File size: 10,957 Bytes
f2e6b6d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 | """Music Transformer with relative attention for chord generation.
Architecture: Transformer decoder (autoregressive) with relative position
encoding (Shaw et al. 2018, efficient skewing from Huang et al. 2018).
Default config (~25M params):
d_model=512, n_heads=8, d_ff=2048, n_layers=8
"""
from __future__ import annotations
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class RelativeMultiHeadAttention(nn.Module):
"""Multi-head self-attention with relative position bias."""
def __init__(
self,
d_model: int,
n_heads: int,
max_seq_len: int,
dropout: float = 0.1,
) -> None:
super().__init__()
assert d_model % n_heads == 0
self.n_heads = n_heads
self.d_k = d_model // n_heads
self.scale = math.sqrt(self.d_k)
self.w_q = nn.Linear(d_model, d_model)
self.w_k = nn.Linear(d_model, d_model)
self.w_v = nn.Linear(d_model, d_model)
self.w_o = nn.Linear(d_model, d_model)
# Learnable relative position embeddings: positions in [-max_len+1, max_len-1]
self.max_seq_len = max_seq_len
self.rel_emb = nn.Embedding(2 * max_seq_len - 1, self.d_k)
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor, mask: torch.Tensor | None = None) -> torch.Tensor:
"""
Args:
x: (B, L, D)
mask: (L, L) bool — True = masked (don't attend)
Returns:
(B, L, D)
"""
B, L, _ = x.shape
H, dk = self.n_heads, self.d_k
Q = self.w_q(x).view(B, L, H, dk).transpose(1, 2) # (B, H, L, dk)
K = self.w_k(x).view(B, L, H, dk).transpose(1, 2)
V = self.w_v(x).view(B, L, H, dk).transpose(1, 2)
# Content attention: Q K^T
content = torch.matmul(Q, K.transpose(-2, -1)) # (B, H, L, L)
# Relative position attention: Q R^T via efficient gather
rel = self._relative_attention(Q, L) # (B, H, L, L)
attn = (content + rel) / self.scale
if mask is not None:
attn = attn.masked_fill(mask.unsqueeze(0).unsqueeze(0), float("-inf"))
attn = self.dropout(F.softmax(attn, dim=-1))
out = torch.matmul(attn, V) # (B, H, L, dk)
out = out.transpose(1, 2).contiguous().view(B, L, -1)
return self.w_o(out)
def _relative_attention(self, Q: torch.Tensor, L: int) -> torch.Tensor:
"""Compute Q @ R^T using relative position embeddings.
Uses the index-gather approach: for each (i, j) pair, the relative
position is j - i, shifted to a non-negative index.
"""
device = Q.device
# Relative position indices: rel[i,j] = j - i + max_seq_len - 1
positions = torch.arange(L, device=device)
rel_idx = positions.unsqueeze(0) - positions.unsqueeze(1) + self.max_seq_len - 1
rel_idx = rel_idx.clamp(0, 2 * self.max_seq_len - 2)
R = self.rel_emb(rel_idx) # (L, L, dk)
# Q: (B, H, L, dk) R: (L, L, dk) → need (B, H, L, L)
# Reshape Q to (B*H, L, dk), bmm with R^T reshaped
BH = Q.shape[0] * Q.shape[1]
Q_flat = Q.reshape(BH, L, self.d_k) # (BH, L, dk)
# For each query position i, we want dot(Q[i], R[i, :, :]) → (BH, L, L)
# R: (L, L, dk) → transpose last two → (L, dk, L)
# Then Q_flat[:, i, :] @ R[i, :, :].T for each i
# Efficient: einsum
rel_score = torch.einsum("bld,lsd->bls", Q_flat, R) # (BH, L, L)
return rel_score.view(Q.shape[0], Q.shape[1], L, L)
class TransformerBlock(nn.Module):
"""Pre-norm Transformer decoder block."""
def __init__(
self,
d_model: int,
n_heads: int,
d_ff: int,
max_seq_len: int,
dropout: float = 0.1,
) -> None:
super().__init__()
self.norm1 = nn.LayerNorm(d_model)
self.attn = RelativeMultiHeadAttention(d_model, n_heads, max_seq_len, dropout)
self.norm2 = nn.LayerNorm(d_model)
self.ffn = nn.Sequential(
nn.Linear(d_model, d_ff),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(d_ff, d_model),
nn.Dropout(dropout),
)
self.drop = nn.Dropout(dropout)
def forward(self, x: torch.Tensor, mask: torch.Tensor | None = None) -> torch.Tensor:
x = x + self.drop(self.attn(self.norm1(x), mask))
x = x + self.ffn(self.norm2(x))
return x
class MusicTransformer(nn.Module):
"""Autoregressive Music Transformer for chord generation."""
