def __init__(self):
self.convert_RGB_to_HSV = ConvertColor(current='RGB', to='HSV')
self.convert_HSV_to_RGB = ConvertColor(current='HSV', to='RGB')
self.convert_to_float32 = ConvertDataType(to='float32')
self.convert_to_uint8 = ConvertDataType(to='uint8')
self.convert_to_3_channels = ConvertTo3Channels()
self.random_brightness = RandomBrightness(lower=-32,
upper=32,
prob=0.5)
self.random_contrast = RandomContrast(lower=0.5, upper=1.5, prob=0.5)
self.random_saturation = RandomSaturation(lower=0.5,
upper=1.5,
prob=0.5)
self.random_hue = RandomHue(max_delta=18, prob=0.5)
self.random_channel_swap = RandomChannelSwap(prob=0.0)
self.sequence1 = [
self.convert_to_3_channels, self.convert_to_float32,
self.random_brightness, self.random_contrast,
self.convert_to_uint8, self.convert_RGB_to_HSV,
self.convert_to_float32, self.random_saturation, self.random_hue,
self.convert_to_uint8, self.convert_HSV_to_RGB,
self.random_channel_swap
]
self.sequence2 = [
self.convert_to_3_channels, self.convert_to_float32,
self.random_brightness, self.convert_to_uint8,
self.convert_RGB_to_HSV, self.convert_to_float32,
self.random_saturation, self.random_hue, self.convert_to_uint8,
self.convert_HSV_to_RGB, self.convert_to_float32,
self.random_contrast, self.convert_to_uint8,
self.random_channel_swap
]
def __init__(self,
resize_height,
resize_width,
random_brightness=(-48, 48, 0.5),
random_contrast=(0.5, 1.8, 0.5),
random_saturation=(0.5, 1.8, 0.5),
random_hue=(18, 0.5),
random_flip=0.5,
random_rotate=([90, 180, 270], 0.5),
min_scale=0.3,
max_scale=2.0,
min_aspect_ratio=0.8,
max_aspect_ratio=1.25,
n_trials_max=3,
clip_boxes=True,
overlap_criterion='area',
bounds_box_filter=(0.3, 1.0),
bounds_validator=(0.5, 1.0),
n_boxes_min=1,
background=(0, 0, 0),
labels_format={
'class_id': 0,
'xmin': 1,
'ymin': 2,
'xmax': 3,
'ymax': 4
}):
self.n_trials_max = n_trials_max
self.clip_boxes = clip_boxes
self.overlap_criterion = overlap_criterion
self.bounds_box_filter = bounds_box_filter
self.bounds_validator = bounds_validator
self.n_boxes_min = n_boxes_min
self.background = background
self.labels_format = labels_format
# Determines which boxes are kept in an image after the transformations have been applied.
self.box_filter_patch = BoxFilter(
check_overlap=True,
check_min_area=False,
check_degenerate=False,
overlap_criterion=self.overlap_criterion,
overlap_bounds=self.bounds_box_filter,
labels_format=self.labels_format)
self.box_filter_resize = BoxFilter(check_overlap=False,
check_min_area=True,
check_degenerate=True,
min_area=16,
labels_format=self.labels_format)
# Determines whether the result of the transformations is a valid training image.
self.image_validator = ImageValidator(
overlap_criterion=self.overlap_criterion,
bounds=self.bounds_validator,
n_boxes_min=self.n_boxes_min,
labels_format=self.labels_format)
# Utility transformations
self.convert_to_3_channels = ConvertTo3Channels(
) # Make sure all images end up having 3 channels.
