def testScaledDotAttention(self):
batch_size = 3
num_heads = 8
values_length = [5, 3, 7]
queries_length = [8, 6, 10]
depth = 20
queries = tf.placeholder_with_default(
np.random.randn(batch_size, num_heads, max(queries_length), depth).astype(np.float32),
shape=(None, num_heads, None, depth))
values = tf.placeholder_with_default(
np.random.randn(batch_size, num_heads, max(values_length), depth).astype(np.float32),
shape=(None, num_heads, None, depth))
keys = values
mask = transformer.build_sequence_mask(values_length, num_heads=num_heads)
context, attn = transformer.dot_product_attention(
queries,
keys,
values,
tf.estimator.ModeKeys.PREDICT,
mask=mask)
with self.test_session() as sess:
context, attn = sess.run([context, attn])
self.assertTupleEqual(
(batch_size, num_heads, max(queries_length), depth), context.shape)
self.assertTupleEqual(
(batch_size, num_heads, max(queries_length), max(values_length)), attn.shape)
for i in range(batch_size):
length = values_length[i]
padding_length = max(values_length) - length
if padding_length > 0:
self.assertEqual(0.0, np.sum(attn[i, :, :, length:max(values_length)]))
def _build_memory_mask(self, memory, memory_sequence_length=None):
if memory_sequence_length is None:
return None
else:
return transformer.build_sequence_mask(
memory_sequence_length,
num_heads=self.num_heads,
maximum_length=tf.shape(memory)[1])
def _build_memory_mask(self, memory, memory_sequence_length=None):
if memory_sequence_length is None:
return None
else:
return transformer.build_sequence_mask(
memory_sequence_length,
num_heads=self.num_heads,
maximum_length=tf.shape(memory)[1],
dtype=memory.dtype)
def cross_attention(self,
table_encodes,
document_encodes,
num_units,
num_heads,
num_layers,
ffn_inner_dim,
sequence_length=None,
mode=tf.estimator.ModeKeys.TRAIN):
table_encodes *= num_units**0.5
# if self.position_encoder is not None:
# inputs = self.position_encoder(inputs)
inputs = tf.layers.dropout(
table_encodes,
rate=FLAGS.attention_dropout,
training=mode == tf.estimator.ModeKeys.TRAIN)
mask = transformer.build_sequence_mask(
sequence_length,
num_heads=num_heads,
maximum_length=tf.shape(document_encodes)[1])
state = ()
for l in range(num_layers):
with tf.variable_scope("layer_{}".format(l)):
with tf.variable_scope("multi_head"):
context = transformer.multi_head_attention(
num_heads,
transformer.norm(inputs),
document_encodes,
mode,
num_units=num_units,
mask=mask,
dropout=FLAGS.attention_dropout)
context = transformer.drop_and_add(inputs,
context,
mode,
dropout=FLAGS.dropout)
with tf.variable_scope("ffn"):
transformed = transformer.feed_forward(
transformer.norm(context),
ffn_inner_dim,
mode,
dropout=FLAGS.attention_dropout)
transformed = transformer.drop_and_add(
context, transformed, mode, dropout=FLAGS.dropout)
inputs = transformed
state += (tf.reduce_mean(inputs, axis=1), )
outputs = transformer.norm(inputs)
# return (outputs, state, sequence_length)
return outputs
def testBuildSequenceMask(self):
num_heads = 4
length = [5, 3, 7]
expected = [[1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0],
[1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0],
[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]]
mask = transformer.build_sequence_mask(tf.constant(length),
num_heads=num_heads)
mask = self.evaluate(mask)
self.assertTupleEqual(mask.shape, (len(length), 1, 1, max(length)))
self.assertAllEqual(np.squeeze(mask), expected)
def encode(self,
inputs,
sequence_length=None,
mode=tf.estimator.ModeKeys.TRAIN):
inputs *= self.num_units**0.5
if self.position_encoder is not None:
inputs = self.position_encoder(inputs)
else:
print("===============================================")
print("no position encoder")
inputs = tf.layers.dropout(
inputs,
rate=self.dropout,
training=mode == tf.estimator.ModeKeys.TRAIN)
mask = transformer.build_sequence_mask(
sequence_length,
num_heads=self.num_heads,
maximum_length=tf.shape(inputs)[1])
state = ()
for l in range(self.num_layers):
with tf.variable_scope("layer_{}".format(l)):
with tf.variable_scope("multi_head"):
context = transformer.multi_head_attention(
self.num_heads,
transformer.norm(inputs),
None,
mode,
num_units=self.num_units,
mask=mask,
dropout=self.attention_dropout)
context = transformer.drop_and_add(inputs,
context,
mode,
dropout=self.dropout)
with tf.variable_scope("ffn"):
transformed = transformer.feed_forward(
transformer.norm(context),
self.ffn_inner_dim,
mode,
dropout=self.relu_dropout)
transformed = transformer.drop_and_add(
context, transformed, mode, dropout=self.dropout)
inputs = transformed
state += (tf.reduce_mean(inputs, axis=1), )
outputs = transformer.norm(inputs)
return (outputs, state, sequence_length)
