File size: 6,004 Bytes
5548515
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
# Copyright (c) 2023 Amphion.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import numpy as np

from .Layers import FFTBlock
from text.symbols import symbols

PAD = 0
UNK = 1
BOS = 2
EOS = 3

PAD_WORD = "<blank>"
UNK_WORD = "<unk>"
BOS_WORD = "<s>"
EOS_WORD = "</s>"


def get_sinusoid_encoding_table(n_position, d_hid, padding_idx=None):
    """Sinusoid position encoding table"""

    def cal_angle(position, hid_idx):
        return position / np.power(10000, 2 * (hid_idx // 2) / d_hid)

    def get_posi_angle_vec(position):
        return [cal_angle(position, hid_j) for hid_j in range(d_hid)]

    sinusoid_table = np.array(
        [get_posi_angle_vec(pos_i) for pos_i in range(n_position)]
    )

    sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2])  # dim 2i
    sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2])  # dim 2i+1

    if padding_idx is not None:
        # zero vector for padding dimension
        sinusoid_table[padding_idx] = 0.0

    return torch.FloatTensor(sinusoid_table)


class Encoder(nn.Module):
    """Encoder"""

    def __init__(self, config):
        super(Encoder, self).__init__()

        n_position = config["max_seq_len"] + 1
        n_src_vocab = len(symbols) + 1
        d_word_vec = config["transformer"]["encoder_hidden"]
        n_layers = config["transformer"]["encoder_layer"]
        n_head = config["transformer"]["encoder_head"]
        d_k = d_v = (
            config["transformer"]["encoder_hidden"]
            // config["transformer"]["encoder_head"]
        )
        d_model = config["transformer"]["encoder_hidden"]
        d_inner = config["transformer"]["conv_filter_size"]
        kernel_size = config["transformer"]["conv_kernel_size"]
        dropout = config["transformer"]["encoder_dropout"]

        self.max_seq_len = config["max_seq_len"]
        self.d_model = d_model

        self.src_word_emb = nn.Embedding(n_src_vocab, d_word_vec, padding_idx=PAD)
        self.position_enc = nn.Parameter(
            get_sinusoid_encoding_table(n_position, d_word_vec).unsqueeze(0),
            requires_grad=False,
        )

        self.layer_stack = nn.ModuleList(
            [
                FFTBlock(
                    d_model, n_head, d_k, d_v, d_inner, kernel_size, dropout=dropout
                )
                for _ in range(n_layers)
            ]
        )

    def forward(self, src_seq, mask, return_attns=False):
        enc_slf_attn_list = []
        batch_size, max_len = src_seq.shape[0], src_seq.shape[1]

        # -- Prepare masks
        slf_attn_mask = mask.unsqueeze(1).expand(-1, max_len, -1)

        # -- Forward
        if not self.training and src_seq.shape[1] > self.max_seq_len:
            enc_output = self.src_word_emb(src_seq) + get_sinusoid_encoding_table(
                src_seq.shape[1], self.d_model
            )[: src_seq.shape[1], :].unsqueeze(0).expand(batch_size, -1, -1).to(
                src_seq.device
            )
        else:
            enc_output = self.src_word_emb(src_seq) + self.position_enc[
                :, :max_len, :
            ].expand(batch_size, -1, -1)

        for enc_layer in self.layer_stack:
            enc_output, enc_slf_attn = enc_layer(
                enc_output, mask=mask, slf_attn_mask=slf_attn_mask
            )
            if return_attns:
                enc_slf_attn_list += [enc_slf_attn]

        return enc_output


class Decoder(nn.Module):
    """Decoder"""

    def __init__(self, config):
        super(Decoder, self).__init__()

        n_position = config["max_seq_len"] + 1
        d_word_vec = config["transformer"]["decoder_hidden"]
        n_layers = config["transformer"]["decoder_layer"]
        n_head = config["transformer"]["decoder_head"]
        d_k = d_v = (
            config["transformer"]["decoder_hidden"]
            // config["transformer"]["decoder_head"]
        )
        d_model = config["transformer"]["decoder_hidden"]
        d_inner = config["transformer"]["conv_filter_size"]
        kernel_size = config["transformer"]["conv_kernel_size"]
        dropout = config["transformer"]["decoder_dropout"]

        self.max_seq_len = config["max_seq_len"]
        self.d_model = d_model

        self.position_enc = nn.Parameter(
            get_sinusoid_encoding_table(n_position, d_word_vec).unsqueeze(0),
            requires_grad=False,
        )

        self.layer_stack = nn.ModuleList(
            [
                FFTBlock(
                    d_model, n_head, d_k, d_v, d_inner, kernel_size, dropout=dropout
                )
                for _ in range(n_layers)
            ]
        )

    def forward(self, enc_seq, mask, return_attns=False):
        dec_slf_attn_list = []
        batch_size, max_len = enc_seq.shape[0], enc_seq.shape[1]

        # -- Forward
        if not self.training and enc_seq.shape[1] > self.max_seq_len:
            # -- Prepare masks
            slf_attn_mask = mask.unsqueeze(1).expand(-1, max_len, -1)
            dec_output = enc_seq + get_sinusoid_encoding_table(
                enc_seq.shape[1], self.d_model
            )[: enc_seq.shape[1], :].unsqueeze(0).expand(batch_size, -1, -1).to(
                enc_seq.device
            )
        else:
            max_len = min(max_len, self.max_seq_len)

            # -- Prepare masks
            slf_attn_mask = mask.unsqueeze(1).expand(-1, max_len, -1)
            dec_output = enc_seq[:, :max_len, :] + self.position_enc[
                :, :max_len, :
            ].expand(batch_size, -1, -1)
            mask = mask[:, :max_len]
            slf_attn_mask = slf_attn_mask[:, :, :max_len]

        for dec_layer in self.layer_stack:
            dec_output, dec_slf_attn = dec_layer(
                dec_output, mask=mask, slf_attn_mask=slf_attn_mask
            )
            if return_attns:
                dec_slf_attn_list += [dec_slf_attn]

        return dec_output, mask