Spaces:
Sleeping
Sleeping
File size: 12,493 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 | """Phrase completion evaluation on JSB chorales.
For each JSB piece, we build (prompt, ground_truth_continuation) pairs by
splitting near the midpoint BAR_END. We then compare:
- model completions
- random-token completions
- ground-truth Bach continuations
using tonal stability, rhythmic regularity, and phrase closure metrics.
"""
from __future__ import annotations
import argparse
import json
import random
import sys
import tempfile
from pathlib import Path
from typing import Any, Dict, List, Tuple
import matplotlib
import numpy as np
import pretty_midi
import torch
import torch.nn.functional as F
matplotlib.use("Agg")
import matplotlib.pyplot as plt # noqa: E402
_SCRIPT_DIR = Path(__file__).resolve().parent
_ROOT = _SCRIPT_DIR.parent
if str(_SCRIPT_DIR) not in sys.path:
sys.path.insert(0, str(_SCRIPT_DIR))
from model import GPT, GPTConfig, default_gpt_config # noqa: E402
from tokenizer import ( # noqa: E402
BAR_END,
ID2TOKEN,
PHRASE_END,
VOCAB_SIZE,
encode,
)
def _pick_device() -> torch.device:
if torch.cuda.is_available():
return torch.device("cuda")
mps = getattr(torch.backends, "mps", None)
if mps is not None and mps.is_available():
return torch.device("mps")
return torch.device("cpu")
def top_k_filter(logits: torch.Tensor, k: int) -> torch.Tensor:
if k <= 0 or k >= logits.size(-1):
return logits
values, _ = torch.topk(logits, k)
threshold = values[:, -1].unsqueeze(-1)
return logits.masked_fill(logits < threshold, float("-inf"))
def _extract_gpt_config_dict(raw: Dict[str, Any]) -> Dict[str, Any]:
keys = set(GPTConfig.__dataclass_fields__.keys())
return {k: raw[k] for k in keys if k in raw}
def _load_config_from_sources(
ckpt: Dict[str, Any], config_path: str
) -> GPTConfig:
cfg = default_gpt_config()
ckpt_cfg = ckpt.get("config")
if isinstance(ckpt_cfg, dict):
for k, v in _extract_gpt_config_dict(ckpt_cfg).items():
setattr(cfg, k, v)
if config_path:
loaded = json.loads(Path(config_path).read_text())
if not isinstance(loaded, dict):
raise ValueError("--config must be a JSON object.")
for k, v in _extract_gpt_config_dict(loaded).items():
setattr(cfg, k, v)
return cfg
def _pitch_class_from_id(tid: int) -> int | None:
tok = ID2TOKEN.get(tid, "")
if not tok.startswith("P"):
return None
try:
return int(tok[1:]) % 12
except ValueError:
return None
def _bar_lengths(ids: List[int]) -> List[int]:
lengths: List[int] = []
count = 0
in_bar = False
for tid in ids:
if tid == BAR_END:
if in_bar:
lengths.append(max(1, count))
in_bar = False
count = 0
continue
count += 1
tok = ID2TOKEN.get(tid, "")
if tok == "BAR_START":
in_bar = True
return lengths
def _infer_tonic(prompt_ids: List[int]) -> int:
pcs = [
pc for tid in prompt_ids if (pc := _pitch_class_from_id(tid)) is not None
]
if not pcs:
return 0
counts = np.bincount(np.array(pcs, dtype=np.int64), minlength=12)
return int(np.argmax(counts))
def _scale_pitch_classes(tonic: int) -> Tuple[set[int], set[int]]:
major = {0, 2, 4, 5, 7, 9, 11}
minor = {0, 2, 3, 5, 7, 8, 10}
maj = {(tonic + x) % 12 for x in major}
minr = {(tonic + x) % 12 for x in minor}
return maj, minr
def tonal_stability(prompt_ids: List[int], gen_ids: List[int]) -> float:
tonic = _infer_tonic(prompt_ids)
maj, minr = _scale_pitch_classes(tonic)
pcs = [pc for tid in gen_ids if (pc := _pitch_class_from_id(tid)) is not None]
if not pcs:
return 0.0
in_maj = sum(1 for pc in pcs if pc in maj) / len(pcs)
in_min = sum(1 for pc in pcs if pc in minr) / len(pcs)
return float(max(in_maj, in_min))
def rhythmic_regularity(prompt_ids: List[int], gen_ids: List[int]) -> float:
p_bars = _bar_lengths(prompt_ids)
g_bars = _bar_lengths(gen_ids)
if not p_bars or not g_bars:
return 0.0