self.convert_RGB_to_HSV = ConvertColor(current='RGB', to='HSV')
self.convert_HSV_to_RGB = ConvertColor(current='HSV', to='RGB')
self.convert_to_float32 = ConvertDataType(to='float32')
self.convert_to_uint8 = ConvertDataType(to='uint8')
self.resize = Resize(height=resize_height,
width=resize_width,
box_filter=self.box_filter_resize,
labels_format=self.labels_format)
# Photometric transformations
self.random_brightness = RandomBrightness(lower=random_brightness[0],
upper=random_brightness[1],
prob=random_brightness[2])
self.random_contrast = RandomContrast(lower=random_contrast[0],
upper=random_contrast[1],
prob=random_contrast[2])
self.random_saturation = RandomSaturation(lower=random_saturation[0],
upper=random_saturation[1],
prob=random_saturation[2])
self.random_hue = RandomHue(max_delta=random_hue[0],
prob=random_hue[1])
# Geometric transformations
self.random_horizontal_flip = RandomFlip(
dim='horizontal',
prob=random_flip,
labels_format=self.labels_format)
self.random_vertical_flip = RandomFlip(
dim='vertical', prob=random_flip, labels_format=self.labels_format)
self.random_rotate = RandomRotate(angles=random_rotate[0],
prob=random_rotate[1],
labels_format=self.labels_format)
self.patch_coord_generator = PatchCoordinateGenerator(
must_match='w_ar',
min_scale=min_scale,
max_scale=max_scale,
scale_uniformly=False,
min_aspect_ratio=min_aspect_ratio,
max_aspect_ratio=max_aspect_ratio)
self.random_patch = RandomPatch(
patch_coord_generator=self.patch_coord_generator,
box_filter=self.box_filter_patch,
image_validator=self.image_validator,
n_trials_max=self.n_trials_max,
clip_boxes=self.clip_boxes,
prob=1.0,
can_fail=False,
labels_format=self.labels_format)
# Define the processing chain.
self.transformations = [
self.convert_to_3_channels, self.convert_to_float32,
self.random_brightness, self.random_contrast,
self.convert_to_uint8, self.convert_RGB_to_HSV,
self.convert_to_float32, self.random_saturation, self.random_hue,
self.convert_to_uint8, self.convert_HSV_to_RGB,
self.random_horizontal_flip, self.random_vertical_flip,
self.random_rotate, self.random_patch, self.resize
]
model.load_weights(weights_path, by_name=True)
# 3: Instantiate an optimizer and the SSD loss function and compile the model.
# If you want to follow the original Caffe implementation, use the preset SGD
# optimizer, otherwise I'd recommend the commented-out Adam optimizer.
#adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
sgd = SGD(lr=0.001, momentum=0.9, decay=0.0, nesterov=False)
ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0)
model.compile(optimizer=sgd, loss=ssd_loss.compute_loss)
val_dataset = DataGenerator(load_images_into_memory=False,
hdf5_dataset_path=val_hdf5_path)
convert_to_3_channels = ConvertTo3Channels()
resize = Resize(height=img_height, width=img_width)
predict_generator = val_dataset.generate(
batch_size=1,
shuffle=True,
transformations=[convert_to_3_channels, resize],
label_encoder=None,
returns={
'processed_images', 'filenames', 'inverse_transform',
'original_images', 'original_labels'
},
keep_images_without_gt=False)
batch_images, batch_filenames, batch_inverse_transforms, batch_original_images, batch_original_labels = next(
predict_generator)
i = 0 # Which batch item to look at
def __init__(self,
random_brightness=(-48, 48, 0.5),
random_contrast=(0.5, 1.8, 0.5),
random_saturation=(0.5, 1.8, 0.5),
random_hue=(18, 0.5),
random_flip=0.5,
random_translate=((0.03, 0.5), (0.03, 0.5), 0.5),
random_scale=(0.5, 2.0, 0.5),
n_trials_max=3,
clip_boxes=True,
overlap_criterion='area',
bounds_box_filter=(0.3, 1.0),
bounds_validator=(0.5, 1.0),
n_boxes_min=1,
background=(0, 0, 0),
labels_format={
'class_id': 0,
'xmin': 1,
'ymin': 2,
'xmax': 3,
'ymax': 4
}):
if (random_scale[0] >= 1) or (random_scale[1] <= 1):
raise ValueError(
"This sequence of transformations only makes sense if the minimum scaling factor is <1 and the maximum scaling factor is >1."