def encode(self, inputs, sequence_length=None, mode=tf.estimator.ModeKeys.TRAIN):
inputs *= self.num_units**0.5
if self.position_encoder is not None:
inputs = self.position_encoder(inputs, sequence_length=sequence_length)
inputs = tf.layers.dropout(
inputs,
rate=self.dropout,
training=mode == tf.estimator.ModeKeys.TRAIN)
mask = transformer.build_sequence_mask(
sequence_length,
num_heads=self.num_heads,
maximum_length=tf.shape(inputs)[1],
dtype=inputs.dtype)
state = ()
for l in range(self.num_layers):
with tf.variable_scope("layer_{}".format(l)):
with tf.variable_scope("multi_head"):
context = transformer.multi_head_attention(
self.num_heads,
transformer.norm(inputs),
None,
mode,
num_units=self.num_units,
mask=mask,
dropout=self.attention_dropout)
context = transformer.drop_and_add(
inputs,
context,
mode,
dropout=self.dropout)
with tf.variable_scope("ffn"):
transformed = transformer.feed_forward(
transformer.norm(context),
self.ffn_inner_dim,
mode,
dropout=self.relu_dropout)
transformed = transformer.drop_and_add(
context,
transformed,
mode,
dropout=self.dropout)
inputs = transformed
state += (tf.reduce_mean(inputs, axis=1),)
outputs = transformer.norm(inputs)
return (outputs, state, sequence_length)
def testBuildSequenceMask(self):
num_heads = 4
length = [5, 3, 7]
expected = [
[1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0],
[1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0],
[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]]
mask = transformer.build_sequence_mask(tf.constant(length), num_heads=num_heads)
with self.test_session() as sess:
mask = sess.run(mask)
mask = np.reshape(mask, (len(length), num_heads, max(length)))
mask = np.transpose(mask, (1, 0, 2))
for b in range(len(length)):
self.assertAllEqual(expected, mask[b])
def testBuildSequenceMaskWithMaxLen(self):
num_heads = 4
length = [5, 3, 6]
maximum_length = 7
expected = [[1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 0.0],
[1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0],
[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.0]]
mask = transformer.build_sequence_mask(tf.constant(length),
num_heads=num_heads,
maximum_length=maximum_length)
with self.test_session() as sess:
mask = sess.run(mask)
self.assertTupleEqual(mask.shape,
(len(length), 1, 1, maximum_length))
self.assertAllEqual(np.squeeze(mask), expected)
def __init__(self,
mode,
num_layers,
num_units,
num_heads,
memory,
memory_sequence_length,
cell_class=tf.contrib.rnn.LayerNormBasicLSTMCell,
dropout=0.3):
super(_RNMTPlusDecoderCell, self).__init__()
self._mode = mode
self._num_units = num_units
self._num_heads = num_heads
self._dropout = dropout
self._cells = [cell_class(num_units) for _ in range(num_layers)]
self._memory = memory
self._memory_mask = build_sequence_mask(
memory_sequence_length,
num_heads=self._num_heads,
maximum_length=tf.shape(memory)[1])
def _self_attention_stack(self,
inputs,
sequence_length=None,
mode=tf.estimator.ModeKeys.TRAIN,
cache=None,
memory=None,
memory_sequence_length=None,
step=None):
inputs *= self.num_units**0.5
if self.position_encoder is not None:
if step is None:
inputs = self.position_encoder(inputs, sequence_length=sequence_length)
else:
inputs = self.position_encoder.apply_one(inputs, step + 1)
inputs = tf.layers.dropout(
inputs,
rate=self.dropout,
training=mode == tf.estimator.ModeKeys.TRAIN)
decoder_mask = None
memory_mask = None
last_attention = None
if self.self_attention_type == "scaled_dot":
if sequence_length is not None:
decoder_mask = transformer.build_future_mask(
sequence_length,
num_heads=self.num_heads,
maximum_length=tf.shape(inputs)[1])
elif self.self_attention_type == "average":
if cache is None:
if sequence_length is None:
sequence_length = tf.fill([tf.shape(inputs)[0]], tf.shape(inputs)[1])
decoder_mask = transformer.cumulative_average_mask(
sequence_length, maximum_length=tf.shape(inputs)[1], dtype=inputs.dtype)
if memory is not None and memory_sequence_length is not None:
memory_mask = transformer.build_sequence_mask(
memory_sequence_length,
num_heads=self.num_heads,
maximum_length=tf.shape(memory)[1])
for l in range(self.num_layers):
layer_name = "layer_{}".format(l)
layer_cache = cache[layer_name] if cache is not None else None
with tf.variable_scope(layer_name):
if self.self_attention_type == "scaled_dot":
with tf.variable_scope("masked_multi_head"):
encoded = transformer.multi_head_attention(
self.num_heads,
transformer.norm(inputs),
None,
mode,
num_units=self.num_units,
mask=decoder_mask,
cache=layer_cache,
dropout=self.attention_dropout)
encoded = transformer.drop_and_add(
inputs,
encoded,
mode,
dropout=self.dropout)
elif self.self_attention_type == "average":
with tf.variable_scope("average_attention"):