p_len = float(np.mean(p_bars))
g_len = float(np.mean(g_bars))
if p_len <= 0:
return 0.0
score = 1.0 - abs(g_len - p_len) / p_len
return float(max(0.0, min(1.0, score)))
def phrase_closure(prompt_ids: List[int], gen_ids: List[int]) -> float:
if not gen_ids:
return 0.0
tail = gen_ids[-8:]
tonic = _infer_tonic(prompt_ids)
tonic_present = any(_pitch_class_from_id(tid) == tonic for tid in tail)
if PHRASE_END in tail or BAR_END in tail or tonic_present:
return 1.0
return 0.0
def _metrics(
prompt_ids: List[int], continuation_ids: List[int]
) -> Dict[str, float]:
return {
"tonal_stability": tonal_stability(prompt_ids, continuation_ids),
"rhythmic_regularity": rhythmic_regularity(prompt_ids, continuation_ids),
"phrase_closure": phrase_closure(prompt_ids, continuation_ids),
}
def _build_phrase_pairs(
n_pairs: int,
min_prompt_tokens: int,
min_cont_tokens: int,
seed: int,
) -> List[Tuple[List[int], List[int]]]:
from music21 import corpus
rng = random.Random(seed)
chorales = list(
corpus.chorales.Iterator(
numberingSystem="bwv",
returnType="stream",
)
)
rng.shuffle(chorales)
pairs: List[Tuple[List[int], List[int]]] = []
for score in chorales:
if len(pairs) >= n_pairs:
break
try:
with tempfile.NamedTemporaryFile(
suffix=".mid",
delete=True,
) as tmp:
score.write("midi", fp=tmp.name)
pm = pretty_midi.PrettyMIDI(tmp.name)
ids = encode(pm)
except Exception:
continue
bar_end_pos = [i for i, tid in enumerate(ids) if tid == BAR_END]
if len(bar_end_pos) < 2:
continue
split = bar_end_pos[len(bar_end_pos) // 2]
prompt = ids[:split + 1]
continuation = ids[split + 1:]
if (
len(prompt) < min_prompt_tokens
or len(continuation) < min_cont_tokens
):
continue
pairs.append((prompt, continuation))
return pairs
@torch.no_grad()
def _generate_continuation(
model: GPT,
prompt_ids: List[int],
n_tokens: int,
temperature: float,
top_k: int,
device: torch.device,
) -> List[int]:
x = torch.tensor([prompt_ids], dtype=torch.long, device=device)
out = x
for _ in range(n_tokens):
ctx = out[:, -model.config.block_size:]
logits = model(ctx)[:, -1, :] / temperature
logits = top_k_filter(logits, top_k)
probs = F.softmax(logits, dim=-1)
nxt = torch.multinomial(probs, num_samples=1)
out = torch.cat([out, nxt], dim=1)
full = out[0].tolist()
return full[len(prompt_ids):]
def _plot_metrics(
system_scores: Dict[str, Dict[str, float]], out_path: Path
) -> None:
metrics = ["tonal_stability", "rhythmic_regularity", "phrase_closure"]
systems = list(system_scores.keys())
x = np.arange(len(metrics))
width = 0.8 / len(systems)
fig, ax = plt.subplots(figsize=(8.0, 4.6))
for idx, name in enumerate(systems):
offs = x - 0.4 + width * (idx + 0.5)
vals = [system_scores[name][m] for m in metrics]
ax.bar(offs, vals, width=width, label=name)
ax.set_xticks(x)
ax.set_xticklabels(["tonal", "rhythmic", "closure"])
ax.set_ylim(0.0, 1.05)
ax.set_ylabel("score")
ax.set_title("Phrase completion metrics")
ax.grid(alpha=0.25, axis="y")
ax.legend()
fig.tight_layout()
out_path.parent.mkdir(parents=True, exist_ok=True)
fig.savefig(out_path, dpi=160)
plt.close(fig)
def _write_summary(
out_path: Path,
n_pairs: int,
n_samples: int,
system_scores: Dict[str, Dict[str, float]],
) -> None:
rows = [
"# Phrase completion evaluation",
"",
f"- Pairs evaluated: {n_pairs}",
f"- Model samples per prompt: {n_samples}",
"",
"| system | tonal_stability | rhythmic_regularity | phrase_closure |",
"|---|---:|---:|---:|",
]
for name, scores in system_scores.items():
rows.append(
f"| {name} | {scores['tonal_stability']:.3f} | "
f"{scores['rhythmic_regularity']:.3f} | {scores['phrase_closure']:.3f} |"
)
out_path.parent.mkdir(parents=True, exist_ok=True)
out_path.write_text("\n".join(rows))
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser(
description="JSB phrase completion evaluation."
