def composite_layer(inputs, mask, hparams, for_output=False):
"""Composite layer."""
x = inputs
# Applies ravanbakhsh on top of each other.
if hparams.composite_layer_type == "ravanbakhsh":
for layer in xrange(hparams.layers_per_layer):
with tf.variable_scope(".%d" % layer):
x = common_layers.ravanbakhsh_set_layer(
hparams.hidden_size,
x,
mask=mask,
sequential=for_output,
dropout=hparams.relu_dropout)
# Transforms elements to get a context, and then uses this in a final layer.
elif hparams.composite_layer_type == "reembedding":
# Transform elements n times and then pool.
for layer in xrange(hparams.layers_per_layer):
with tf.variable_scope("sub_layer_%d" % layer):
x = common_layers.linear_set_layer(
hparams.hidden_size, x, dropout=hparams.relu_dropout)
if for_output:
context = common_layers.running_global_pool_1d(x)
else:
context = common_layers.global_pool_1d(x, mask=mask)
# Final layer.
x = common_layers.linear_set_layer(hparams.hidden_size,
x,
context=context,
dropout=hparams.relu_dropout)
return x
def testGlobalPool1d(self):
x1 = np.random.rand(5, 4, 11)
no_mask = np.ones((5, 4))
full_mask = np.zeros((5, 4))
x1_ = tf.Variable(x1, dtype=tf.float32)
no_mask_ = tf.Variable(no_mask, dtype=tf.float32)
full_mask_ = tf.Variable(full_mask, dtype=tf.float32)
none_mask_max = common_layers.global_pool_1d(x1_)
no_mask_max = common_layers.global_pool_1d(x1_, mask=no_mask_)
result1 = tf.reduce_sum(none_mask_max - no_mask_max)
full_mask_max = common_layers.global_pool_1d(x1_, mask=full_mask_)
result2 = tf.reduce_sum(full_mask_max)
none_mask_avr = common_layers.global_pool_1d(x1_, "AVR")
no_mask_avr = common_layers.global_pool_1d(x1_, "AVR", no_mask_)
result3 = tf.reduce_sum(none_mask_avr - no_mask_avr)
full_mask_avr = common_layers.global_pool_1d(x1_, "AVR", full_mask_)
result4 = tf.reduce_sum(full_mask_avr)
self.evaluate(tf.global_variables_initializer())
actual = self.evaluate([result1, result2, result3, result4])
self.assertAllEqual(actual[:3], [0.0, 0.0, 0.0])
def testGlobalPool1d(self):
x1 = np.random.rand(5, 4, 11)
no_mask = np.ones((5, 4))
full_mask = np.zeros((5, 4))
with self.test_session() as session:
x1_ = tf.Variable(x1, dtype=tf.float32)
no_mask_ = tf.Variable(no_mask, dtype=tf.float32)
full_mask_ = tf.Variable(full_mask, dtype=tf.float32)
none_mask_max = common_layers.global_pool_1d(x1_)
no_mask_max = common_layers.global_pool_1d(x1_, mask=no_mask_)
result1 = tf.reduce_sum(none_mask_max - no_mask_max)
full_mask_max = common_layers.global_pool_1d(x1_, mask=full_mask_)
result2 = tf.reduce_sum(full_mask_max)
none_mask_avr = common_layers.global_pool_1d(x1_, "AVR")
no_mask_avr = common_layers.global_pool_1d(x1_, "AVR", no_mask_)
result3 = tf.reduce_sum(none_mask_avr - no_mask_avr)
full_mask_avr = common_layers.global_pool_1d(
x1_, "AVR", full_mask_)
result4 = tf.reduce_sum(full_mask_avr)
session.run(tf.global_variables_initializer())
actual = session.run([result1, result2, result3, result4])
self.assertAllEqual(actual[:3], [0.0, 0.0, 0.0])
