def test_anchors_for_shape_odd_input():
pyramid_levels = [3]
image_shape = (
20, 20
) # this shape causes rounding errors when downsampling using convolutions
sizes = [32]
strides = [8]
ratios = np.array([1], keras.backend.floatx())
scales = np.array([1], keras.backend.floatx())
anchor_params = AnchorParameters(sizes, strides, ratios, scales)
anchors = anchors_for_shape(image_shape,
pyramid_levels=pyramid_levels,
anchor_params=anchor_params)
expected_anchors = np.array([
[-14, -14, 18, 18],
[-6, -14, 26, 18],
[2, -14, 34, 18],
[-14, -6, 18, 26],
[-6, -6, 26, 26],
[2, -6, 34, 26],
[-14, 2, 18, 34],
[-6, 2, 26, 34],
[2, 2, 34, 34],
])
np.testing.assert_equal(anchors, expected_anchors)
def __init__(self,
backbone="resnet50",
random_transform=False,
batch_size=2,
step=100,
epoch=40,
anchor_settings=None):
""" Initializer
Args:
backbone
string indicate which backbone network to use (only for ones defined in
keras-retinanet)
random_transform
boolean indicate whether image is randomly transformed when training
batch_size
number of images sent for each step during training
step
number of steps(batches) to run for each epoch during training
epoch
number of epochs for this training
anchor_settings
a list of lists indicate anchor box parameters. please reference
keras-retinanet's document for how to set up anchor box parameters
"""
check_keras_version()
self._backbone_name = backbone
self._backbone = models.backbone(backbone)
self._img_preprocessor = self._backbone.preprocess_image
self._batch_size = batch_size
self._step = step
self._epoch = epoch
if random_transform:
self._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,
)
else:
self._transform_generator = random_transform_generator(
flip_x_chance=0.5)
self._common_args = {
'batch_size': batch_size,
'preprocess_image': self._img_preprocessor
}
if anchor_settings:
self._anchor_params = AnchorParameters(*anchor_settings)
else:
self._anchor_params = AnchorParameters.default
def parse_anchor_parameters(config):
ratios = np.array(
list(map(float, config['anchor_parameters']['ratios'].split(' '))),
keras.backend.floatx())
scales = np.array(
list(map(float, config['anchor_parameters']['scales'].split(' '))),
keras.backend.floatx())
sizes = list(map(int, config['anchor_parameters']['sizes'].split(' ')))
strides = list(map(int, config['anchor_parameters']['strides'].split(' ')))
return AnchorParameters(sizes, strides, ratios, scales)
def calculate_config(values, ratio_count):
split_point = int((ratio_count - 1) / 2)
ratios = [1]
for i in range(split_point):
ratios.append(values[i])
ratios.append(1 / values[i])
scales = values[split_point:]
return AnchorParameters(SIZES, STRIDES, ratios, scales)
def test_anchors_for_shape_dimensions():
sizes = [32, 64, 128]
strides = [8, 16, 32]
ratios = np.array([0.5, 1, 2, 3], tensorflow.keras.backend.floatx())
scales = np.array([1, 1.2, 1.6], tensorflow.keras.backend.floatx())
anchor_params = AnchorParameters(sizes, strides, ratios, scales)
pyramid_levels = [3, 4, 5]
image_shape = (64, 64)
all_anchors = anchors_for_shape(image_shape, pyramid_levels=pyramid_levels, anchor_params=anchor_params)
assert all_anchors.shape == (1008, 4)
def calculate_config(values,
ratio_count,
SIZES=[12, 24, 48, 96, 192],
STRIDES=[8, 16, 32, 64, 128]):
split_point = int((ratio_count - 1) / 2)
ratios = [1]
for i in range(split_point):
ratios.append(values[i])
ratios.append(1 / values[i])
scales = values[split_point:]
return AnchorParameters(SIZES, STRIDES, ratios, scales)
def test_anchors_for_shape_values():
sizes = [12]
strides = [8]
ratios = np.array([1, 2], tf.keras.backend.floatx())
scales = np.array([1, 2], tf.keras.backend.floatx())
anchor_params = AnchorParameters(sizes, strides, ratios, scales)
pyramid_levels = [3]
image_shape = (16, 16)
all_anchors = anchors_for_shape(image_shape,
pyramid_levels=pyramid_levels,
anchor_params=anchor_params)
# using almost_equal for floating point imprecisions
np.testing.assert_almost_equal(all_anchors[0, :], [
strides[0] / 2 - (sizes[0] * scales[0] / np.sqrt(ratios[0])) / 2,
strides[0] / 2 - (sizes[0] * scales[0] * np.sqrt(ratios[0])) / 2,
strides[0] / 2 + (sizes[0] * scales[0] / np.sqrt(ratios[0])) / 2,
strides[0] / 2 + (sizes[0] * scales[0] * np.sqrt(ratios[0])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[1, :], [
strides[0] / 2 - (sizes[0] * scales[1] / np.sqrt(ratios[0])) / 2,
strides[0] / 2 - (sizes[0] * scales[1] * np.sqrt(ratios[0])) / 2,
strides[0] / 2 + (sizes[0] * scales[1] / np.sqrt(ratios[0])) / 2,
strides[0] / 2 + (sizes[0] * scales[1] * np.sqrt(ratios[0])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[2, :], [