# Cumulative average.
x = transformer.norm(inputs)
y = transformer.cumulative_average(
x, decoder_mask if cache is None else step, cache=layer_cache)
# FFN.
y = transformer.feed_forward(
y, self.ffn_inner_dim, mode, dropout=self.relu_dropout)
# Gating layer.
z = tf.layers.dense(tf.concat([x, y], -1), self.num_units * 2)
i, f = tf.split(z, 2, axis=-1)
y = tf.sigmoid(i) * x + tf.sigmoid(f) * y
encoded = transformer.drop_and_add(
inputs, y, mode, dropout=self.dropout)
if memory is not None:
with tf.variable_scope("multi_head"):
context, last_attention = transformer.multi_head_attention(
self.num_heads,
transformer.norm(encoded),
memory,
mode,
mask=memory_mask,
cache=layer_cache,
dropout=self.attention_dropout,
return_attention=True)
context = transformer.drop_and_add(
encoded,
context,
mode,
dropout=self.dropout)
else:
context = encoded
with tf.variable_scope("ffn"):
transformed = transformer.feed_forward(
transformer.norm(context),
self.ffn_inner_dim,
mode,
dropout=self.relu_dropout)
transformed = transformer.drop_and_add(
context,
transformed,
mode,
dropout=self.dropout)
inputs = transformed
if last_attention is not None:
# The first head of the last layer is returned.
first_head_attention = last_attention[:, 0]
else:
first_head_attention = None
outputs = transformer.norm(inputs)
return outputs, first_head_attention
def _self_attention_stack(self,
inputs, # batch, max_dec_len, emb_dim
sequence_length=None, # [batch]
mode=tf.estimator.ModeKeys.TRAIN,
cache=None,
memory=None, # [batch, enc_len, num_units]
memory_sequence_length=None, # [batch]
step=None):
inputs *= self.num_units ** 0.5
if self.position_encoder is not None:
inputs = self.position_encoder(inputs, position=step + 1 if step is not None else None)
# inputs [batch, max_dec_len, emb_dim]
inputs = tf.layers.dropout( # batch, max_dec_len, emb_dim
inputs,
rate=self.dropout,
training=mode == tf.estimator.ModeKeys.TRAIN)
decoder_mask = None
memory_mask = None
last_attention = None
if self.self_attention_type == "scaled_dot":
if sequence_length is not None: # sequence_length is None when decode, not None at train
decoder_mask = transformer.build_future_mask( # [batch, 1, max_dec_len, max_dec_len]
sequence_length,
num_heads=self.num_heads,
maximum_length=tf.shape(inputs)[1])
elif self.self_attention_type == "average":
if cache is None:
if sequence_length is None:
sequence_length = tf.fill([tf.shape(inputs)[0]], tf.shape(inputs)[1])
decoder_mask = transformer.cumulative_average_mask(
sequence_length, maximum_length=tf.shape(inputs)[1], dtype=inputs.dtype)
if memory is not None and not tf.contrib.framework.nest.is_sequence(memory):
memory = (memory,)
if memory_sequence_length is not None:
if not tf.contrib.framework.nest.is_sequence(memory_sequence_length):
memory_sequence_length = (memory_sequence_length,)
memory_mask = [ # [batch, 1, 1, enc_len]
transformer.build_sequence_mask(
length, num_heads=self.num_heads, maximum_length=tf.shape(m)[1])
for m, length in zip(memory, memory_sequence_length)]
for l in range(self.num_layers):
layer_name = "layer_{}".format(l)
layer_cache = cache[layer_name] if cache is not None else None # train的时候没有cache，decode的时候有cache
with tf.variable_scope(layer_name):
# self attention encode the decoder input (training) or last step output (decode)
if self.self_attention_type == "scaled_dot":
with tf.variable_scope("masked_multi_head"):
encoded = transformer.multi_head_attention( # [batch, decode_len, hidden]
self.num_heads,
transformer.norm(inputs),
None,
mode,
num_units=self.num_units,
mask=decoder_mask, # [batch, 1, len, len]
cache=layer_cache,
dropout=self.attention_dropout)
last_context = transformer.drop_and_add( # [batch, decode_len, hidden]
inputs,
encoded,
mode,
dropout=self.dropout)
elif self.self_attention_type == "average":
with tf.variable_scope("average_attention"):
