You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
PaddleSpeech/paddlespeech/t2s/modules/transformer/attention.py

349 lines
13 KiB

# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified from espnet(https://github.com/espnet/espnet)
"""Multi-Head Attention layer definition."""
import math
import numpy
import paddle
from paddle import nn
from paddlespeech.t2s.modules.masked_fill import masked_fill
class MultiHeadedAttention(nn.Layer):
"""Multi-Head Attention layer.
Args:
n_head (int):
The number of heads.
n_feat (int):
The number of features.
dropout_rate (float):
Dropout rate.
"""
def __init__(self, n_head, n_feat, dropout_rate):
"""Construct an MultiHeadedAttention object."""
super().__init__()
assert n_feat % n_head == 0
# We assume d_v always equals d_k
self.d_k = n_feat // n_head
self.h = n_head
self.linear_q = nn.Linear(n_feat, n_feat, bias_attr=True)
self.linear_k = nn.Linear(n_feat, n_feat, bias_attr=True)
self.linear_v = nn.Linear(n_feat, n_feat, bias_attr=True)
self.linear_out = nn.Linear(n_feat, n_feat, bias_attr=True)
self.attn = None
self.dropout = nn.Dropout(p=dropout_rate)
def forward_qkv(self, query, key, value):
"""Transform query, key and value.
Args:
query(Tensor):
query tensor (#batch, time1, size).
key(Tensor):
Key tensor (#batch, time2, size).
value(Tensor):
Value tensor (#batch, time2, size).
Returns:
Tensor:
Transformed query tensor (#batch, n_head, time1, d_k).
Tensor:
Transformed key tensor (#batch, n_head, time2, d_k).
Tensor:
Transformed value tensor (#batch, n_head, time2, d_k).
"""
n_batch = paddle.shape(query)[0]
q = paddle.reshape(
self.linear_q(query), [n_batch, -1, self.h, self.d_k])
k = paddle.reshape(self.linear_k(key), [n_batch, -1, self.h, self.d_k])
v = paddle.reshape(
self.linear_v(value), [n_batch, -1, self.h, self.d_k])
# (batch, head, time1, d_k)
q = q.transpose((0, 2, 1, 3))
# (batch, head, time2, d_k)
k = k.transpose((0, 2, 1, 3))
# (batch, head, time2, d_k)
v = v.transpose((0, 2, 1, 3))
return q, k, v
def forward_attention(self, value, scores, mask=None):
"""Compute attention context vector.
Args:
value(Tensor):
Transformed value (#batch, n_head, time2, d_k).
scores(Tensor):
Attention score (#batch, n_head, time1, time2).
mask(Tensor, optional):
Mask (#batch, 1, time2) or (#batch, time1, time2). (Default value = None)
Returns:
Tensor: Transformed value (#batch, time1, d_model) weighted by the attention score (#batch, time1, time2).
"""
n_batch = paddle.shape(value)[0]
softmax = paddle.nn.Softmax(axis=-1)
if mask is not None:
mask = mask.unsqueeze(1)
mask = paddle.logical_not(mask)
# assume scores.dtype==paddle.float32, we only use "float32" here
dtype = str(scores.dtype).split(".")[-1]
min_value = numpy.finfo(dtype).min
scores = masked_fill(scores, mask, min_value)
# (batch, head, time1, time2)
self.attn = softmax(scores)
self.attn = masked_fill(self.attn, mask, 0.0)
else:
# (batch, head, time1, time2)
self.attn = softmax(scores)
# (batch, head, time1, time2)
p_attn = self.dropout(self.attn)
# (batch, head, time1, time2) * (batch, head, time2, d_k) -> # (batch, head, time1, d_k)
x = paddle.matmul(p_attn, value)
# (batch, time1, d_model)
x = (paddle.reshape(
x.transpose((0, 2, 1, 3)), (n_batch, -1, self.h * self.d_k)))
# (batch, time1, d_model)
return self.linear_out(x)
def forward(self, query, key, value, mask=None):
"""Compute scaled dot product attention.
