensors)
detections = process_output(output_result)
draw_bounding_boxes(frame, detections, resize_factor, labels)
cv2.imshow('PyArmNN Object Detection Demo', frame)
if cv2.waitKey(1) == 27:
print('\nExit key activated. Closing video...')
break
video.release(), cv2.destroyAllWindows()
def main(args):
enable_profile = args.profiling_enabled == "true"
action_profiler = Profiling(enable_profile)
overall_profiler = Profiling(enable_profile)
overall_profiler.profiling_start()
action_profiler.profiling_start()
if args.tflite_delegate_path is not None:
from network_executor_tflite import TFLiteNetworkExecutor as NetworkExecutor
exec_input_args = (args.model_file_path, args.preferred_backends,
args.tflite_delegate_path)
else:
from network_executor import ArmnnNetworkExecutor as NetworkExecutor
exec_input_args = (args.model_file_path, args.preferred_backends)
executor = NetworkExecutor(*exec_input_args)
action_profiler.profiling_stop_and_print_us("Executor initialization")
action_profiler.profiling_start()
video, video_writer, frame_count = init_video_file_capture(
args.video_file_path, args.output_video_file_path)
process_output, resize_factor = get_model_processing(
args.model_name, video, executor.get_shape())
action_profiler.profiling_stop_and_print_us("Video initialization")
labels = dict_labels(args.label_path, include_rgb=True)
if all(element is not None for element in [
args.style_predict_model_file_path,
args.style_transfer_model_file_path, args.style_image_path,
args.style_transfer_class
]):
style_image = cv2.imread(args.style_image_path)
action_profiler.profiling_start()
style_transfer_executor = style_transfer.StyleTransfer(
args.style_predict_model_file_path,
args.style_transfer_model_file_path, style_image,
args.preferred_backends, args.tflite_delegate_path)
action_profiler.profiling_stop_and_print_us(
"Style Transfer Executor initialization")
for _ in tqdm(frame_count, desc='Processing frames'):
frame_present, frame = video.read()
if not frame_present:
continue
model_name = args.model_name
if model_name == "ssd_mobilenet_v1":
input_data = preprocess(frame, executor.get_data_type(),
executor.get_shape(), True)
else:
input_data = preprocess(frame, executor.get_data_type(),
executor.get_shape(), False)
action_profiler.profiling_start()
output_result = executor.run([input_data])
action_profiler.profiling_stop_and_print_us("Running inference")
detections = process_output(output_result)
if all(element is not None for element in [
args.style_predict_model_file_path,
args.style_transfer_model_file_path, args.style_image_path,
args.style_transfer_class
]):
action_profiler.profiling_start()
frame = style_transfer.create_stylized_detection(
style_transfer_executor, args.style_transfer_class, frame,
detections, resize_factor, labels)
action_profiler.profiling_stop_and_print_us(
"Running Style Transfer")
else:
draw_bounding_boxes(frame, detections, resize_factor, labels)
video_writer.write(frame)
print('Finished processing frames')
overall_profiler.profiling_stop_and_print_us("Total compute time")
video.release(), video_writer.release()
def main(args):
enable_profile = args.profiling_enabled == "true"
action_profiler = Profiling(enable_profile)
action_profiler.profiling_start()
if args.tflite_delegate_path is not None:
from network_executor_tflite import TFLiteNetworkExecutor as NetworkExecutor
exec_input_args = (args.model_file_path, args.preferred_backends,
args.tflite_delegate_path)
else:
from network_executor import ArmnnNetworkExecutor as NetworkExecutor
exec_input_args = (args.model_file_path, args.preferred_backends)
executor = NetworkExecutor(*exec_input_args)
action_profiler.profiling_stop_and_print_us("Executor initialization")
action_profiler.profiling_start()
video = init_video_stream_capture(args.video_source)
action_profiler.profiling_stop_and_print_us("Video initialization")
model_name = args.model_name
process_output, resize_factor = get_model_processing(
args.model_name, video, executor.get_shape())
labels = dict_labels(args.label_path, include_rgb=True)
if all(element is not None for element in [
args.style_predict_model_file_path,
args.style_transfer_model_file_path, args.style_image_path,
args.style_transfer_class
]):
style_image = cv2.imread(args.style_image_path)
action_profiler.profiling_start()
style_transfer_executor = style_transfer.StyleTransfer(
args.style_predict_model_file_path,
args.style_transfer_model_file_path, style_image,
args.preferred_backends, args.tflite_delegate_path)
action_profiler.profiling_stop_and_print_us(
"Style Transfer Executor initialization")
while True:
frame_present, frame = video.read()
frame = cv2.flip(frame, 1) # Horizontally flip the frame
if not frame_present:
raise RuntimeError('Error reading frame from video stream')
action_profiler.profiling_start()
if model_name == "ssd_mobilenet_v1":
input_data = preprocess(frame, executor.get_data_type(),
executor.get_shape(), True)
else:
input_data = preprocess(frame, executor.get_data_type(),
executor.get_shape(), False)
output_result = executor.run([input_data])
if not enable_profile:
print("Running inference...")
action_profiler.profiling_stop_and_print_us("Running inference...")
detections = process_output(output_result)
if all(element is not None for element in [
args.style_predict_model_file_path,
args.style_transfer_model_file_path, args.style_image_path,
args.style_transfer_class
]):
action_profiler.profiling_start()
frame = style_transfer.create_stylized_detection(
style_transfer_executor, args.style_transfer_class, frame,
detections, resize_factor, labels)
action_profiler.profiling_stop_and_print_us(
"Running Style Transfer")
else:
draw_bounding_boxes(frame, detections, resize_factor, labels)
cv2.imshow('PyArmNN Object Detection Demo', frame)
if cv2.waitKey(1) == 27:
print('\nExit key activated. Closing video...')
break
video.release(), cv2.destroyAllWindows()