Spaces:
Sleeping
Sleeping
File size: 18,526 Bytes
6b7b403 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 | """Minimal decoder-only GPT for MIDI token LM (interpretability-friendly).
Architecture (Pre-LN, GPT-2 style):
tok_emb + pos_emb
→ repeat: x += attn(LN(x)); x += mlp(LN(x))
→ LN → logits
Causal self-attention is implemented explicitly so attention weights can be
returned for probing (``return_attn=True``).
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import List, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from tokenizer import (
ID2TOKEN,
N_POS_BINS,
PITCH_MAX,
PITCH_MIN,
VOCAB_SIZE,
)
SCALE_DEGREE_NONE = 12 # sentinel for "no key context / not a pitch"
@dataclass
class GPTConfig:
vocab_size: int = VOCAB_SIZE
block_size: int = 1024
d_model: int = 512
n_layers: int = 6
n_heads: int = 8
d_ff: int = 2048
dropout: float = 0.1
# Compound-embedding axes (additive on top of token embedding).
use_pitch_class_embed: bool = True # adds 13 (pc 0..11 + sentinel)
use_octave_embed: bool = True # adds 9 (oct 0..7 + sentinel)
use_interval_embed: bool = True # adds 27 (-13..13 + sentinel) for melodic interval
use_beat_cyclic_embed: bool = True # adds N_POS_BINS+1 for beat-within-bar
use_scale_degree_embed: bool = True # adds 13 (chromatic 0..11 + sentinel) relative to current key
def default_gpt_config() -> GPTConfig:
"""Recommended starter config (~10M params with weight tying)."""
return GPTConfig()
# --- Static per-token feature lookups -----------------------------------------
# Sentinel index for "no pitch class / octave applies."
PC_NONE = 12
OCT_NONE = 8
def _build_token_pitch_feature_tables() -> Tuple[torch.Tensor, torch.Tensor]:
"""Per-token-id buffers giving (pitch_class, octave) for pitch tokens
and sentinels otherwise.
"""
pc = torch.full((VOCAB_SIZE,), PC_NONE, dtype=torch.long)
oct_ = torch.full((VOCAB_SIZE,), OCT_NONE, dtype=torch.long)
for tid, name in ID2TOKEN.items():
if (
name.startswith("P")
and not name.startswith("POS")
and name[1:].isdigit()
):
midi = int(name[1:])
if PITCH_MIN <= midi <= PITCH_MAX:
pc[tid] = midi % 12
oct_[tid] = max(0, min(7, midi // 12 - 1))
return pc, oct_
def _is_pitch_token_mask() -> torch.Tensor:
"""Boolean mask of length VOCAB_SIZE: True for pitch tokens."""
mask = torch.zeros(VOCAB_SIZE, dtype=torch.bool)
for tid, name in ID2TOKEN.items():
if (
name.startswith("P")
and not name.startswith("POS")
and name[1:].isdigit()
):
mask[tid] = True
return mask
def _midi_for_pitch_token(tid: int) -> int:
"""MIDI number for a pitch token id, or -1 if not a pitch token."""
name = ID2TOKEN.get(tid, "")
if name.startswith("P") and not name.startswith("POS") and name[1:].isdigit():
return int(name[1:])
return -1
def _build_pitch_to_midi() -> torch.Tensor:
"""Per-token-id MIDI value for pitch tokens, -1 elsewhere."""
arr = torch.full((VOCAB_SIZE,), -1, dtype=torch.long)
for tid in range(VOCAB_SIZE):
arr[tid] = _midi_for_pitch_token(tid)
return arr
def _build_pos_token_value() -> torch.Tensor:
"""For each token id, the POS bin value if it's a POS token, else -1."""
arr = torch.full((VOCAB_SIZE,), -1, dtype=torch.long)
for tid, name in ID2TOKEN.items():
if name.startswith("POS") and name[3:].isdigit():
arr[tid] = int(name[3:])
return arr
def _build_key_token_root() -> torch.Tensor:
"""For each token id, the root pitch class for KEY tokens (0..11),
else -1. KEY_0..11 are major keys C..B; KEY_12..23 are minor keys C..B.