Args:
query(Tensor):
Query tensor (#batch, time1, size).
key(Tensor):
Key tensor (#batch, time2, size).
value(Tensor):
Value tensor (#batch, time2, size).
mask(Tensor, optional):
Mask tensor (#batch, 1, time2) or (#batch, time1, time2). (Default value = None)
Returns:
Tensor: Output tensor (#batch, time1, d_model).
"""
q, k, v = self.forward_qkv(query, key, value)
scores = paddle.matmul(q, k.transpose(
(0, 1, 3, 2))) / math.sqrt(self.d_k)
return self.forward_attention(v, scores, mask)
class RelPositionMultiHeadedAttention(MultiHeadedAttention):
"""Multi-Head Attention layer with relative position encoding (new implementation).
Details can be found in https://github.com/espnet/espnet/pull/2816.
Paper: https://arxiv.org/abs/1901.02860
Args:
n_head (int):
The number of heads.
n_feat (int):
The number of features.
dropout_rate (float):
Dropout rate.
zero_triu (bool):
Whether to zero the upper triangular part of attention matrix.
"""
def __init__(self, n_head, n_feat, dropout_rate, zero_triu=False):
"""Construct an RelPositionMultiHeadedAttention object."""
super().__init__(n_head, n_feat, dropout_rate)
self.zero_triu = zero_triu
# linear transformation for positional encoding
self.linear_pos = nn.Linear(n_feat, n_feat, bias_attr=False)
# these two learnable bias are used in matrix c and matrix d
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
self.pos_bias_u = paddle.create_parameter(
shape=(self.h, self.d_k),
dtype='float32',
default_initializer=paddle.nn.initializer.XavierUniform())
self.pos_bias_v = paddle.create_parameter(
shape=(self.h, self.d_k),
dtype='float32',
default_initializer=paddle.nn.initializer.XavierUniform())
def rel_shift(self, x):
"""Compute relative positional encoding.
Args:
x(Tensor):
Input tensor (batch, head, time1, 2*time1-1).
Returns:
Tensor: Output tensor.
"""
b, h, t1, t2 = paddle.shape(x)
zero_pad = paddle.zeros((b, h, t1, 1))
x_padded = paddle.concat([zero_pad, x], axis=-1)
x_padded = x_padded.reshape([b, h, t2 + 1, t1])
# only keep the positions from 0 to time2
x = x_padded[:, :, 1:].reshape([b, h, t1, t2])[:, :, :, :t2 // 2 + 1]
if self.zero_triu:
ones = paddle.ones((t1, t2))
x = x * paddle.tril(ones, t2 - 1)[None, None, :, :]
return x
def forward(self, query, key, value, pos_emb, mask):
"""Compute 'Scaled Dot Product Attention' with rel. positional encoding.
Args:
query(Tensor):
Query tensor (#batch, time1, size).
key(Tensor):
Key tensor (#batch, time2, size).
value(Tensor):
Value tensor (#batch, time2, size).
pos_emb(Tensor):
Positional embedding tensor (#batch, 2*time1-1, size).
mask(Tensor):
Mask tensor (#batch, 1, time2) or (#batch, time1, time2).
Returns:
Tensor: Output tensor (#batch, time1, d_model).
"""
q, k, v = self.forward_qkv(query, key, value)
# (batch, time1, head, d_k)
q = q.transpose([0, 2, 1, 3])
n_batch_pos = paddle.shape(pos_emb)[0]
p = self.linear_pos(pos_emb).reshape(
[n_batch_pos, -1, self.h, self.d_k])
# (batch, head, 2*time1-1, d_k)
p = p.transpose([0, 2, 1, 3])
# (batch, head, time1, d_k)
q_with_bias_u = (q + self.pos_bias_u).transpose([0, 2, 1, 3])
# (batch, head, time1, d_k)
q_with_bias_v = (q + self.pos_bias_v).transpose([0, 2, 1, 3])
# compute attention score
# first compute matrix a and matrix c
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
# (batch, head, time1, time2)
matrix_ac = paddle.matmul(q_with_bias_u, k.transpose([0, 1, 3, 2]))
# compute matrix b and matrix d
# (batch, head, time1, 2*time1-1)
matrix_bd = paddle.matmul(q_with_bias_v, p.transpose([0, 1, 3, 2]))
matrix_bd = self.rel_shift(matrix_bd)
# (batch, head, time1, time2)
scores = (matrix_ac + matrix_bd) / math.sqrt(self.d_k)
return self.forward_attention(v, scores, mask)
class LegacyRelPositionMultiHeadedAttention(MultiHeadedAttention):
"""Multi-Head Attention layer with relative position encoding (old version).
