def test_iou_zero_target(self):
self.assertAllClose(iou_utils.iou_loss(self.pb, self.zeros, 'iou'),
[0.0, 0.0])
self.assertAllClose(iou_utils.iou_loss(self.pb, self.zeros, 'ciou'),
[0.0, 0.0])
self.assertAllClose(iou_utils.iou_loss(self.pb, self.zeros, 'diou'),
[0.0, 0.0])
self.assertAllClose(iou_utils.iou_loss(self.pb, self.zeros, 'giou'),
[0.0, 0.0])
def call(self, y_true, box_outputs):
num_positives, box_targets = y_true
normalizer = num_positives * 4.0
box_iou_loss = iou_utils.iou_loss(box_outputs, box_targets,
self.iou_loss_type)
box_iou_loss = tf.reduce_sum(box_iou_loss) / normalizer
return box_iou_loss
def test_ciou_grad(self):
pb = tf.concat([self.pb, self.zeros], axis=-1)
tb = tf.concat([self.tb, self.zeros], axis=-1)
with tf.GradientTape() as tape:
tape.watch([pb, tb])
loss = iou_utils.iou_loss(pb, tb, 'ciou')
grad = tape.gradient(loss, [tb, pb])
self.assertAlmostEqual(tf.reduce_sum(grad[0]).numpy(), 0.16687772)
self.assertAlmostEqual(tf.reduce_sum(grad[1]).numpy(), -0.16687769)
def call(self, y_true, box_outputs):
anchor_boxes = tf.tile(
self.input_anchors.boxes,
[box_outputs.shape[0] // self.input_anchors.boxes.shape[0], 1])
num_positives, box_targets = y_true
box_outputs = anchors.decode_box_outputs(box_outputs, anchor_boxes)
box_targets = anchors.decode_box_outputs(box_targets, anchor_boxes)
normalizer = num_positives * 4.0
box_iou_loss = iou_utils.iou_loss(box_outputs, box_targets,
self.iou_loss_type)
box_iou_loss = tf.reduce_sum(box_iou_loss) / normalizer
return box_iou_loss
def call(self, y_true, box_outputs):
input_anchors = BoxList(
tf.tile(self.input_anchors.boxes,
[box_outputs.shape[0] // self.input_anchors.boxes.shape[0], 1]))
num_positives, box_targets = y_true
box_outputs = self.box_coder.decode(box_outputs, input_anchors)
box_targets = self.box_coder.decode(box_targets, input_anchors)
normalizer = num_positives * 4.0
box_iou_loss = iou_utils.iou_loss(box_outputs.data['boxes'],
box_targets.data['boxes'],
self.iou_loss_type)
box_iou_loss = tf.reduce_sum(box_iou_loss) / normalizer
return box_iou_loss
def test_iou_multiple_anchors_mixed(self):
pb = tf.concat([self.pb, self.zeros], axis=-1)
tb = tf.concat([self.tb, self.zeros], axis=-1)
self.assertAllClose(iou_utils.iou_loss(pb, tb, 'iou'), [0.875, 1.0])
def test_iou_multiple_anchors(self):
pb = tf.tile(self.pb, [1, 2])
tb = tf.tile(self.tb, [1, 2])
self.assertAllClose(iou_utils.iou_loss(pb, tb, 'iou'), [1.75, 2.0])
def test_giou(self):
self.assertAllClose(iou_utils.iou_loss(self.pb, self.tb, 'giou'),
[1.075000, 1.933333])
def test_diou(self):
self.assertAllClose(iou_utils.iou_loss(self.pb, self.tb, 'diou'),
[1.406532, 1.531532])
def test_ciou(self):
self.assertAllClose(iou_utils.iou_loss(self.pb, self.tb, 'ciou'),
[1.408893, 1.548753])
def test_iou(self):
self.assertAllClose(iou_utils.iou_loss(self.pb, self.tb, 'iou'),
[0.875, 1.])
def _box_iou_loss(box_outputs, box_targets, num_positives, iou_loss_type):
"""Computes box iou loss."""
normalizer = num_positives * 4.0
box_iou_loss = iou_utils.iou_loss(box_outputs, box_targets, iou_loss_type)
box_iou_loss = tf.reduce_sum(box_iou_loss) / normalizer
return box_iou_loss
def test_diou(self):
self.assertAllClose(iou_utils.iou_loss(self.pb, self.tb, 'diou'),
[0.9969512, 1.6243094])
def test_ciou(self):
self.assertAllClose(iou_utils.iou_loss(self.pb, self.tb, 'ciou'),
[0.99931306, 1.6415315])