File size: 4,883 Bytes
9c7e38e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Transformer model for multi-instrument grid-based music generation.
"""

from typing import Dict

import torch
import torch.nn as nn
import torch.nn.functional as F

import config


class PositionalEncoding(nn.Module):
    """
    Standard sinusoidal positional encoding for Transformer inputs.
    Expects input of shape [batch, seq_len, d_model].
    """

    def __init__(self, d_model: int, max_len: int = 2048):
        super().__init__()
        self.d_model = d_model

        pe = torch.zeros(max_len, d_model)  # [max_len, d_model]
        position = torch.arange(0, max_len, dtype=torch.float32).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, d_model, 2, dtype=torch.float32)
            * (-torch.log(torch.tensor(10000.0)) / d_model)
        )

        pe[:, 0::2] = torch.sin(position * div_term)  # even indices
        pe[:, 1::2] = torch.cos(position * div_term)  # odd indices

        pe = pe.unsqueeze(0)  # [1, max_len, d_model]
        self.register_buffer("pe", pe)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Add positional encoding to input.

        Args:
            x: [batch, seq_len, d_model]

        Returns:
            [batch, seq_len, d_model]
        """
        seq_len = x.size(1)
        # pe[:, :seq_len] has shape [1, seq_len, d_model]
        return x + self.pe[:, :seq_len]


class MusicTransformer(nn.Module):
    """
    Transformer encoder with per-instrument output heads.

    Input:
        x: [batch, seq_len, INPUT_FEATURES] (multi-hot grid)

    Output:
        dict of logits:
            {
                "strings":  [batch, seq_len, 12],
                "woodwinds":[batch, seq_len, 12],
                "brass":    [batch, seq_len, 12],
                "perc":     [batch, seq_len, 3],
            }
    """

    def __init__(
        self,
        input_dim: int = config.INPUT_FEATURES,
        d_model: int = config.D_MODEL,
        n_layers: int = config.N_LAYERS,
        n_heads: int = config.N_HEADS,
        ffn_dim: int = config.FFN_DIM,
        dropout: float = config.DROPOUT,
    ):
        super().__init__()

        self.input_dim = input_dim
        self.d_model = d_model

        # Project 39-dim multi-hot input to d_model
        self.input_proj = nn.Linear(input_dim, d_model)

        # Positional encoding
        self.pos_encoding = PositionalEncoding(d_model=d_model)

        # Transformer encoder
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=d_model,
            nhead=n_heads,
            dim_feedforward=ffn_dim,
            dropout=dropout,
            batch_first=True,  # [batch, seq, d_model]
        )
        self.transformer_encoder = nn.TransformerEncoder(
            encoder_layer,
            num_layers=n_layers,
        )

        # Per-instrument heads
        self.strings_head = nn.Linear(d_model, config.PITCHES_MELODIC)  # 12
        self.woodwinds_head = nn.Linear(d_model, config.PITCHES_MELODIC)  # 12
        self.brass_head = nn.Linear(d_model, config.PITCHES_MELODIC)  # 12
        self.perc_head = nn.Linear(d_model, config.PITCHES_PERC)  # 3

    def forward(self, x: torch.Tensor) -> Dict[str, torch.Tensor]:
        """
        Forward pass.

        Args:
            x: Tensor of shape [batch, seq_len, INPUT_FEATURES]

        Returns:
            Dict of logits, each [batch, seq_len, pitches]:
                {
                    "strings": logits_S,
                    "woodwinds": logits_W,
                    "brass": logits_B,
                    "perc": logits_P,
                }
        """
        # Project input features to d_model
        h = self.input_proj(x)  # [batch, seq_len, d_model]

        # Add positional encoding
        h = self.pos_encoding(h)  # [batch, seq_len, d_model]

        # Transformer encoder
        # (no src_key_padding_mask here; you can add later if needed)
        h = self.transformer_encoder(h)  # [batch, seq_len, d_model]

        # Per-instrument logits
        logits_strings = self.strings_head(h)    # [batch, seq_len, 12]
        logits_woodwinds = self.woodwinds_head(h)
        logits_brass = self.brass_head(h)
        logits_perc = self.perc_head(h)          # [batch, seq_len, 3]

        return {
            "strings": logits_strings,
            "woodwinds": logits_woodwinds,
            "brass": logits_brass,
            "perc": logits_perc,
        }


def build_model(cfg=config) -> MusicTransformer:
    """
    Convenience function to build a MusicTransformer from config values.

    Returns:
        MusicTransformer instance with parameters from config.py.
    """
    model = MusicTransformer(
        input_dim=cfg.INPUT_FEATURES,
        d_model=cfg.D_MODEL,
        n_layers=cfg.N_LAYERS,
        n_heads=cfg.N_HEADS,
        ffn_dim=cfg.FFN_DIM,
        dropout=cfg.DROPOUT,
    )
    return model