def test_detect(self):
modules.load_known_modules()
detector = ImageObjectDetector.create("example_detector")
image = Image()
image_container = ImageContainer(image)
detections = detector.detect(image_container)
nose.tools.ok_(detections is not None,
"Unexpected empty detections" )
nose.tools.assert_equal(len(detections), 1)
def test_config(self):
modules.load_known_modules()
detector = ImageObjectDetector.create("example_detector")
# Verify that 6 config values are present in example_detector
nose.tools.assert_equal(len(detector.get_configuration()), 6)
test_cfg = _dummy_detector_cfg()
# Verify that the detector has different configuration before setting to test
nose.tools.assert_not_equal(detector.check_configuration(test_cfg), False)
detector.set_configuration(test_cfg)
# Verify that the config value is being set properly
nose.tools.assert_equal(detector.check_configuration(test_cfg), True)
def test_nested_config(self):
modules.load_known_modules()
detector = ImageObjectDetector.create("example_detector")
nested_cfg = config.empty_config()
ImageObjectDetector.get_nested_algo_configuration( "detector",
nested_cfg,
detector )
# Verify that test cfg is set to configuration inside detector
# nested configuration uses the name of a detector as an additional configuration
# key thus it is checked against 7 rather than 6
#nose.tools.assert_equal(len(nested_cfg), 7)
#test_cfg = _dummy_detector_cfg()
#test_cfg.set_value("example_detector:type", "example_detector")
#ImageObjectDetector.set_nested_algo_configuration( "example_detector",
# test_cfg,
# detector )
nose.tools.assert_equal(ImageObjectDetector.check_nested_algo_configuration(
"detector",
nested_cfg), True)
def test_bad_detect(self):
modules.load_known_modules()
detector = ImageObjectDetector.create("example_detector")
detector.detect("Image")
def test_bad_create(self):
# Should fail to create an algorithm without a factory
ImageObjectDetector.create("")
def set_configuration(self, cfg_in):
cfg = self.get_configuration()
cfg.merge_config(cfg_in)
# Read configs from file
self._identifier = str(cfg.get_value("identifier"))
self._mode = str(cfg.get_value("mode"))
self._arch = str(cfg.get_value("arch"))
self._seed_model = str(cfg.get_value("seed_model"))
self._train_directory = str(cfg.get_value("train_directory"))
self._output_directory = str(cfg.get_value("output_directory"))
self._output_prefix = str(cfg.get_value("output_prefix"))
self._output_plots = strtobool(cfg.get_value("output_plots"))
self._pipeline_template = str(cfg.get_value("pipeline_template"))
self._gpu_count = int(cfg.get_value("gpu_count"))
self._gt_frames_only = strtobool(cfg.get_value("gt_frames_only"))
self._augmentation = str(cfg.get_value("augmentation"))
self._chip_width = str(cfg.get_value("chip_width"))
self._chip_overlap = str(cfg.get_value("chip_overlap"))
self._chip_method = str(cfg.get_value("chip_method"))
self._max_epochs = str(cfg.get_value("max_epochs"))
self._batch_size = str(cfg.get_value("batch_size"))
self._learning_rate = str(cfg.get_value("learning_rate"))
self._timeout = str(cfg.get_value("timeout"))
self._backbone = str(cfg.get_value("backbone"))
self._pipeline_template = str(cfg.get_value("pipeline_template"))
self._max_scale_wrt_chip = float(cfg.get_value("max_scale_wrt_chip"))
self._no_format = strtobool(cfg.get_value("no_format"))
self._allow_unicode = str(cfg.get_value("allow_unicode"))
self._aux_image_labels = str(cfg.get_value("aux_image_labels"))
self._aux_image_extensions = str(cfg.get_value("aux_image_extensions"))
self._area_lower_bound = float(cfg.get_value("area_lower_bound"))
self._area_upper_bound = float(cfg.get_value("area_upper_bound"))
self._border_exclude = float(cfg.get_value("border_exclude"))
self._detector_model = str(cfg.get_value("detector_model"))
self._max_neg_per_frame = float(cfg.get_value("max_neg_per_frame"))
self._negative_category = str(cfg.get_value("negative_category"))
self._reduce_category = str(cfg.get_value("reduce_category"))
self._scale_type_file = str(cfg.get_value("scale_type_file"))
# Check GPU-related variables
gpu_memory_available = 0
