mirror of
https://github.com/babysor/MockingBird.git
synced 2024-03-22 13:11:31 +08:00
b617a87ee4
* Init ppg extractor and ppg2mel * add preprocess and training * FIx known issues * Update __init__.py Allow to gen audio * Fix length issue * Fix bug of preparing fid * Fix sample issues * Add UI usage of PPG-vc
184 lines
6.9 KiB
Python
184 lines
6.9 KiB
Python
#!/usr/bin/env python3
|
|
# -*- coding: utf-8 -*-
|
|
|
|
# Copyright 2019 Shigeki Karita
|
|
# Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
|
|
|
|
"""Multi-Head Attention layer definition."""
|
|
|
|
import math
|
|
|
|
import numpy
|
|
import torch
|
|
from torch import nn
|
|
|
|
|
|
class MultiHeadedAttention(nn.Module):
|
|
"""Multi-Head Attention layer.
|
|
|
|
:param int n_head: the number of head s
|
|
:param int n_feat: the number of features
|
|
:param float dropout_rate: dropout rate
|
|
|
|
"""
|
|
|
|
def __init__(self, n_head, n_feat, dropout_rate):
|
|
"""Construct an MultiHeadedAttention object."""
|
|
super(MultiHeadedAttention, self).__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)
|
|
self.linear_k = nn.Linear(n_feat, n_feat)
|
|
self.linear_v = nn.Linear(n_feat, n_feat)
|
|
self.linear_out = nn.Linear(n_feat, n_feat)
|
|
self.attn = None
|
|
self.dropout = nn.Dropout(p=dropout_rate)
|
|
|
|
def forward_qkv(self, query, key, value):
|
|
"""Transform query, key and value.
|
|
|
|
:param torch.Tensor query: (batch, time1, size)
|
|
:param torch.Tensor key: (batch, time2, size)
|
|
:param torch.Tensor value: (batch, time2, size)
|
|
:return torch.Tensor transformed query, key and value
|
|
|
|
"""
|
|
n_batch = query.size(0)
|
|
q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
|
|
k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
|
|
v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
|
|
q = q.transpose(1, 2) # (batch, head, time1, d_k)
|
|
k = k.transpose(1, 2) # (batch, head, time2, d_k)
|
|
v = v.transpose(1, 2) # (batch, head, time2, d_k)
|
|
|
|
return q, k, v
|
|
|
|
def forward_attention(self, value, scores, mask):
|
|
"""Compute attention context vector.
|
|
|
|
:param torch.Tensor value: (batch, head, time2, size)
|
|
:param torch.Tensor scores: (batch, head, time1, time2)
|
|
:param torch.Tensor mask: (batch, 1, time2) or (batch, time1, time2)
|
|
:return torch.Tensor transformed `value` (batch, time1, d_model)
|
|
weighted by the attention score (batch, time1, time2)
|
|
|
|
"""
|
|
n_batch = value.size(0)
|
|
if mask is not None:
|
|
mask = mask.unsqueeze(1).eq(0) # (batch, 1, *, time2)
|
|
min_value = float(
|
|
numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min
|
|
)
|
|
scores = scores.masked_fill(mask, min_value)
|
|
self.attn = torch.softmax(scores, dim=-1).masked_fill(
|
|
mask, 0.0
|
|
) # (batch, head, time1, time2)
|
|
else:
|
|
self.attn = torch.softmax(scores, dim=-1) # (batch, head, time1, time2)
|
|
|
|
p_attn = self.dropout(self.attn)
|
|
x = torch.matmul(p_attn, value) # (batch, head, time1, d_k)
|
|
x = (
|
|
x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
|
|
) # (batch, time1, d_model)
|
|
|
|
return self.linear_out(x) # (batch, time1, d_model)
|
|
|
|
def forward(self, query, key, value, mask):
|
|
"""Compute 'Scaled Dot Product Attention'.