Details can be found in https://github.com/espnet/espnet/pull/2816.
Paper: https://arxiv.org/abs/1901.02860
Args:
n_head (int):
The number of heads.
n_feat (int):
The number of features.
dropout_rate (float):
Dropout rate.
zero_triu (bool):
Whether to zero the upper triangular part of attention matrix.
"""
def __init__(self, n_head, n_feat, dropout_rate, zero_triu=False):
"""Construct an RelPositionMultiHeadedAttention object."""
super().__init__(n_head, n_feat, dropout_rate)
self.zero_triu = zero_triu
# linear transformation for positional encoding
self.linear_pos = nn.Linear(n_feat, n_feat, bias_attr=False)
# these two learnable bias are used in matrix c and matrix d
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
self.pos_bias_u = paddle.create_parameter(
shape=(self.h, self.d_k),
dtype='float32',
default_initializer=paddle.nn.initializer.XavierUniform())
self.pos_bias_v = paddle.create_parameter(
shape=(self.h, self.d_k),
dtype='float32',
default_initializer=paddle.nn.initializer.XavierUniform())
def rel_shift(self, x):
"""Compute relative positional encoding.
Args:
x(Tensor):
Input tensor (batch, head, time1, time2).
Returns:
Tensor:Output tensor.
"""
b, h, t1, t2 = paddle.shape(x)
zero_pad = paddle.zeros((b, h, t1, 1))
x_padded = paddle.concat([zero_pad, x], axis=-1)
x_padded = paddle.reshape(x_padded, [b, h, t2 + 1, t1])
# only keep the positions from 0 to time2
x = paddle.reshape(x_padded[:, :, 1:], [b, h, t1, t2])
if self.zero_triu:
ones = paddle.ones((t1, t2))
x = x * paddle.tril(ones, t2 - 1)[None, None, :, :]
return x
def forward(self, query, key, value, pos_emb, mask):
"""Compute 'Scaled Dot Product Attention' with rel. positional encoding.
Args:
query(Tensor): Query tensor (#batch, time1, size).
key(Tensor): Key tensor (#batch, time2, size).
value(Tensor): Value tensor (#batch, time2, size).
pos_emb(Tensor): Positional embedding tensor (#batch, time1, size).
mask(Tensor): Mask tensor (#batch, 1, time2) or (#batch, time1, time2).
Returns:
Tensor: Output tensor (#batch, time1, d_model).
"""
q, k, v = self.forward_qkv(query, key, value)
# (batch, time1, head, d_k)
q = paddle.transpose(q, [0, 2, 1, 3])
n_batch_pos = paddle.shape(pos_emb)[0]
p = paddle.reshape(
self.linear_pos(pos_emb), [n_batch_pos, -1, self.h, self.d_k])
# (batch, head, time1, d_k)
p = paddle.transpose(p, [0, 2, 1, 3])
# (batch, head, time1, d_k)
q_with_bias_u = paddle.transpose((q + self.pos_bias_u), [0, 2, 1, 3])
# (batch, head, time1, d_k)
q_with_bias_v = paddle.transpose((q + self.pos_bias_v), [0, 2, 1, 3])
# compute attention score
# first compute matrix a and matrix c
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
# (batch, head, time1, time2)
matrix_ac = paddle.matmul(q_with_bias_u,
paddle.transpose(k, [0, 1, 3, 2]))
# compute matrix b and matrix d
# (batch, head, time1, time1)
matrix_bd = paddle.matmul(q_with_bias_v,
paddle.transpose(p, [0, 1, 3, 2]))
matrix_bd = self.rel_shift(matrix_bd)
# (batch, head, time1, time2)
scores = (matrix_ac + matrix_bd) / math.sqrt(self.d_k)
return self.forward_attention(v, scores, mask)