"""
arr = torch.full((VOCAB_SIZE,), -1, dtype=torch.long)
for tid, name in ID2TOKEN.items():
if name.startswith("KEY_") and name[4:].isdigit():
arr[tid] = int(name[4:]) % 12
return arr
# Interval embedding: clipped to [-13..13] with 27 = sentinel (no interval).
INTERVAL_RANGE = 13
INTERVAL_NONE = 2 * INTERVAL_RANGE + 1 # = 27
class CausalSelfAttention(nn.Module):
"""Multi-head causal self-attention with optional weight return."""
def __init__(
self,
d_model: int,
n_heads: int,
block_size: int,
dropout: float,
) -> None:
super().__init__()
if d_model % n_heads != 0:
raise ValueError("d_model must be divisible by n_heads")
self.d_model = d_model
self.n_heads = n_heads
self.head_dim = d_model // n_heads
self.scale = self.head_dim**-0.5
self.qkv = nn.Linear(d_model, 3 * d_model)
self.proj = nn.Linear(d_model, d_model)
self.attn_drop = nn.Dropout(dropout)
self.resid_drop = nn.Dropout(dropout)
def forward(
self,
x: torch.Tensor,
return_attn: bool = False,
use_cache: bool = False,
past_kv: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
) -> Tuple[
torch.Tensor,
Optional[torch.Tensor],
Optional[Tuple[torch.Tensor, torch.Tensor]],
]:
B, Tq, C = x.shape
qkv = self.qkv(x)
qkv = qkv.view(B, Tq, 3, self.n_heads, self.head_dim)
qkv = qkv.permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
if past_kv is not None:
past_k, past_v = past_kv
k = torch.cat([past_k, k], dim=2)
v = torch.cat([past_v, v], dim=2)
att = (q @ k.transpose(-2, -1)) * self.scale
Tk = k.size(2)
past_len = Tk - Tq
key_pos = torch.arange(Tk, device=x.device).unsqueeze(0)
query_pos = (
torch.arange(Tq, device=x.device).unsqueeze(1) + past_len
)
causal = key_pos <= query_pos
att = att.masked_fill(~causal.unsqueeze(0).unsqueeze(0), float("-inf"))
att_weights = F.softmax(att, dim=-1)
att_weights = self.attn_drop(att_weights)
out = att_weights @ v
out = out.transpose(1, 2).contiguous().view(B, Tq, C)
out = self.resid_drop(self.proj(out))
present = (k, v) if use_cache else None
if return_attn:
return out, att_weights, present
return out, None, present
class TransformerBlock(nn.Module):
def __init__(self, config: GPTConfig) -> None:
super().__init__()
self.ln1 = nn.LayerNorm(config.d_model)
self.attn = CausalSelfAttention(
d_model=config.d_model,
n_heads=config.n_heads,
block_size=config.block_size,
dropout=config.dropout,
)
self.ln2 = nn.LayerNorm(config.d_model)
self.mlp = nn.Sequential(
nn.Linear(config.d_model, config.d_ff),
nn.GELU(),
nn.Linear(config.d_ff, config.d_model),
nn.Dropout(config.dropout),
)
def forward(
self,
x: torch.Tensor,
return_attn: bool = False,
use_cache: bool = False,
past_kv: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
) -> Tuple[
torch.Tensor,
Optional[torch.Tensor],
Optional[Tuple[torch.Tensor, torch.Tensor]],
]:
h, attn_w, present = self.attn(
self.ln1(x),
return_attn=return_attn,
use_cache=use_cache,
past_kv=past_kv,
)
x = x + h
x = x + self.mlp(self.ln2(x))
return x, attn_w, present
class GPT(nn.Module):
"""Decoder-only transformer LM with optional attention outputs."""
def __init__(self, config: GPTConfig) -> None:
super().__init__()
self.config = config
self.wte = nn.Embedding(config.vocab_size, config.d_model)
self.wpe = nn.Embedding(config.block_size, config.d_model)
self.drop = nn.Dropout(config.dropout)
self.blocks = nn.ModuleList(
TransformerBlock(config) for _ in range(config.n_layers)
)
self.ln_f = nn.LayerNorm(config.d_model)
