def call(self, inputs, mask=None, **kwargs):
if isinstance(inputs, list):
query, key, value = inputs
else:
query = key = value = inputs
if isinstance(mask, list):
mask = mask[1]
feature_dim = K.shape(query)[-1]
e = K.batch_dot(query, key, axes=2) / K.sqrt(
K.cast(feature_dim, dtype=K.floatx()))
e = K.exp(e - K.max(e, axis=-1, keepdims=True))
if self.history_only:
query_len, key_len = K.shape(query)[1], K.shape(key)[1]
indices = K.expand_dims(K.arange(0, key_len), axis=0)
upper = K.expand_dims(K.arange(0, query_len), axis=-1)
e *= K.expand_dims(K.cast(indices <= upper, K.floatx()), axis=0)
if mask is not None:
e *= K.cast(K.expand_dims(mask, axis=-2), K.floatx())
s = K.sum(e, axis=-1, keepdims=True)
s += K.cast(K.equal(s, 0.0), K.floatx())
a = e / s
v = K.batch_dot(a, value)
if self.return_attention:
return [v, a]
return v
def call(self, inputs, training=None, **kwargs):
inputs, memory = inputs
batch_size = K.shape(inputs)[0]
seq_len = K.shape(inputs)[1]
mem_mask = K.tile(K.ones_like(memory[:, :, :1], dtype=K.floatx()), [1, 1, seq_len])
# Build content mask with random permutation
ranges = K.tile(K.expand_dims(K.arange(0, seq_len), axis=-1), [1, batch_size])
if self.enabled:
shuffle = random_shuffle(ranges)
else:
shuffle = ranges
if self.directional:
shuffled = K.in_train_phase(shuffle, ranges, training)
else:
if self.enabled:
shuffled = K.in_train_phase(shuffle, ranges + seq_len, training)
else:
shuffled = ranges + seq_len
ranges = K.expand_dims(K.permute_dimensions(ranges, [1, 0]), axis=-1)
shuffled = K.expand_dims(K.permute_dimensions(shuffled, [1, 0]), axis=1)
content_mask = K.cast(ranges <= shuffled, dtype=K.floatx())
# Build query mask based on content mask
ranges = K.arange(0, seq_len)
eye = K.equal(K.expand_dims(ranges, axis=0), K.expand_dims(ranges, axis=-1))
eye = K.expand_dims(K.cast(eye, dtype=K.floatx()), axis=0)
query_mask = content_mask * (1.0 - eye)
content_mask = K.concatenate([mem_mask, content_mask], axis=1)
query_mask = K.concatenate([mem_mask, query_mask], axis=1)
return [
K.permute_dimensions(content_mask, [0, 2, 1]),
K.permute_dimensions(query_mask, [0, 2, 1]),
]
def call(self, inputs, **kwargs):
inputs, memory_length = inputs
memory_length = K.cast(memory_length[0][0], 'int32')
batch_size = K.cast(K.shape(inputs)[0], 'int32')
seq_len = K.cast(K.shape(inputs)[1], 'int32')
# Build new memory
pad = K.tile(inputs[0:1, ...], (self.batch_size - batch_size, 1, 1))
padded = K.concatenate([inputs, pad], axis=0) # (self.batch_size, seq_len, output_dim)
new_memory = K.concatenate([self.memory, padded], axis=1) # (self.batch_size, self.memory_len + seq_len, ...)
