def over_composite(rgbas):
"""Combines a list of rgba images using the over operation.
Combines RGBA images from back to front (where back is index 0 in list)
with the over operation.
Args:
rgbas: A list of rgba images, these are combined from *back to front*.
Returns:
Returns an RGB image.
"""
alphas = rgbas[:, :, :, :, -1:]
colors = rgbas[:, :, :, :, :-1]
transmittance = tf.cumprod(
1.0 - alphas + 1.0e-8, axis=3, exclusive=True, reverse=True) * alphas
output = tf.reduce_sum(transmittance * colors, axis=3)
return output
def prefix_accuracy(predictions,
labels,
weights_fn=common_layers.weights_nonzero):
"""Average # of correct tokens at start of sequences, ignoring padding 0s.
See section 4.3 of Learning to Transduce with Unbounded Memory,
Grefenstette et al., 2015.
Args:
predictions: Tensor of shape [`batch_size`, `length`, 1, `num_classes`] and
type tf.float32 representing the logits, 0-padded.
labels: Tensor of shape [`batch_size`, `length`, 1, 1] and type tf.int32
representing the labels of same length as logits and 0-padded.
weights_fn: ignored. The weights returned are the total length of the ground
truth labels, excluding 0-paddings.
Returns:
(prefix accuracy, 1.0)
Raises:
ValueError: if weights_fn is not common_layers.weights_nonzero.
"""
if weights_fn is not common_layers.weights_nonzero:
raise ValueError("Only weights_nonzero can be used for this metric.")
predictions = tf.to_int32(
tf.squeeze(tf.argmax(predictions, axis=-1), axis=2))
labels = tf.squeeze(labels, axis=(2, 3))
seq_len = tf.reduce_sum(tf.cast(tf.not_equal(labels, tf.constant(0)),
dtype=tf.float32),
axis=1)
matching_elements = tf.equal(labels, predictions)
prefix_len = tf.reduce_sum(tf.cumprod(tf.cast(matching_elements,
tf.float32),
axis=1),
axis=1)
return tf.reduce_mean(prefix_len / seq_len), tf.constant(1.0)
def discounted_reduce_sum(X, discount, axis=-1):
if discount != 1.0:
disc = tf.cumprod(discount * tf.ones_like(X), axis=axis)
else:
disc = 1.0
return tf.reduce_sum(X * disc, axis=axis)