)
self.n_trials_max = n_trials_max
self.clip_boxes = clip_boxes
self.overlap_criterion = overlap_criterion
self.bounds_box_filter = bounds_box_filter
self.bounds_validator = bounds_validator
self.n_boxes_min = n_boxes_min
self.background = background
self.labels_format = labels_format
# Determines which boxes are kept in an image after the transformations have been applied.
self.box_filter = BoxFilter(check_overlap=True,
check_min_area=True,
check_degenerate=True,
overlap_criterion=self.overlap_criterion,
overlap_bounds=self.bounds_box_filter,
min_area=16,
labels_format=self.labels_format)
# Determines whether the result of the transformations is a valid training image.
self.image_validator = ImageValidator(
overlap_criterion=self.overlap_criterion,
bounds=self.bounds_validator,
n_boxes_min=self.n_boxes_min,
labels_format=self.labels_format)
# Utility distortions
self.convert_RGB_to_HSV = ConvertColor(current='RGB', to='HSV')
self.convert_HSV_to_RGB = ConvertColor(current='HSV', to='RGB')
self.convert_to_float32 = ConvertDataType(to='float32')
self.convert_to_uint8 = ConvertDataType(to='uint8')
self.convert_to_3_channels = ConvertTo3Channels(
) # Make sure all images end up having 3 channels.
# Photometric transformations
self.random_brightness = RandomBrightness(lower=random_brightness[0],
upper=random_brightness[1],
prob=random_brightness[2])
self.random_contrast = RandomContrast(lower=random_contrast[0],
upper=random_contrast[1],
prob=random_contrast[2])
self.random_saturation = RandomSaturation(lower=random_saturation[0],
upper=random_saturation[1],
prob=random_saturation[2])
self.random_hue = RandomHue(max_delta=random_hue[0],
prob=random_hue[1])
# Geometric transformations
self.random_flip = RandomFlip(dim='horizontal',
prob=random_flip,
labels_format=self.labels_format)
self.random_translate = RandomTranslate(
dy_minmax=random_translate[0],
dx_minmax=random_translate[1],
prob=random_translate[2],
clip_boxes=self.clip_boxes,
box_filter=self.box_filter,
image_validator=self.image_validator,
n_trials_max=self.n_trials_max,
background=self.background,
labels_format=self.labels_format)
self.random_zoom_in = RandomScale(min_factor=1.0,
max_factor=random_scale[1],
prob=random_scale[2],
clip_boxes=self.clip_boxes,
box_filter=self.box_filter,
image_validator=self.image_validator,
n_trials_max=self.n_trials_max,
background=self.background,
labels_format=self.labels_format)
self.random_zoom_out = RandomScale(
min_factor=random_scale[0],
max_factor=1.0,
prob=random_scale[2],
clip_boxes=self.clip_boxes,
box_filter=self.box_filter,
image_validator=self.image_validator,
n_trials_max=self.n_trials_max,
background=self.background,
labels_format=self.labels_format)
# If we zoom in, do translation before scaling.
self.sequence1 = [
self.convert_to_3_channels, self.convert_to_float32,
self.random_brightness, self.random_contrast,
self.convert_to_uint8, self.convert_RGB_to_HSV,
self.convert_to_float32, self.random_saturation, self.random_hue,
self.convert_to_uint8, self.convert_HSV_to_RGB,
self.random_translate, self.random_zoom_in, self.random_flip
]
# If we zoom out, do scaling before translation.
self.sequence2 = [
self.convert_to_3_channels, self.convert_to_float32,
self.random_brightness, self.convert_to_uint8,
self.convert_RGB_to_HSV, self.convert_to_float32,
self.random_saturation, self.random_hue, self.convert_to_uint8,
self.convert_HSV_to_RGB, self.convert_to_float32,
self.random_contrast, self.convert_to_uint8, self.random_zoom_out,
self.random_translate, self.random_flip
]