new_memory = tf.slice( # (self.batch_size, self.memory_len, output_dim)
new_memory,
(0, seq_len, 0),
(self.batch_size, self.memory_len + self.target_len, self.output_dim),
)
self.add_update(K.update(self.memory, new_memory), inputs)
# Build output
old_memory = tf.slice( # (batch_size, memory_length, output_dim)
new_memory,
(0, K.maximum(0, self.memory_len + self.target_len - seq_len - memory_length), 0),
(batch_size, K.minimum(self.memory_len, memory_length), self.output_dim),
)
return old_memory
def _attention_regularizer(self, attention):
batch_size = K.cast(K.shape(attention)[0], K.floatx())
input_len = K.shape(attention)[-1]
indices = K.expand_dims(K.arange(0, input_len), axis=0)
diagonal = K.expand_dims(K.arange(0, input_len), axis=-1)
eye = K.cast(K.equal(indices, diagonal), K.floatx())
return self.attention_regularizer_weight * K.sum(K.square(K.batch_dot(
attention,
K.permute_dimensions(attention, (0, 2, 1))) - eye)) / batch_size
def _relative_shift(x, key_len_expected=-1):
batch_size, q_len, k_len = K.shape(x)[0], K.shape(x)[1], K.shape(x)[2]
x = K.reshape(
x, (batch_size, k_len,
q_len)) # (batch * n_head, prev_len + seq_len + 1, seq_len)
x = x[:, 1:, :] # (batch * n_head, prev_len + seq_len, seq_len)
x = K.reshape(x, (batch_size, q_len, k_len -
1)) # (batch * n_head, seq_len, prev_len + seq_len)
x = tf.slice(
x, (0, 0, 0),
(-1, -1,
key_len_expected)) # (batch * n_head, seq_len, key_len_expected)
return x
def call(self, inputs, **kwargs):
length = K.shape(inputs[0])[1] + K.shape(inputs[1])[1]
inputs = K.tile(
K.expand_dims(K.arange(length - 1, -1, -1, dtype=K.floatx()),
axis=0),
[K.shape(inputs[0])[0], 1],
)
if self.clamp_len is not None:
inputs = K.clip(inputs, min_value=0, max_value=self.clamp_len)
inputs = K.expand_dims(inputs, axis=-1)
output_dim = K.cast(self.output_dim, K.floatx())
ranges = K.expand_dims(K.arange(0.0, self.output_dim, 2.0),
axis=0) / output_dim
inverse = 1.0 / K.pow(10000.0, ranges)
positions = inputs * inverse
return K.concatenate([K.sin(positions), K.cos(positions)], axis=-1)
def _reshape_mask(mask, head_num):
if mask is None:
return mask
seq_len = K.shape(mask)[1]
mask = K.expand_dims(mask, axis=1)
mask = K.tile(mask, [1, head_num, 1])
return K.reshape(mask, (-1, seq_len))
def _reshape_to_batches(self, x):
input_shape = K.shape(x)
batch_size, seq_len, feature_dim = input_shape[0], input_shape[
1], input_shape[2]
x = K.reshape(x, (batch_size, seq_len, self.num_head, self.units_head))
x = K.permute_dimensions(x, [0, 2, 1, 3])
return K.reshape(
x, (batch_size * self.num_head, seq_len, self.units_head))
def _reshape_to_batches(x, head_num):
input_shape = K.shape(x)
batch_size, seq_len, feature_dim = input_shape[0], input_shape[
1], input_shape[2]
head_dim = feature_dim // head_num
x = K.reshape(x, (batch_size, seq_len, head_num, head_dim))
x = K.permute_dimensions(x, [0, 2, 1, 3])
return K.reshape(x, (batch_size * head_num, seq_len, head_dim))
def call(self, inputs, **kwargs):
q_len, m_len = K.shape(inputs[0])[1], K.shape(inputs[1])[1]
k_len = q_len + m_len
start, stop = k_len, -1
if not self.directional:
stop = -q_len
inputs = K.tile(
K.expand_dims(K.arange(start, stop, -1, dtype=K.floatx()), axis=0),
[K.shape(inputs[0])[0], 1],
)
if self.clamp_len is not None:
inputs = K.clip(inputs, min_value=0, max_value=self.clamp_len)
inputs = K.expand_dims(inputs, axis=-1)
output_dim = K.cast(self.output_dim, K.floatx())
ranges = K.expand_dims(K.arange(0.0, self.output_dim, 2.0),
axis=0) / output_dim
inverse = 1.0 / K.pow(10000.0, ranges)
positions = inputs * inverse
return K.concatenate([K.sin(positions), K.cos(positions)], axis=-1)
def call(self, inputs, mask=None):
input_shape = K.shape(inputs)
if self.mode == self.MODE_ADD:
batch_size, seq_len, output_dim = input_shape[0], input_shape[
1], input_shape[2]
