def add_label_loss(
model: tf.keras.Model,
grid_fixed: tf.Tensor,
fixed_label: (tf.Tensor, None),
pred_fixed_label: (tf.Tensor, None),
loss_config: dict,
) -> tf.keras.Model:
"""
Add label dissimilarity loss of ddf into model.
:param model: tf.keras.Model
:param grid_fixed: tensor of shape (f_dim1, f_dim2, f_dim3, 3)
:param fixed_label: tensor of shape (batch, f_dim1, f_dim2, f_dim3)
:param pred_fixed_label: tensor of shape (batch, f_dim1, f_dim2, f_dim3)
:param loss_config: config for loss
"""
if fixed_label is not None:
loss_label = tf.reduce_mean(
label_loss.get_dissimilarity_fn(
config=loss_config["dissimilarity"]["label"])(
y_true=fixed_label, y_pred=pred_fixed_label))
weighted_loss_label = (loss_label *
loss_config["dissimilarity"]["label"]["weight"])
model.add_loss(weighted_loss_label)
model.add_metric(loss_label,
name="loss/label_dissimilarity",
aggregation="mean")
model.add_metric(
weighted_loss_label,
name="loss/weighted_label_dissimilarity",
aggregation="mean",
)
# metrics
dice_binary = label_loss.dice_score(y_true=fixed_label,
y_pred=pred_fixed_label,
binary=True)
dice_float = label_loss.dice_score(y_true=fixed_label,
y_pred=pred_fixed_label,
binary=False)
tre = label_loss.compute_centroid_distance(y_true=fixed_label,
y_pred=pred_fixed_label,
grid=grid_fixed)
foreground_label = label_loss.foreground_proportion(y=fixed_label)
foreground_pred = label_loss.foreground_proportion(y=pred_fixed_label)
model.add_metric(dice_binary,
name="metric/dice_binary",
aggregation="mean")
model.add_metric(dice_float,
name="metric/dice_float",
aggregation="mean")
model.add_metric(tre, name="metric/tre", aggregation="mean")
model.add_metric(foreground_label,
name="metric/foreground_label",
aggregation="mean")
model.add_metric(foreground_pred,
name="metric/foreground_pred",
aggregation="mean")
return model
def test_foreground_prop_binary():
"""
Test foreground function with a
tensor of zeros with some ones, asserting
equal to known precomputed tensor.
Testing with binary case.
"""
array_eye = np.identity((3))
tensor_eye = np.zeros((3, 3, 3, 3))
tensor_eye[:, :, 0:3, 0:3] = array_eye
expect = [1.0 / 3, 1.0 / 3, 1.0 / 3]
get = label.foreground_proportion(tensor_eye)
assert assertTensorsEqual(get, expect)
def test_foreground_prop_binary():
"""
Test foreground function with a
tensor of zeros with some ones, asserting
equal to known precomputed tensor.
Testing with binary case.
"""
array_eye = np.identity(3, dtype=np.float32)
tensor_eye = np.zeros((3, 3, 3, 3), dtype=np.float32)
tensor_eye[:, :, 0:3, 0:3] = array_eye
expect = tf.convert_to_tensor([1.0 / 3, 1.0 / 3, 1.0 / 3],
dtype=tf.float32)
get = label.foreground_proportion(tensor_eye)
assert is_equal_tf(get, expect)
def test_foreground_prop_simple():
"""
Test foreground functions with a tensor
of zeros with some ones and some values below
one to assert the thresholding works.
"""
array_eye = np.identity((3))
tensor_eye = np.zeros((3, 3, 3, 3))
tensor_eye[:, 0, :, :] = 0.4 * array_eye # 0
tensor_eye[:, 1, :, :] = array_eye
tensor_eye[:, 2, :, :] = array_eye
expect = [54 / (27 * 9), 54 / (27 * 9), 54 / (27 * 9)]
get = label.foreground_proportion(tensor_eye)
assert assertTensorsEqual(get, expect)
def test_foreground_prop_simple():
"""
Test foreground functions with a tensor
of zeros with some ones and some values below
one to assert the thresholding works.
"""
array_eye = np.identity(3, dtype=np.float32)
tensor_eye = np.zeros((3, 3, 3, 3), dtype=np.float32)
tensor_eye[:, 0, :, :] = 0.4 * array_eye # 0
tensor_eye[:, 1, :, :] = array_eye
tensor_eye[:, 2, :, :] = array_eye
tensor_eye = tf.convert_to_tensor(tensor_eye, dtype=tf.float32)
expect = [54 / (27 * 9), 54 / (27 * 9), 54 / (27 * 9)]
get = label.foreground_proportion(tensor_eye)
assert is_equal_tf(get, expect)