def show_annotations():
generator = PascalVocGenerator(data_dir='datasets/voc_trainval/VOC0712',
set_name='val')
for image_group, annotation_group, targets in generator:
locations = targets[0]
batch_regr_targets = targets[1]
batch_cls_targets = targets[2]
batch_centerness_targets = targets[3]
for image, annotation, regr_targets, cls_targets, centerness_targets in zip(
image_group, annotation_group, batch_regr_targets,
batch_cls_targets, batch_centerness_targets):
gt_boxes = annotation['bboxes']
for gt_box in gt_boxes:
gt_xmin, gt_ymin, gt_xmax, gt_ymax = gt_box
cv2.rectangle(image, (int(gt_xmin), int(gt_ymin)),
(int(gt_xmax), int(gt_ymax)), (0, 255, 0), 2)
pos_indices = np.where(centerness_targets[:, 1] == 1)[0]
for pos_index in pos_indices:
cx, cy = locations[pos_index]
l, t, r, b, *_ = regr_targets[pos_index]
xmin = cx - l
ymin = cy - t
xmax = cx + r
ymax = cy + b
class_id = np.argmax(cls_targets[pos_index])
centerness = centerness_targets[pos_index][0]
# cv2.putText(image, '{:.2f}'.format(centerness), (cx, cy), cv2.FONT_HERSHEY_SIMPLEX, 2.0, (255, 0, 255), 2)
cv2.putText(image, str(class_id), (xmin, ymin),
cv2.FONT_HERSHEY_SIMPLEX, 2.0, (0, 0, 0), 3)
cv2.circle(image, (round(cx), round(cy)), 5, (255, 0, 0), -1)
cv2.rectangle(image, (round(xmin), round(ymin)),
(round(xmax), round(ymax)), (0, 0, 255), 2)
cv2.namedWindow('image', cv2.WINDOW_NORMAL)
cv2.imshow('image', image)
cv2.waitKey(0)
def verify_no_negative_regr():
transform_generator = random_transform_generator(
min_rotation=-0.1,
max_rotation=0.1,
min_translation=(-0.1, -0.1),
max_translation=(0.1, 0.1),
min_shear=-0.1,
max_shear=0.1,
min_scaling=(0.9, 0.9),
max_scaling=(1.1, 1.1),
flip_x_chance=0.5,
flip_y_chance=0.5,
)
visual_effect_generator = random_visual_effect_generator(
contrast_range=(0.9, 1.1),
brightness_range=(-.1, .1),
hue_range=(-0.05, 0.05),
saturation_range=(0.95, 1.05)
)
common_args = {
'batch_size': 1,
'image_min_side': 800,
'image_max_side': 1333,
'preprocess_image': preprocess_image,
}
generator = PascalVocGenerator(
'datasets/voc_trainval/VOC0712',
'trainval',
transform_generator=transform_generator,
visual_effect_generator=visual_effect_generator,
skip_difficult=True,
**common_args
)
i = 0
for image_group, targets in generator:
i += 1
if i > 20000:
break
def create_generators(args, preprocess_image):
"""
Create generators for training and validation.
Args
args: parseargs object containing configuration for generators.
preprocess_image: Function that preprocesses an image for the network.
"""
common_args = {
'batch_size': args.batch_size,
'config': args.config,
'image_min_side': args.image_min_side,
'image_max_side': args.image_max_side,
'preprocess_image': preprocess_image,
}
# create random transform generator for augmenting training data
if args.random_transform:
transform_generator = random_transform_generator(
min_rotation=-0.1,
max_rotation=0.1,
min_translation=(-0.1, -0.1),
max_translation=(0.1, 0.1),
min_shear=-0.1,
max_shear=0.1,
min_scaling=(0.9, 0.9),
max_scaling=(1.1, 1.1),
flip_x_chance=0.5,
flip_y_chance=0.5,
)
visual_effect_generator = random_visual_effect_generator(
contrast_range=(0.9, 1.1),
brightness_range=(-.1, .1),
hue_range=(-0.05, 0.05),
saturation_range=(0.95, 1.05)
)
else:
transform_generator = random_transform_generator(flip_x_chance=0.5)
visual_effect_generator = None
if args.dataset_type == 'pascal':
train_generator = PascalVocGenerator(
args.pascal_path,
'trainval',
transform_generator=transform_generator,
visual_effect_generator=visual_effect_generator,
skip_difficult=True,
**common_args
)
validation_generator = PascalVocGenerator(
args.pascal_path,
'val',
shuffle_groups=False,
skip_difficult=True,
**common_args
)
elif args.dataset_type == 'csv':
train_generator = CSVGenerator(
args.annotations_path,
args.classes_path,
transform_generator=transform_generator,
visual_effect_generator=visual_effect_generator,
**common_args
)
if args.val_annotations_path:
validation_generator = CSVGenerator(
args.val_annotations_path,
args.classes_path,
shuffle_groups=False,
**common_args
)
else:
validation_generator = None
elif args.dataset_type == 'coco':
# import here to prevent unnecessary dependency on cocoapi
from generators.coco_generator import CocoGenerator
train_generator = CocoGenerator(
args.coco_path,
'train2017',
transform_generator=transform_generator,
visual_effect_generator=visual_effect_generator,
**common_args
)
validation_generator = CocoGenerator(
args.coco_path,
'val2017',
shuffle_groups=False,
**common_args
)
else:
raise ValueError('Invalid data type received: {}'.format(args.dataset_type))
return train_generator, validation_generator
'image_min_side': 800,
'image_max_side': 1333,
'preprocess_image': preprocess_image,
}
# generator = PascalVocGenerator(
# 'datasets/voc_trainval/VOC0712',
# 'val',
# shuffle_groups=False,
# skip_truncated=False,
# skip_difficult=True,
# **common_args
# )
generator = PascalVocGenerator(
'datasets/voc_test/VOC2007',
'test',
shuffle_groups=False,
skip_truncated=False,
skip_difficult=True,
**common_args
)
model_path = '/home/adam/workspace/github/xuannianz/carrot/fsaf/snapshots/2019-10-05/resnet101_pascal_47_0.7652.h5'
# load retinanet model
# import keras.backend as K
# K.set_learning_phase(1)
from models.resnet import resnet_fsaf
from models.retinanet import fsaf_bbox
fsaf = resnet_fsaf(num_classes=20, backbone='resnet101')
model = fsaf_bbox(fsaf)
model.load_weights(model_path, by_name=True)
average_precisions = evaluate(generator, model, visualize=False)
# compute per class average precision
total_instances = []
'batch_size': 1,
'image_min_side': 800,
'image_max_side': 1333,
'preprocess_image': preprocess_image,
}
# generator = PascalVocGenerator(
# 'datasets/voc_trainval/VOC0712',
# 'val',
# shuffle_groups=False,
# skip_truncated=False,
# skip_difficult=True,
# **common_args
# )
generator = PascalVocGenerator('datasets/VOC2007',
'test',
shuffle_groups=False,
skip_truncated=False,
skip_difficult=True,
**common_args)
model_path = 'snapshots/2019-08-25/resnet101_pascal_07_0.7352.h5'
# load retinanet model
import keras.backend as K
K.clear_session()
K.set_learning_phase(1)
model = models.load_model(model_path, backbone_name='resnet101')
# if the model is not converted to an inference model, use the line below
# see: https://github.com/fizyr/keras-retinanet#converting-a-training-model-to-inference-model
model = models.convert_model(model)
average_precisions = evaluate(generator, model, epoch=0)
# compute per class average precision
total_instances = []
precisions = []