# Cumulative average.
x = transformer.norm(inputs)
y = transformer.cumulative_average(
x, decoder_mask if cache is None else step, cache=layer_cache)
# FFN.
y = transformer.feed_forward(
y, self.ffn_inner_dim, mode, dropout=self.relu_dropout)
# Gating layer.
z = tf.layers.dense(tf.concat([x, y], -1), self.num_units * 2)
i, f = tf.split(z, 2, axis=-1)
y = tf.sigmoid(i) * x + tf.sigmoid(f) * y
last_context = transformer.drop_and_add(
inputs, y, mode, dropout=self.dropout)
# attending to encoder memory using decoder context
if memory is not None:
for i, (mem, mask) in enumerate(zip(memory, memory_mask)):
memory_cache = layer_cache["memory"][
i] if layer_cache is not None else None # train的时候没有cache，decode的时候有cache
with tf.variable_scope("multi_head" if i == 0 else "multi_head_%d" % i):
context, last_attention = transformer.multi_head_attention(
self.num_heads,
transformer.norm(last_context),
mem, # [batch, enc_len, dim]
mode,
mask=mask, # [batch, 1, 1, len]
cache=memory_cache,
dropout=self.attention_dropout,
return_attention=True)
# context [batch, decode_len, num_units], last_attention train[batch, head, dec_len, enc_len]
last_context = transformer.drop_and_add(
last_context, # [batch, decode_len, num_units]
context,
mode,
dropout=self.dropout)
if i > 0: # Do not return attention in case of multi source.
last_attention = None
with tf.variable_scope("ffn"):
transformed = transformer.feed_forward( # [batch, decode_len, num_units]
transformer.norm(last_context),
self.ffn_inner_dim,
mode,
dropout=self.relu_dropout)
transformed = transformer.drop_and_add( # [batch, decode_len, num_units]
last_context,
transformed,
mode,
dropout=self.dropout)
inputs = transformed
if last_attention is not None:
# The first head of the last layer is returned.
first_head_attention = last_attention[:, 0]
else:
first_head_attention = None
outputs = transformer.norm(inputs) # [batch, decode_len, num_units]
return outputs, first_head_attention
def encode(self,
inputs,
sequence_length=None,
mode=tf.estimator.ModeKeys.TRAIN):
"""
:param inputs: [batch, enc_len, emb_dim]
:param sequence_length: [batch]
:param mode:
:return: outputs: [batch, len, dim] last layer output
state: a tuple ([batch, dim]) * num_layers, contains the sum over len of each layer outputs
sequence_length [batch]
"""
inputs *= self.num_units**0.5
if self.position_encoder is not None:
inputs = self.position_encoder(inputs)
inputs = tf.layers.dropout(
inputs,
rate=self.dropout,
training=mode == tf.estimator.ModeKeys.TRAIN)
mask = transformer.build_sequence_mask( # [batch, 1, 1, enc_len]
sequence_length,
num_heads=self.num_heads,
maximum_length=tf.shape(inputs)[1])
state = ()
for l in range(self.num_layers):
with tf.variable_scope("layer_{}".format(l)):
with tf.variable_scope("multi_head"):
context = transformer.multi_head_attention( # [batch, len, dim]
self.num_heads,
transformer.norm(inputs),
None,
mode,
num_units=self.num_units,
mask=mask,
dropout=self.attention_dropout)
context = transformer.drop_and_add( # [batch, len, dim]
inputs,
context,
mode,
dropout=self.dropout)
with tf.variable_scope("ffn"):
transformed = transformer.feed_forward( # [batch, len, dim]
transformer.norm(context),
self.ffn_inner_dim,
mode,
dropout=self.relu_dropout)
transformed = transformer.drop_and_add( # [batch, len, dim]
context,
transformed,
mode,
dropout=self.dropout)
inputs = transformed
state += (tf.reduce_mean(inputs, axis=1), )
outputs = transformer.norm(inputs) # [batch, len, dim]
return (outputs, state, sequence_length)