pos_input = K.tile(K.expand_dims(K.arange(0, seq_len), axis=0),
[batch_size, 1])
elif self.mode == self.MODE_CONCAT:
batch_size, seq_len, output_dim = input_shape[0], input_shape[
1], self.output_dim
pos_input = K.tile(K.expand_dims(K.arange(0, seq_len), axis=0),
[batch_size, 1])
else:
output_dim = self.output_dim
pos_input = inputs
if K.dtype(pos_input) != K.floatx():
pos_input = K.cast(pos_input, K.floatx())
evens = K.arange(0, output_dim // 2) * 2
odds = K.arange(0, output_dim // 2) * 2 + 1
even_embd = K.sin(
K.dot(
K.expand_dims(pos_input, -1),
K.expand_dims(
1.0 / K.pow(
10000.0,
K.cast(evens, K.floatx()) /
K.cast(output_dim, K.floatx())), 0)))
odd_embd = K.cos(
K.dot(
K.expand_dims(pos_input, -1),
K.expand_dims(
1.0 / K.pow(
10000.0,
K.cast((odds - 1), K.floatx()) /
K.cast(output_dim, K.floatx())), 0)))
embd = K.stack([even_embd, odd_embd], axis=-1)
output = K.reshape(embd, [-1, K.shape(inputs)[1], output_dim])
if self.mode == self.MODE_CONCAT:
output = K.concatenate([inputs, output], axis=-1)
if self.mode == self.MODE_ADD:
output += inputs
return output
def call(self, x, mask=None):
logits = K.dot(x, self.W)
if self.use_bias:
logits += self.b
x_shape = K.shape(x)
logits = K.reshape(logits, (x_shape[0], x_shape[1]))
ai = K.exp(logits - K.max(logits, axis=-1, keepdims=True))
if mask is not None:
mask = K.cast(mask, K.floatx())
ai = ai * mask
att_weights = ai / (K.sum(ai, axis=1, keepdims=True) + K.epsilon())
weighted_input = x * K.expand_dims(att_weights)
result = K.sum(weighted_input, axis=1)
if self.return_attention:
return [result, att_weights]
return result
def _call_additive_emission(self, inputs):
input_shape = K.shape(inputs)
batch_size, input_len = input_shape[0], input_shape[1]
# h_{t, t'} = \tanh(x_t^T W_t + x_{t'}^T W_x + b_h)
q = K.expand_dims(K.dot(inputs, self.Wt), 2)
k = K.expand_dims(K.dot(inputs, self.Wx), 1)
if self.use_additive_bias:
h = K.tanh(q + k + self.bh)
else:
h = K.tanh(q + k)
# e_{t, t'} = W_a h_{t, t'} + b_a
if self.use_attention_bias:
e = K.reshape(K.dot(h, self.Wa) + self.ba, (batch_size, input_len, input_len))
else:
e = K.reshape(K.dot(h, self.Wa), (batch_size, input_len, input_len))
return e
def call(self, inputs, mask=None, **kwargs):
input_len = K.shape(inputs)[1]
if self.attention_type == SeqSelfAttention.ATTENTION_TYPE_ADD:
e = self._call_additive_emission(inputs)
elif self.attention_type == SeqSelfAttention.ATTENTION_TYPE_MUL:
e = self._call_multiplicative_emission(inputs)
if self.attention_activation is not None:
e = self.attention_activation(e)
e = K.exp(e - K.max(e, axis=-1, keepdims=True))
if self.attention_width is not None:
if self.history_only:
lower = K.arange(0, input_len) - (self.attention_width - 1)
else:
lower = K.arange(0, input_len) - self.attention_width // 2
lower = K.expand_dims(lower, axis=-1)
upper = lower + self.attention_width
indices = K.expand_dims(K.arange(0, input_len), axis=0)
e = e * K.cast(lower <= indices, K.floatx()) * K.cast(indices < upper, K.floatx())
if mask is not None:
mask = K.cast(mask, K.floatx())
mask = K.expand_dims(mask)
e = K.permute_dimensions(K.permute_dimensions(e * mask, (0, 2, 1)) * mask, (0, 2, 1))
# a_{t} = \text{softmax}(e_t)
s = K.sum(e, axis=-1, keepdims=True)
a = e / (s + K.epsilon())
# l_t = \sum_{t'} a_{t, t'} x_{t'}
v = K.batch_dot(a, inputs)
if self.attention_regularizer_weight > 0.0:
self.add_loss(self._attention_regularizer(a))
if self.return_attention:
return [v, a]
return v
def call(self, inputs, mask=None, training=None):
(inputs, content, memories, segment_mat, segment_embed, relatives,
bias_context, bias_relative, bias_segment, permutation) = inputs
full = K.concatenate([memories, content],
axis=1) # (batch, prev_len + seq_len, units)
kernel_q = self.kernel[:, :self.units]
kernel_kv = self.kernel[:, self.units:self.units * 3]
kernel_r = self.kernel[:, self.units * 3:self.units * 4]