gpu_param_adj = 1
if torch.cuda.is_available():
if self._gpu_count < 0:
self._gpu_count = torch.cuda.device_count()
gpu_param_adj = self._gpu_count
for i in range(self._gpu_count):
single_gpu_mem = torch.cuda.get_device_properties(
i).total_memory
if gpu_memory_available == 0:
gpu_memory_available = single_gpu_mem
else:
gpu_memory_available = min(gpu_memory_available,
single_gpu_mem)
if self._mode == "detector":
if self._batch_size == "auto":
if len(self._aux_image_labels) > 0:
if gpu_memory_available >= 22e9:
self._batch_size = str(2 * gpu_param_adj)
else:
self._batch_size = str(1 * gpu_param_adj)
elif gpu_memory_available > 9.5e9:
self._batch_size = str(4 * gpu_param_adj)
elif gpu_memory_available >= 7.5e9:
self._batch_size = str(3 * gpu_param_adj)
elif gpu_memory_available >= 4.5e9:
self._batch_size = str(2 * gpu_param_adj)
else:
self._batch_size = str(1 * gpu_param_adj)
if self._learning_rate == "auto":
self._learning_rate = str(1e-3)
elif self._mode == "frame_classifier" or self._mode == "detection_refiner":
if self._batch_size == "auto":
if gpu_memory_available > 9.5e9:
self._batch_size = str(64 * gpu_param_adj)
elif gpu_memory_available >= 7.5e9:
self._batch_size = str(32 * gpu_param_adj)
elif gpu_memory_available >= 4.5e9:
self._batch_size = str(16 * gpu_param_adj)
else:
self._batch_size = str(8 * gpu_param_adj)
if self._learning_rate == "auto":
self._learning_rate = str(5e-3)
else:
print("Invalid mode string " + self._mode)
return False
# Make required directories and file streams
if self._train_directory is not None:
if not os.path.exists(self._train_directory):
os.mkdir(self._train_directory)
self._training_file = os.path.join(self._train_directory,
self._tmp_training_file)
self._validation_file = os.path.join(self._train_directory,
self._tmp_validation_file)
self._chip_directory = os.path.join(self._train_directory,
"image_chips")
else:
self._training_file = self._tmp_training_file
self._validation_file = self._tmp_validation_file
if self._output_directory is not None:
if not os.path.exists(self._output_directory):
os.mkdir(self._output_directory)
from kwiver.vital.modules import load_known_modules
load_known_modules()
if not self._no_format:
self._training_writer = \
DetectedObjectSetOutput.create( "coco" )
self._validation_writer = \
DetectedObjectSetOutput.create( "coco" )
writer_conf = self._training_writer.get_configuration()
writer_conf.set_value("aux_image_labels", self._aux_image_labels)
writer_conf.set_value("aux_image_extensions",
self._aux_image_extensions)
self._training_writer.set_configuration(writer_conf)
writer_conf = self._validation_writer.get_configuration()
writer_conf.set_value("aux_image_labels", self._aux_image_labels)
writer_conf.set_value("aux_image_extensions",
self._aux_image_extensions)
self._validation_writer.set_configuration(writer_conf)
self._training_writer.open(self._training_file)
self._validation_writer.open(self._validation_file)
# Set default architecture if unset
if not self._arch:
if self._mode == "frame_classifier" or self._mode == "detection_refiner":
self._arch = "resnet50"
else:
self._arch = "cascade"
if self._mode == "detection_refiner" and not os.path.exists(
self._chip_directory):
os.mkdir(self._chip_directory)
# Load object detector if enabled
if self._detector_model:
self._detector = ImageObjectDetector.create("netharn")
detector_config = self._detector.get_configuration()
detector_config.set_value("deployed", self._detector_model)
if not self._detector.set_configuration(detector_config):
print("Unable to configure detector")
return False
# Load scale based on type file if enabled
self._target_type_scales = dict()
if self._scale_type_file:
fin = open(self._scale_type_file, 'r')
for line in fin.readlines():
line = line.rstrip()
parsed_line = line.split()
if len(parsed_line) < 1:
continue
target_area = float(parsed_line[-1])
type_str = str(' '.join(parsed_line[:-1]))
self._target_type_scales[type_str] = target_area
# Initialize persistent variables
self._training_data = []
self._validation_data = []
self._sample_count = 0
return True