def bytenet_internal(inputs, targets, hparams, train):
"""ByteNet, main step used for training."""
with tf.variable_scope("bytenet"):
# Flatten inputs and extend length by 50%.
inputs = tf.expand_dims(common_layers.flatten4d3d(inputs), axis=2)
extend_length = tf.to_int32(0.5 * tf.to_float(tf.shape(inputs)[1]))
inputs_shape = inputs.shape.as_list()
inputs = tf.pad(inputs, [[0, 0], [0, extend_length], [0, 0], [0, 0]])
inputs_shape[1] = None
inputs.set_shape(
inputs_shape) # Don't lose the other shapes when padding.
# Pad inputs and targets to be the same length, divisible by 50.
inputs, targets = common_layers.pad_to_same_length(
inputs, targets, final_length_divisible_by=50)
final_encoder = residual_dilated_conv(inputs, hparams.num_block_repeat,
"SAME", "encoder", hparams,
train)
shifted_targets = common_layers.shift_left(targets)
kernel = (hparams.kernel_height, hparams.kernel_width)
decoder_start = common_layers.conv_block(
tf.concat([final_encoder, shifted_targets], axis=3),
hparams.hidden_size, [((1, 1), kernel)],
padding="LEFT")
return residual_dilated_conv(decoder_start, hparams.num_block_repeat,
"LEFT", "decoder", hparams, train)
def testPadToSameLength(self):
x1 = np.random.rand(5, 7, 11)
x2 = np.random.rand(5, 9, 11)
with self.test_session() as session:
a, b = common_layers.pad_to_same_length(
tf.constant(x1, dtype=tf.float32), tf.constant(x2, dtype=tf.float32))
c, d = common_layers.pad_to_same_length(
tf.constant(x1, dtype=tf.float32),
tf.constant(x2, dtype=tf.float32),
final_length_divisible_by=4)
res1, res2 = session.run([a, b])
res1a, res2a = session.run([c, d])
self.assertEqual(res1.shape, (5, 9, 11))
self.assertEqual(res2.shape, (5, 9, 11))
self.assertEqual(res1a.shape, (5, 12, 11))
self.assertEqual(res2a.shape, (5, 12, 11))
def bytenet_internal(inputs, targets, hparams, train):
"""ByteNet, main step used for training."""
with tf.variable_scope("bytenet"):
# Flatten inputs and extend length by 50%.
inputs = tf.expand_dims(common_layers.flatten4d3d(inputs), axis=2)
extend_length = tf.to_int32(0.5 * tf.to_float(tf.shape(inputs)[1]))
inputs_shape = inputs.shape.as_list()
inputs = tf.pad(inputs, [[0, 0], [0, extend_length], [0, 0], [0, 0]])
inputs_shape[1] = None
inputs.set_shape(inputs_shape) # Don't lose the other shapes when padding.
# Pad inputs and targets to be the same length, divisible by 50.
inputs, targets = common_layers.pad_to_same_length(
inputs, targets, final_length_divisible_by=50)
final_encoder = residual_dilated_conv(
inputs, hparams.num_block_repeat, "SAME", "encoder", hparams, train)
shifted_targets = common_layers.shift_left(targets)
kernel = (hparams.kernel_height, hparams.kernel_width)
decoder_start = common_layers.conv_block(
tf.concat([final_encoder, shifted_targets], axis=3),
hparams.hidden_size, [((1, 1), kernel)],
padding="LEFT")
return residual_dilated_conv(
decoder_start, hparams.num_block_repeat,
"LEFT", "decoder", hparams, train)
def testPadToSameLength(self):
x1 = np.random.rand(5, 7, 11)
x2 = np.random.rand(5, 9, 11)
with self.test_session() as session:
a, b = common_layers.pad_to_same_length(
tf.constant(x1, dtype=tf.float32),
tf.constant(x2, dtype=tf.float32))
c, d = common_layers.pad_to_same_length(
tf.constant(x1, dtype=tf.float32),
tf.constant(x2, dtype=tf.float32),
final_length_divisible_by=4)
res1, res2 = session.run([a, b])
res1a, res2a = session.run([c, d])
self.assertEqual(res1.shape, (5, 9, 11))
self.assertEqual(res2.shape, (5, 9, 11))
self.assertEqual(res1a.shape, (5, 12, 11))
self.assertEqual(res2a.shape, (5, 12, 11))
def similarity_cost(inputs_encoded, targets_encoded):
"""Loss telling to be more similar to your own targets than to others."""
# This is a first very simple version: handle variable-length by padding
# to same length and putting everything into batch. In need of a better way.
x, y = common_layers.pad_to_same_length(inputs_encoded, targets_encoded)
depth = tf.shape(inputs_encoded)[3]
x, y = tf.reshape(x, [-1, depth]), tf.reshape(y, [-1, depth])
return rank_loss(x, y)
def similarity_cost(inputs_encoded, targets_encoded): """Loss telling to be more similar to your own targets than to others.""" # This is a first very simple version: handle variable-length by padding # to same length and putting everything into batch. In need of a better way. x, y = common_layers.pad_to_same_length(inputs_encoded, targets_encoded) depth = tf.shape(inputs_encoded)[3] x, y = tf.reshape(x, [-1, depth]), tf.reshape(y, [-1, depth]) return rank_loss(x, y)