def finite_mean_metric(value, name):
"""Compute Mean Metric"""
# Variables
acc = base.get_metric_variable(name=f"{name}_acc", shape=(), dtype=tf.float32)
num = base.get_metric_variable(name=f"{name}_num", shape=(), dtype=tf.int64)
# New Variables Values
is_finite = tf.math.is_finite(value)
new_acc = tf.cond(is_finite, lambda: acc + value, lambda: acc)
new_num = tf.cond(is_finite, lambda: num + 1, lambda: num)
# Return value and update op
update_op = tf.group(tf.assign(acc, new_acc), tf.assign(num, new_num))
val = tf.div_no_nan(acc, tf.to_float(num))
return (val, update_op)
def __call__(self, tensors: Dict[str, tf.Tensor]) -> Dict[str, Tuple]:
# pylint: disable=unused-argument
value = base.get_metric_variable(name=self.name,
shape=(),
dtype=tf.int64)
update_op = tf.compat.v1.assign(value, value + 1)
return {self.name: (value, update_op)}
def decay_mean_metric(value, decay: float, name: str):
last = base.get_metric_variable(name=f"{name}_decayed_mean",
shape=(),
dtype=value.dtype)
new_value = tf.cond(tf.equal(last, 0), lambda: value,
lambda: decay * last + (1.0 - decay) * value)
update_op = tf.assign(last, new_value)
return (last, update_op)
def max_value_metric(value, name):
max_value = base.get_metric_variable(name=f"{name}_max",
shape=(),
dtype=value.dtype)
update_op = tf.assign(max_value, tf.maximum(value, max_value))
return (max_value, update_op)
def last_value_metric(value, name):
last_value = base.get_metric_variable(name=f"{name}_last",
shape=(),
dtype=value.dtype)
update_op = tf.assign(last_value, value)
return (last_value, update_op)