def __init__(
self,
vocab_size: int,
d_model: int = 512,
n_heads: int = 8,
d_ff: int = 2048,
n_layers: int = 8,
max_seq_len: int = 512,
dropout: float = 0.1,
pad_id: int = 0,
) -> None:
super().__init__()
self.d_model = d_model
self.max_seq_len = max_seq_len
self.pad_id = pad_id
self.token_emb = nn.Embedding(vocab_size, d_model, padding_idx=pad_id)
self.drop = nn.Dropout(dropout)
self.layers = nn.ModuleList([
TransformerBlock(d_model, n_heads, d_ff, max_seq_len, dropout)
for _ in range(n_layers)
])
self.norm = nn.LayerNorm(d_model)
self.out_proj = nn.Linear(d_model, vocab_size, bias=False)
# Weight tying (embedding ↔ output projection)
self.out_proj.weight = self.token_emb.weight
self._init_weights()
def _init_weights(self) -> None:
for name, p in self.named_parameters():
if p.dim() > 1 and "token_emb" not in name:
nn.init.xavier_uniform_(p)
# Embedding std=1/sqrt(d_model) so that after *sqrt(d_model) scaling
# inputs have unit variance, and weight-tied output logits stay small
nn.init.normal_(self.token_emb.weight, mean=0.0, std=self.d_model ** -0.5)
@staticmethod
def _causal_mask(L: int, device: torch.device) -> torch.Tensor:
"""Upper-triangular causal mask (True = masked)."""
return torch.triu(torch.ones(L, L, device=device, dtype=torch.bool), diagonal=1)
def forward(self, input_ids: torch.Tensor) -> torch.Tensor:
"""
Args:
input_ids: (B, L) token IDs
Returns:
logits: (B, L, vocab_size)
"""
B, L = input_ids.shape
x = self.token_emb(input_ids) * math.sqrt(self.d_model)
x = self.drop(x)
mask = self._causal_mask(L, input_ids.device)
for layer in self.layers:
x = layer(x, mask)
return self.out_proj(self.norm(x))
def count_parameters(self) -> int:
return sum(p.numel() for p in self.parameters() if p.requires_grad)
@torch.no_grad()
def generate(
self,
prompt_ids: torch.Tensor,
max_new_tokens: int = 64,
temperature: float = 1.0,
top_k: int = 0,
top_p: float = 0.9,
eos_id: int = 2,
repetition_penalty: float = 1.0,
no_repeat_ngram_size: int = 0,
ignore_repeat_token_ids: set[int] | None = None,
) -> torch.Tensor:
"""Autoregressive generation from a prompt.
Args:
prompt_ids: (1, L) token IDs including [BOS] and context.
max_new_tokens: maximum tokens to generate.
temperature: sampling temperature (lower = more deterministic).
top_k: keep only top-k logits (0 = disabled).
top_p: nucleus sampling threshold.
eos_id: stop token.
repetition_penalty: divide logits of previously-seen tokens by
this factor (HF convention). > 1.0 discourages repeats.
1.0 disables. Typical: 1.2–1.5.
no_repeat_ngram_size: ban candidate tokens that would complete
an n-gram already present in the current sequence (n =
this value). 0 disables. Typical: 3 for chord sequences.
ignore_repeat_token_ids: token ids exempt from the two repetition
controls above — e.g. [BAR] or other separators that
*should* recur. If None, no exemptions.
Returns:
(1, L') full sequence including prompt and generated tokens.
"""
self.eval()
ids = prompt_ids.clone()
exempt = ignore_repeat_token_ids or set()
for _ in range(max_new_tokens):
ctx = ids[:, -self.max_seq_len :]
logits = self(ctx)[:, -1, :] / max(temperature, 1e-8)
# Repetition penalty (HuggingFace-style): scale already-seen token
# logits so they are less attractive. Positive logits get divided,
# negative logits get multiplied (stays "less attractive" either sign).
if repetition_penalty != 1.0:
seen = set(ids[0].tolist()) - exempt
if seen:
idx = torch.tensor(list(seen), device=logits.device, dtype=torch.long)
vals = logits[0, idx]
vals = torch.where(
vals > 0,
vals / repetition_penalty,
vals * repetition_penalty,
)
logits[0, idx] = vals
# No-repeat n-gram: block any candidate token that would complete
# an n-gram already present earlier in the sequence.
if no_repeat_ngram_size > 0 and ids.shape[1] >= no_repeat_ngram_size:
n = no_repeat_ngram_size
seq = ids[0].tolist()
prefix = tuple(seq[-(n - 1):]) if n > 1 else ()
banned: set[int] = set()
for i in range(len(seq) - n + 1):
if tuple(seq[i : i + n - 1]) == prefix:
banned.add(seq[i + n - 1])
banned -= exempt
if banned:
bidx = torch.tensor(list(banned), device=logits.device, dtype=torch.long)
logits[0, bidx] = float("-inf")
# Top-k
if top_k > 0:
topk_vals, _ = torch.topk(logits, min(top_k, logits.size(-1)))
logits[logits < topk_vals[:, -1:]] = float("-inf")
# Top-p (nucleus)
if 0 < top_p < 1.0:
sorted_logits, sorted_idx = torch.sort(logits, descending=True)
cum_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
remove = cum_probs - F.softmax(sorted_logits, dim=-1) > top_p
sorted_logits[remove] = float("-inf")
logits = sorted_logits.scatter(1, sorted_idx, sorted_logits)
probs = F.softmax(logits, dim=-1)
next_id = torch.multinomial(probs, num_samples=1)
ids = torch.cat([ids, next_id], dim=-1)
if (next_id == eos_id).all():
break
return ids
|