|
|
|
|
:param torch.Tensor query: (batch, time1, size)
|
|
:param torch.Tensor key: (batch, time2, size)
|
|
:param torch.Tensor value: (batch, time2, size)
|
|
:param torch.Tensor mask: (batch, 1, time2) or (batch, time1, time2)
|
|
:param torch.nn.Dropout dropout:
|
|
:return torch.Tensor: attention output (batch, time1, d_model)
|
|
"""
|
|
q, k, v = self.forward_qkv(query, key, value)
|
|
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
|
|
return self.forward_attention(v, scores, mask)
|
|
|
|
|
|
class RelPositionMultiHeadedAttention(MultiHeadedAttention):
|
|
"""Multi-Head Attention layer with relative position encoding.
|
|
|
|
Paper: https://arxiv.org/abs/1901.02860
|
|
|
|
:param int n_head: the number of head s
|
|
:param int n_feat: the number of features
|
|
:param float dropout_rate: dropout rate
|
|
|
|
"""
|
|
|
|
def __init__(self, n_head, n_feat, dropout_rate):
|
|
"""Construct an RelPositionMultiHeadedAttention object."""
|
|
super().__init__(n_head, n_feat, dropout_rate)
|
|
# linear transformation for positional ecoding
|
|
self.linear_pos = nn.Linear(n_feat, n_feat, bias=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 = nn.Parameter(torch.Tensor(self.h, self.d_k))
|
|
self.pos_bias_v = nn.Parameter(torch.Tensor(self.h, self.d_k))
|
|
torch.nn.init.xavier_uniform_(self.pos_bias_u)
|
|
torch.nn.init.xavier_uniform_(self.pos_bias_v)
|
|
|
|
def rel_shift(self, x, zero_triu=False):
|
|
"""Compute relative positinal encoding.
|
|
|
|
:param torch.Tensor x: (batch, time, size)
|
|
:param bool zero_triu: return the lower triangular part of the matrix
|
|
"""
|
|
zero_pad = torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype)
|
|
x_padded = torch.cat([zero_pad, x], dim=-1)
|
|
|
|
x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2))
|
|
x = x_padded[:, :, 1:].view_as(x)
|
|
|
|
if zero_triu:
|
|
ones = torch.ones((x.size(2), x.size(3)))
|
|
x = x * torch.tril(ones, x.size(3) - x.size(2))[None, None, :, :]
|
|
|
|
return x
|
|
|
|
def forward(self, query, key, value, pos_emb, mask):
|
|
"""Compute 'Scaled Dot Product Attention' with rel. positional encoding.
|
|
|
|
:param torch.Tensor query: (batch, time1, size)
|
|
:param torch.Tensor key: (batch, time2, size)
|
|
:param torch.Tensor value: (batch, time2, size)
|
|
:param torch.Tensor pos_emb: (batch, time1, size)
|
|
:param torch.Tensor mask: (batch, time1, time2)
|
|
:param torch.nn.Dropout dropout:
|
|
:return torch.Tensor: attention output (batch, time1, d_model)
|
|
"""
|
|
q, k, v = self.forward_qkv(query, key, value)
|
|
q = q.transpose(1, 2) # (batch, time1, head, d_k)
|
|
|
|
n_batch_pos = pos_emb.size(0)
|
|
p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k)
|
|
p = p.transpose(1, 2) # (batch, head, time1, d_k)
|
|
|
|
# (batch, head, time1, d_k)
|
|
q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
|
|
# (batch, head, time1, d_k)
|
|
q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)
|
|
|
|
# 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 = torch.matmul(q_with_bias_u, k.transpose(-2, -1))
|
|
|
|
# compute matrix b and matrix d
|
|
# (batch, head, time1, time2)
|
|
matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
|
|
matrix_bd = self.rel_shift(matrix_bd)
|
|
|
|
scores = (matrix_ac + matrix_bd) / math.sqrt(
|
|
self.d_k
|
|
) # (batch, head, time1, time2)
|
|
|
|
return self.forward_attention(v, scores, mask)
|