self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
self.lm_head.weight = self.wte.weight
# --- Compound (2D) embedding axes ------------------------------------
# Static per-token-id feature lookups (computed from tokenizer vocab).
pc_tab, oct_tab = _build_token_pitch_feature_tables()
self.register_buffer("tok_to_pc", pc_tab, persistent=False)
self.register_buffer("tok_to_octave", oct_tab, persistent=False)
self.register_buffer(
"tok_to_midi", _build_pitch_to_midi(), persistent=False
)
self.register_buffer(
"tok_to_pos_value", _build_pos_token_value(), persistent=False
)
self.register_buffer(
"tok_to_key_root", _build_key_token_root(), persistent=False
)
self.register_buffer(
"tok_is_pitch", _is_pitch_token_mask(), persistent=False
)
if config.use_pitch_class_embed:
self.embed_pc = nn.Embedding(PC_NONE + 1, config.d_model)
if config.use_octave_embed:
self.embed_octave = nn.Embedding(OCT_NONE + 1, config.d_model)
if config.use_interval_embed:
self.embed_interval = nn.Embedding(INTERVAL_NONE + 1, config.d_model)
if config.use_beat_cyclic_embed:
# N_POS_BINS bins + sentinel for "no bar context".
self.embed_beat = nn.Embedding(N_POS_BINS + 1, config.d_model)
if config.use_scale_degree_embed:
self.embed_scale_degree = nn.Embedding(SCALE_DEGREE_NONE + 1, config.d_model)
self.apply(self._init_weights)
@staticmethod
def _init_weights(module: nn.Module) -> None:
if isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
torch.nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(
self,
idx: Optional[torch.Tensor] = None,
inputs_embeds: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
return_attn: bool = False,
use_cache: bool = False,
past_key_values: Optional[
List[Tuple[torch.Tensor, torch.Tensor]]
] = None,
) -> Union[
torch.Tensor,
Tuple[torch.Tensor, List[torch.Tensor]],
Tuple[torch.Tensor, List[Tuple[torch.Tensor, torch.Tensor]]],
Tuple[
torch.Tensor,
List[torch.Tensor],
List[Tuple[torch.Tensor, torch.Tensor]],
],
]:
"""Compute logits for token inputs.
Provide exactly one of:
- ``idx``: token ids of shape (B, T), or
- ``inputs_embeds``: precomputed embeddings of shape (B, T, d_model)
If ``return_attn`` is True, also returns a list of attention weight
tensors, one per layer, each shaped (B, n_heads, T, T) after softmax.
"""
if (idx is None) == (inputs_embeds is None):
raise ValueError("Provide exactly one of idx or inputs_embeds.")
if inputs_embeds is not None:
B, T, C = inputs_embeds.shape
if C != self.config.d_model:
raise ValueError(
"inputs_embeds last dim "
f"{C} != d_model {self.config.d_model}"
)
else:
assert idx is not None
B, T = idx.shape
if T > self.config.block_size:
raise ValueError(
f"Sequence length {T} exceeds block_size {self.config.block_size}"
)
if position_ids is None:
if idx is not None:
device = idx.device
else:
assert inputs_embeds is not None
device = inputs_embeds.device
pos = torch.arange(0, T, device=device, dtype=torch.long)
pos_e = self.wpe(pos).unsqueeze(0)
else:
if position_ids.shape[-1] != T:
raise ValueError(
"position_ids length must match sequence length."
)
pos_e = self.wpe(position_ids)
if pos_e.dim() == 2:
pos_e = pos_e.unsqueeze(0)
if pos_e.shape[0] == 1 and B > 1:
pos_e = pos_e.expand(B, -1, -1)
tok = self.wte(idx) if idx is not None else inputs_embeds
x = tok + pos_e
if idx is not None:
x = x + self._compound_embeds(idx)
x = self.drop(x)
attn_list: List[torch.Tensor] = []
present_key_values: List[Tuple[torch.Tensor, torch.Tensor]] = []
for block in self.blocks:
block_idx = len(present_key_values)
past_kv = None
if past_key_values is not None and block_idx < len(past_key_values):
past_kv = past_key_values[block_idx]
x, aw, present = block(
x,
return_attn=return_attn,
use_cache=use_cache,
past_kv=past_kv,
)
if aw is not None:
attn_list.append(aw)
if present is not None:
present_key_values.append(present)
x = self.ln_f(x)
logits = self.lm_head(x)
if return_attn and use_cache:
return logits, attn_list, present_key_values
if return_attn:
return logits, attn_list
if use_cache:
return logits, present_key_values
return logits
def _compound_embeds(self, idx: torch.Tensor) -> torch.Tensor:
"""Sum of all enabled compound (2D) embedding axes for a batch of
token ids. Returns a tensor of shape (B, T, d_model) — zero if all
axes are disabled."""