strides[0] / 2 - (sizes[0] * scales[0] / np.sqrt(ratios[1])) / 2,
strides[0] / 2 - (sizes[0] * scales[0] * np.sqrt(ratios[1])) / 2,
strides[0] / 2 + (sizes[0] * scales[0] / np.sqrt(ratios[1])) / 2,
strides[0] / 2 + (sizes[0] * scales[0] * np.sqrt(ratios[1])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[3, :], [
strides[0] / 2 - (sizes[0] * scales[1] / np.sqrt(ratios[1])) / 2,
strides[0] / 2 - (sizes[0] * scales[1] * np.sqrt(ratios[1])) / 2,
strides[0] / 2 + (sizes[0] * scales[1] / np.sqrt(ratios[1])) / 2,
strides[0] / 2 + (sizes[0] * scales[1] * np.sqrt(ratios[1])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[4, :], [
strides[0] * 3 / 2 - (sizes[0] * scales[0] / np.sqrt(ratios[0])) / 2,
strides[0] / 2 - (sizes[0] * scales[0] * np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[0] / np.sqrt(ratios[0])) / 2,
strides[0] / 2 + (sizes[0] * scales[0] * np.sqrt(ratios[0])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[5, :], [
strides[0] * 3 / 2 - (sizes[0] * scales[1] / np.sqrt(ratios[0])) / 2,
strides[0] / 2 - (sizes[0] * scales[1] * np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[1] / np.sqrt(ratios[0])) / 2,
strides[0] / 2 + (sizes[0] * scales[1] * np.sqrt(ratios[0])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[6, :], [
strides[0] * 3 / 2 - (sizes[0] * scales[0] / np.sqrt(ratios[1])) / 2,
strides[0] / 2 - (sizes[0] * scales[0] * np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[0] / np.sqrt(ratios[1])) / 2,
strides[0] / 2 + (sizes[0] * scales[0] * np.sqrt(ratios[1])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[7, :], [
strides[0] * 3 / 2 - (sizes[0] * scales[1] / np.sqrt(ratios[1])) / 2,
strides[0] / 2 - (sizes[0] * scales[1] * np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[1] / np.sqrt(ratios[1])) / 2,
strides[0] / 2 + (sizes[0] * scales[1] * np.sqrt(ratios[1])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[8, :], [
strides[0] / 2 - (sizes[0] * scales[0] / np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 - (sizes[0] * scales[0] * np.sqrt(ratios[0])) / 2,
strides[0] / 2 + (sizes[0] * scales[0] / np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[0] * np.sqrt(ratios[0])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[9, :], [
strides[0] / 2 - (sizes[0] * scales[1] / np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 - (sizes[0] * scales[1] * np.sqrt(ratios[0])) / 2,
strides[0] / 2 + (sizes[0] * scales[1] / np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[1] * np.sqrt(ratios[0])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[10, :], [
strides[0] / 2 - (sizes[0] * scales[0] / np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 - (sizes[0] * scales[0] * np.sqrt(ratios[1])) / 2,
strides[0] / 2 + (sizes[0] * scales[0] / np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[0] * np.sqrt(ratios[1])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[11, :], [
strides[0] / 2 - (sizes[0] * scales[1] / np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 - (sizes[0] * scales[1] * np.sqrt(ratios[1])) / 2,
strides[0] / 2 + (sizes[0] * scales[1] / np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[1] * np.sqrt(ratios[1])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[12, :], [
strides[0] * 3 / 2 - (sizes[0] * scales[0] / np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 - (sizes[0] * scales[0] * np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[0] / np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[0] * np.sqrt(ratios[0])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[13, :], [
strides[0] * 3 / 2 - (sizes[0] * scales[1] / np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 - (sizes[0] * scales[1] * np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[1] / np.sqrt(ratios[0])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[1] * np.sqrt(ratios[0])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[14, :], [
strides[0] * 3 / 2 - (sizes[0] * scales[0] / np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 - (sizes[0] * scales[0] * np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[0] / np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[0] * np.sqrt(ratios[1])) / 2,
],
decimal=6)
np.testing.assert_almost_equal(all_anchors[15, :], [
strides[0] * 3 / 2 - (sizes[0] * scales[1] / np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 - (sizes[0] * scales[1] * np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[1] / np.sqrt(ratios[1])) / 2,
strides[0] * 3 / 2 + (sizes[0] * scales[1] * np.sqrt(ratios[1])) / 2,
],
decimal=6)
