File size: 7,056 Bytes
7116323
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
import torch
import torch.nn as nn
from torch.nn.modules.normalization import LayerNorm
import random

from utilities.constants import *
from utilities.device import get_device

from .positional_encoding import PositionalEncoding
from .rpr import TransformerEncoderRPR, TransformerEncoderLayerRPR


# MusicTransformer
class MusicTransformer(nn.Module):
    """
    ----------
    Author: Damon Gwinn
    ----------
    Music Transformer reproduction from https://arxiv.org/abs/1809.04281. Arguments allow for
    tweaking the transformer architecture (https://arxiv.org/abs/1706.03762) and the rpr argument
    toggles Relative Position Representations (RPR - https://arxiv.org/abs/1803.02155).

    Supports training and generation using Pytorch's nn.Transformer class with dummy decoder to
    make a decoder-only transformer architecture

    For RPR support, there is modified Pytorch 1.2.0 code in rpr.py. Modified source will be
    kept up to date with Pytorch revisions only as necessary.
    ----------
    """

    def __init__(self, n_layers=6, num_heads=8, d_model=512, dim_feedforward=1024,
                 dropout=0.1, max_sequence=2048, rpr=False):
        super(MusicTransformer, self).__init__()

        self.dummy      = DummyDecoder()

        self.nlayers    = n_layers
        self.nhead      = num_heads
        self.d_model    = d_model
        self.d_ff       = dim_feedforward
        self.dropout    = dropout
        self.max_seq    = max_sequence
        self.rpr        = rpr

        # Input embedding
        self.embedding = nn.Embedding(VOCAB_SIZE, self.d_model)

        # Positional encoding
        self.positional_encoding = PositionalEncoding(self.d_model, self.dropout, self.max_seq)

        # Base transformer
        if(not self.rpr):
            # To make a decoder-only transformer we need to use masked encoder layers
            # Dummy decoder to essentially just return the encoder output
            self.transformer = nn.Transformer(
                d_model=self.d_model, nhead=self.nhead, num_encoder_layers=self.nlayers,
                num_decoder_layers=0, dropout=self.dropout, # activation=self.ff_activ,
                dim_feedforward=self.d_ff, custom_decoder=self.dummy
            )
        # RPR Transformer
        else:
            encoder_norm = LayerNorm(self.d_model)
            encoder_layer = TransformerEncoderLayerRPR(self.d_model, self.nhead, self.d_ff, self.dropout, er_len=self.max_seq)
            encoder = TransformerEncoderRPR(encoder_layer, self.nlayers, encoder_norm)
            self.transformer = nn.Transformer(
                d_model=self.d_model, nhead=self.nhead, num_encoder_layers=self.nlayers,
                num_decoder_layers=0, dropout=self.dropout, # activation=self.ff_activ,
                dim_feedforward=self.d_ff, custom_decoder=self.dummy, custom_encoder=encoder
            )

        # Final output is a softmaxed linear layer
        self.Wout       = nn.Linear(self.d_model, VOCAB_SIZE)
        self.softmax    = nn.Softmax(dim=-1)

    # forward
    def forward(self, x, mask=True):
        """
        ----------
        Author: Damon Gwinn
        ----------
        Takes an input sequence and outputs predictions using a sequence to sequence method.

        A prediction at one index is the "next" prediction given all information seen previously.
        ----------
        """

        if(mask is True):
            mask = self.transformer.generate_square_subsequent_mask(x.shape[1]).to(get_device())
        else:
            mask = None

        x = self.embedding(x)

        # Input shape is (max_seq, batch_size, d_model)
        x = x.permute(1,0,2)

        x = self.positional_encoding(x)

        # Since there are no true decoder layers, the tgt is unused
        # Pytorch wants src and tgt to have some equal dims however
        x_out = self.transformer(src=x, tgt=x, src_mask=mask)

        # Back to (batch_size, max_seq, d_model)
        x_out = x_out.permute(1,0,2)

        y = self.Wout(x_out)
        # y = self.softmax(y)

        del mask

        # They are trained to predict the next note in sequence (we don't need the last one)
        return y

    # generate
    def generate(self, primer=None, target_seq_length=1024, beam=0, beam_chance=1.0):
        """
        ----------
        Author: Damon Gwinn
        ----------
        Generates midi given a primer sample. Music can be generated using a probability distribution over
        the softmax probabilities (recommended) or by using a beam search.
        ----------
        """

        assert (not self.training), "Cannot generate while in training mode"

        print("Generating sequence of max length:", target_seq_length)

        gen_seq = torch.full((1,target_seq_length), TOKEN_PAD, dtype=TORCH_LABEL_TYPE, device=get_device())

        num_primer = len(primer)
        gen_seq[..., :num_primer] = primer.type(TORCH_LABEL_TYPE).to(get_device())


        # print("primer:",primer)
        # print(gen_seq)
        cur_i = num_primer
        while(cur_i < target_seq_length):
            # gen_seq_batch     = gen_seq.clone()
            y = self.softmax(self.forward(gen_seq[..., :cur_i]))[..., :TOKEN_END]
            token_probs = y[:, cur_i-1, :]

            if(beam == 0):
                beam_ran = 2.0
            else:
                beam_ran = random.uniform(0,1)

            if(beam_ran <= beam_chance):
                token_probs = token_probs.flatten()
                top_res, top_i = torch.topk(token_probs, beam)

                beam_rows = top_i // VOCAB_SIZE
                beam_cols = top_i % VOCAB_SIZE

                gen_seq = gen_seq[beam_rows, :]
                gen_seq[..., cur_i] = beam_cols

            else:
                distrib = torch.distributions.categorical.Categorical(probs=token_probs)
                next_token = distrib.sample()
                # print("next token:",next_token)
                gen_seq[:, cur_i] = next_token


                # Let the transformer decide to end if it wants to
                if(next_token == TOKEN_END):
                    print("Model called end of sequence at:", cur_i, "/", target_seq_length)
                    break

            cur_i += 1
            if(cur_i % 50 == 0):
                print(cur_i, "/", target_seq_length)

        return gen_seq[:, :cur_i]

# Used as a dummy to nn.Transformer
# DummyDecoder
class DummyDecoder(nn.Module):
    """
    ----------
    Author: Damon Gwinn
    ----------
    A dummy decoder that returns its input. Used to make the Pytorch transformer into a decoder-only
    architecture (stacked encoders with dummy decoder fits the bill)
    ----------
    """

    def __init__(self):
        super(DummyDecoder, self).__init__()

    def forward(self, tgt, memory, tgt_mask, memory_mask,tgt_key_padding_mask,memory_key_padding_mask, **kwargs):
        """
        ----------
        Author: Damon Gwinn
        ----------
        Returns the input (memory)
        ----------
        """

        return memory