def true_loss(y_true, y_pred):
# print(holder)
y_true2, y_pred2 = data_util.get_target(
y_true[:, :-1, :], y_pred[:, 1:, :]) # is this shit right??
y_true2 = tf.concat(
[tf.zeros([args.batch_size, 1, 1], tf.float32), y_true2],
axis=-2)
y_pred2 = tf.concat(
[tf.zeros([args.batch_size, 1, 1], tf.float32), y_pred2],
axis=-2)
n_batch, n_inter, n_outputs = y_true.shape
waviness_norm_1 = tf.reduce_sum(
tf.abs(y_pred[:, 1:, :] - y_pred[:, :-1, :])) / (n_inter *
n_outputs)
waviness_norm_2 = tf.reduce_sum(
tf.square(y_pred[:, 1:, :] - y_pred[:, :-1, :])) / (n_inter *
n_outputs)
y_true, y_pred = data_util.get_target(y_true, y_pred)
def get_loss_value(a, b, c, d, e, f):
return tf.keras.losses.binary_crossentropy(
a, b) + l_r * tf.keras.losses.binary_crossentropy(
c, d) + l_w1 * e + l_w2 * f # (y_true, y_pred)
return get_loss_value(y_true, y_pred, y_true2, y_pred2,
waviness_norm_1, waviness_norm_2)
def custom_loss(y_true, y_pred):
y_true, y_pred = data_util.get_target(y_true, y_pred)
return tf.keras.losses.binary_crossentropy(y_true, y_pred)
def update_state(self, y_true, y_pred, sample_weight=None):
true, pred = data_util.get_target(y_true, y_pred)
super(TrueNegatives, self).update_state(y_true=true,
y_pred=pred,
sample_weight=sample_weight)
def update_state(self, y_true, y_pred, sample_weight=None):
true, pred = data_util.get_target(y_true, y_pred)
super(BinaryAccuracy, self).update_state(y_true=true,
y_pred=pred,
sample_weight=sample_weight)
def update_state(self, y_true, y_pred, sample_weight=None):
true, pred = data_util.get_target(y_true, y_pred)
super(SpecificityAtSensitivity, self).update_state(y_true=true,
y_pred=pred,
sample_weight=sample_weight)
def update_state(self, y_true, y_pred, sample_weight=None):
true, pred = data_util.get_target(y_true, y_pred)
super(Precision, self).update_state(y_true=true,
y_pred=pred,
sample_weight=sample_weight)