kernel_o = self.kernel[:, self.units * 4:self.units * 5]
bias_q, bias_kv, bias_r, bias_o = (None, ) * 4
if self.use_bias:
bias_q = self.bias[:self.units]
bias_kv = self.bias[self.units:self.units * 3]
bias_r = self.bias[self.units * 3:self.units * 4]
bias_o = self.bias[self.units * 4:self.units * 5]
w_q = K.dot(inputs, kernel_q) # (batch, seq_len, units)
w_kv = K.dot(full, kernel_kv) # (batch, prev_len + seq_len, units * 2)
w_r = K.dot(relatives, kernel_r) # (batch, prev_len + seq_len, units)
if self.use_bias:
w_q = K.bias_add(w_q, bias_q)
w_kv = K.bias_add(w_kv, bias_kv)
w_r = K.bias_add(w_r, bias_r)
if self.activation is not None:
w_q = self.activation(w_q)
w_kv = self.activation(w_kv)
w_r = self.activation(w_r)
w_k = w_kv[:, :, :self.units] # (batch, prev_len + seq_len, units)
w_v = w_kv[:, :, self.units:] # (batch, prev_len + seq_len, units)
batch_size, q_len, k_len = K.shape(inputs)[0], K.shape(
w_q)[1], K.shape(w_k)[1]
w_qc = K.bias_add(w_q, bias_context)
w_qc = self._reshape_to_batches(
w_qc) # (batch * n_head, seq_len, units_head)
w_k = self._reshape_to_batches(
w_k) # (batch * n_head, prev_len + seq_len, units_head)
a_context = K.batch_dot(
w_qc, w_k, axes=2) # (batch * n_head, seq_len, prev_len + seq_len)
w_qr = K.bias_add(w_q, bias_relative)
w_qr = self._reshape_to_batches(
w_qr) # (batch * n_head, seq_len, units_head)
w_r = self._reshape_to_batches(
w_r) # (batch * n_head, prev_len + seq_len, units_head)
a_relative = K.batch_dot(
w_qr, w_r, axes=2) # (batch * n_head, seq_len, prev_len + seq_len)
a_relative = self._relative_shift( # (batch * n_head, seq_len, prev_len + seq_len)
a_relative,
key_len_expected=K.shape(a_context)[-1],
)
w_qs = K.bias_add(w_q, bias_segment)
w_qs = K.reshape(w_qs, (-1, q_len, self.num_head, self.units_head))
w_qs = K.permute_dimensions(
w_qs, (2, 0, 1, 3)) # (n_head, batch, seq_len, units_head)
segment_embed = K.reshape(K.transpose(segment_embed),
(self.num_head, 1, self.units_head, 2))
segment_embed = K.tile(segment_embed, (1, batch_size, 1, 1))
a_segment = K.batch_dot(w_qs, segment_embed,
axes=(3, 2)) # (n_head, batch, seq_len, 2)
a_segment = K.permute_dimensions(
a_segment, (1, 2, 3, 0)) # (batch, seq_len, 2, n_head)
a_segment = K.batch_dot(
segment_mat, a_segment,
axes=(3, 2)) # (batch, seq_len, prev_len + seq_len, n_head)
a_segment = K.reshape(K.permute_dimensions(a_segment, (0, 3, 1, 2)),
(-1, q_len, k_len))
att = (a_context + a_relative + a_segment) / K.sqrt(
K.constant(self.units_head, dtype=K.floatx()))
exp = K.exp(att - K.max(att, axis=-1, keepdims=True))
permutation = K.tile(K.expand_dims(permutation, axis=1),
[1, self.num_head, 1, 1])
permutation = K.reshape(permutation, (-1, q_len, k_len))
exp *= permutation
if mask is not None and mask[0] is not None:
mask = K.cast(mask[0], K.floatx())
mask = K.concatenate([K.ones_like(memories[:, :, 0]), mask],
axis=1)
exp *= K.expand_dims(self._reshape_mask(mask), axis=1)
att = exp / (K.sum(exp, axis=-1, keepdims=True) + K.epsilon())
if self.att_drop_layer is not None:
att = self.att_drop_layer(att, training=training)
w_v = self._reshape_to_batches(
w_v) # (batch * n_head, prev_len + seq_len, units_head)
w_o = K.batch_dot(att, w_v) # (batch * n_head, seq_len, units_head)
w_o = self._reshape_from_batches(w_o) # (batch, seq_len, units)
w_o = K.dot(w_o, kernel_o) # (batch, seq_len, units)
if self.use_bias:
w_o = K.bias_add(w_o, bias_o)
if self.activation is not None:
w_o = self.activation(w_o)
if TF_KERAS:
# Add shape information to tensor when using `tf.keras`
input_shape = K.int_shape(inputs)
if input_shape[1] is not None:
w_o = K.reshape(w_o, (-1, ) + input_shape[1:])
return w_o
def _reshape_mask(self, mask):
seq_len = K.shape(mask)[1]
mask = K.expand_dims(mask, axis=1)
mask = K.tile(mask, [1, self.num_head, 1])
return K.reshape(mask, (-1, seq_len))
