def compute_regression_loss(output, target, labels):
condition = keras.backend.not_equal(labels, -1)
indices = tensorflow_backend.where(condition)
output = tensorflow_backend.gather_nd(output, indices)
target = tensorflow_backend.gather_nd(target, indices)
labels = tensorflow_backend.gather_nd(labels, indices)
condition = keras.backend.greater(labels, 0)
x = keras.backend.zeros_like(labels) + 1
y = keras.backend.zeros_like(labels)
p_star_i = tensorflow_backend.where(condition, x, y)
p_star_i = keras.backend.expand_dims(p_star_i, 0)
output = keras.backend.expand_dims(output, 0)
target = keras.backend.expand_dims(target, 0)
a_y = tensorflow_backend.smooth_l1(output, target, anchored=True)
a = p_star_i * a_y
a = keras.backend.sum(a)
# Divided by anchor overlaps
b = keras.backend.sum(p_star_i + keras.backend.epsilon())
loss = 1.0 * (a / b)
return loss
def compute_loss(output, target):
condition = keras.backend.not_equal(target, -1)
indices = tensorflow_backend.where(condition)
output = tensorflow_backend.gather_nd(output, indices)
target = tensorflow_backend.gather_nd(target, indices)
loss = keras.backend.binary_crossentropy(target, output)
loss = keras.backend.mean(loss)
return loss
def unmap(data, count, inds_inside, fill=0):
""" Unmap a subset of item (data) back to the original set of items (of
size count) """
if keras.backend.ndim(data) == 1:
ret = tensorflow.ones((count, ), dtype=keras.backend.floatx()) * fill
inds_nd = keras.backend.expand_dims(inds_inside)
else:
ret = (count, keras.backend.shape(data)[1])
ret = tensorflow.ones(ret, dtype=keras.backend.floatx()) * fill
data = keras.backend.transpose(data)
data = keras.backend.reshape(data, (-1, ))
inds_ii = keras.backend.tile(inds_inside, [4])
inds_ii = keras.backend.expand_dims(inds_ii)
ones = keras.backend.expand_dims(keras.backend.ones_like(inds_inside),
1)
inds_coords = keras.backend.concatenate(
[ones * 0, ones, ones * 2, ones * 3], 0)
inds_nd = keras.backend.concatenate([inds_ii, inds_coords], 1)
inverse_ret = tensorflow_backend.gather_nd(-1 * ret, inds_nd)
inverse_ret = tensorflow_backend.squeeze(inverse_ret)
updates = inverse_ret + data
ret = tensorflow_backend.scatter_add_tensor(ret, inds_nd, updates)
return ret
def overlapping(anchors, gt_boxes, inds_inside):
"""
overlaps between the anchors and the gt boxes
:param anchors: Generated anchors
:param gt_boxes: Ground truth bounding boxes
:param inds_inside:
:return:
"""
assert keras.backend.ndim(anchors) == 2
assert keras.backend.ndim(gt_boxes) == 2
reference = common.overlap(anchors, gt_boxes)
gt_argmax_overlaps_inds = keras.backend.argmax(reference, axis=0)
argmax_overlaps_inds = keras.backend.argmax(reference, axis=1)
arranged = keras.backend.arange(0, keras.backend.shape(inds_inside)[0])
indices = keras.backend.stack(
[arranged, keras.backend.cast(argmax_overlaps_inds, "int32")], axis=0)
indices = keras.backend.transpose(indices)
max_overlaps = tensorflow_backend.gather_nd(reference, indices)
return argmax_overlaps_inds, max_overlaps, gt_argmax_overlaps_inds
def detections(num_output):
proposals, deltas, scores, metadata = x[0], x[1], x[2], x[3]
proposals = keras.backend.reshape(proposals, (-1, 4))
# unscale back to raw image space
boxes = proposals / metadata[0][2]
num_objects = keras.backend.shape(proposals)[0]
deltas = keras.backend.reshape(deltas, (num_objects, -1))
# Apply bounding-box regression deltas
pred_boxes = common.bbox_transform_inv(boxes, deltas)
pred_boxes = common.clip(pred_boxes, metadata[0][:2])
scores = keras.backend.reshape(scores, (num_objects, -1))
# Arg max
inds = keras.backend.expand_dims(
keras.backend.arange(0, num_objects, dtype='int64'))
top_classes = keras.backend.expand_dims(
keras.backend.argmax(scores, axis=1))
coordinate_0 = keras.backend.concatenate([inds, top_classes * 4],
1)
coordinate_1 = keras.backend.concatenate(
[inds, top_classes * 4 + 1], 1)
coordinate_2 = keras.backend.concatenate(
[inds, top_classes * 4 + 2], 1)
coordinate_3 = keras.backend.concatenate(
[inds, top_classes * 4 + 3], 1)
pred_boxes = tensorflow_backend.gather_nd(
pred_boxes,
keras.backend.reshape(
keras.backend.concatenate([
coordinate_0, coordinate_1, coordinate_2, coordinate_3
], 1), (-1, 2)))
pred_boxes = keras.backend.reshape(pred_boxes, (-1, 4))
max_scores = keras.backend.max(scores, axis=1)
nms_indices = tensorflow_backend.non_maximum_suppression(
boxes=pred_boxes,
scores=max_scores,
maximum=num_objects,
threshold=0.7)
pred_boxes = keras.backend.gather(pred_boxes, nms_indices)
scores = keras.backend.gather(scores, nms_indices)
detections = [
keras.backend.expand_dims(pred_boxes, 0),
keras.backend.expand_dims(scores, 0)
]
return detections[num_output]
def set_label_background(self, labels):
# Clamp labels for the background RoIs to 0
update_indices = keras.backend.arange(self.fg_rois_per_this_image,
keras.backend.shape(labels)[0])
update_indices = keras.backend.reshape(update_indices, (-1, 1))
# By making the label = background
inverse_labels = tensorflow_backend.gather_nd(labels,
update_indices) * -1
labels = tensorflow_backend.scatter_add_tensor(labels, update_indices,
inverse_labels)
return labels