B, T = idx.shape
out = torch.zeros(B, T, self.config.d_model, device=idx.device, dtype=self.wte.weight.dtype)
if self.config.use_pitch_class_embed:
out = out + self.embed_pc(self.tok_to_pc[idx])
if self.config.use_octave_embed:
out = out + self.embed_octave(self.tok_to_octave[idx])
if self.config.use_interval_embed:
out = out + self.embed_interval(self._compute_interval_ids(idx))
if self.config.use_beat_cyclic_embed:
out = out + self.embed_beat(self._compute_beat_ids(idx))
if self.config.use_scale_degree_embed:
out = out + self.embed_scale_degree(self._compute_scale_degree_ids(idx))
return out
def _compute_scale_degree_ids(self, idx: torch.Tensor) -> torch.Tensor:
"""For each pitch position, the chromatic scale degree
(pitch_class - current_key_root) mod 12 — where current key is the
most recent KEY token seen. Non-pitch positions and positions
before the first KEY token get the sentinel.
"""
B, T = idx.shape
pc = self.tok_to_pc[idx] # (B, T) PC_NONE if not pitch
key_root = self.tok_to_key_root[idx] # (B, T) -1 if not KEY
arange = torch.arange(T, device=idx.device).expand(B, T)
cand = torch.where(key_root >= 0, arange, torch.full_like(arange, -1))
last_key_idx = cand.cummax(dim=1).values
safe_idx = last_key_idx.clamp(min=0)
cur_root = torch.gather(key_root, 1, safe_idx)
# Compute (pc - root) mod 12 for pitch positions with a known key.
is_pitch = pc != PC_NONE
sd = (pc - cur_root) % 12
valid = is_pitch & (last_key_idx >= 0)
return torch.where(
valid, sd, torch.full_like(sd, SCALE_DEGREE_NONE)
)
def _compute_interval_ids(self, idx: torch.Tensor) -> torch.Tensor:
"""For each pitch-token position, the clipped melodic interval to the
previous pitch token in the same row. Non-pitch positions and the
first pitch get the sentinel INTERVAL_NONE. Vectorized via cummax.
"""
B, T = idx.shape
midi = self.tok_to_midi[idx] # (B, T) -1 if not pitch
is_pitch = midi >= 0 # (B, T)
arange = torch.arange(T, device=idx.device).expand(B, T)
cand = torch.where(is_pitch, arange, torch.full_like(arange, -1))
# Shift right by 1: previous-pitch-up-to-t-1
shifted = torch.cat(
[torch.full_like(cand[:, :1], -1), cand[:, :-1]], dim=1
)
last_idx = shifted.cummax(dim=1).values # (B, T)
safe_idx = last_idx.clamp(min=0)
prev_midi = torch.gather(midi, 1, safe_idx)
delta = (midi - prev_midi).clamp(-INTERVAL_RANGE, INTERVAL_RANGE) + INTERVAL_RANGE
valid = is_pitch & (last_idx >= 0)
return torch.where(valid, delta, torch.full_like(delta, INTERVAL_NONE))
def _compute_beat_ids(self, idx: torch.Tensor) -> torch.Tensor:
"""For each position, the most recent POS<n> bin value seen so far,
or N_POS_BINS (sentinel) if no POS token has been emitted yet.
Vectorized via cummax over POS positions."""
B, T = idx.shape
pos_val = self.tok_to_pos_value[idx] # (B, T) -1 if not POS
arange = torch.arange(T, device=idx.device).expand(B, T)
cand = torch.where(pos_val >= 0, arange, torch.full_like(arange, -1))
last_idx = cand.cummax(dim=1).values # (B, T)
safe_idx = last_idx.clamp(min=0)
gathered = torch.gather(pos_val, 1, safe_idx)
return torch.where(
last_idx >= 0, gathered, torch.full_like(gathered, N_POS_BINS)
)
@torch.no_grad()
def count_parameters(self) -> int:
return sum(p.numel() for p in self.parameters())
if __name__ == "__main__":
cfg = default_gpt_config()
model = GPT(cfg)
n_params = model.count_parameters()
print(f"Config: {cfg}")
print(f"Parameter count: {n_params:,} (~{n_params / 1e6:.2f}M)")
x = torch.randint(0, cfg.vocab_size, (2, min(64, cfg.block_size)))
logits = model(x)
assert logits.shape == (2, x.shape[1], cfg.vocab_size)
logits2, attn = model(x, return_attn=True)
assert len(attn) == cfg.n_layers
assert attn[0].shape == (2, cfg.n_heads, x.shape[1], x.shape[1])
print("Forward + return_attn smoke test OK.")
|