)
p.add_argument(
"--checkpoint",
type=str,
default=str(_ROOT / "results" / "checkpoints" / "best_model.pt"),
)
p.add_argument("--config", type=str, default="")
p.add_argument("--n-pairs", type=int, default=60)
p.add_argument("--samples-per-prompt", type=int, default=5)
p.add_argument("--temperature", type=float, default=1.0)
p.add_argument("--top-k", type=int, default=40)
p.add_argument("--seed", type=int, default=42)
p.add_argument(
"--plot-out",
type=str,
default=str(_ROOT / "figures" / "phrase_completion_metrics.png"),
)
p.add_argument(
"--summary-out",
type=str,
default=str(_ROOT / "results" / "phrase_completion_summary.md"),
)
return p.parse_args()
@torch.no_grad()
def main() -> None:
args = parse_args()
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
device = _pick_device()
ckpt = torch.load(args.checkpoint, map_location=device, weights_only=True)
cfg = _load_config_from_sources(ckpt, args.config)
model = GPT(cfg).to(device)
state = (
ckpt["model"] if isinstance(ckpt, dict) and "model" in ckpt else ckpt
)
model.load_state_dict(state)
model.eval()
pairs = _build_phrase_pairs(
n_pairs=args.n_pairs,
min_prompt_tokens=32,
min_cont_tokens=32,
seed=args.seed,
)
if not pairs:
raise RuntimeError("No phrase pairs extracted from JSB.")
metric_names = ["tonal_stability", "rhythmic_regularity", "phrase_closure"]
systems: Dict[str, Dict[str, List[float]]] = {
"random": {k: [] for k in metric_names},
"model_t1.0_topk40": {k: [] for k in metric_names},
"ground_truth": {k: [] for k in metric_names},
}
rng = random.Random(args.seed + 123)
for prompt_ids, gt_cont in pairs:
target_len = len(gt_cont)
gt_scores = _metrics(prompt_ids, gt_cont)
for k, v in gt_scores.items():
systems["ground_truth"][k].append(v)
rand_ids = [rng.randrange(VOCAB_SIZE) for _ in range(target_len)]
rand_scores = _metrics(prompt_ids, rand_ids)
for k, v in rand_scores.items():
systems["random"][k].append(v)
for _ in range(args.samples_per_prompt):
gen_ids = _generate_continuation(
model=model,
prompt_ids=prompt_ids,
n_tokens=target_len,
temperature=args.temperature,
top_k=args.top_k,
device=device,
)
gen_scores = _metrics(prompt_ids, gen_ids)
for k, v in gen_scores.items():
systems["model_t1.0_topk40"][k].append(v)
system_means: Dict[str, Dict[str, float]] = {}
for name, metrics in systems.items():
system_means[name] = {
k: float(np.mean(v)) if v else 0.0 for k, v in metrics.items()
}
_plot_metrics(system_means, out_path=Path(args.plot_out))
_write_summary(
out_path=Path(args.summary_out),
n_pairs=len(pairs),
n_samples=args.samples_per_prompt,
system_scores=system_means,
)
print(
f"[probe_completion] pairs={len(pairs)} "
f"samples_per_prompt={args.samples_per_prompt}"
)
print(f"[probe_completion] plot -> {args.plot_out}")
print(f"[probe_completion] summary -> {args.summary_out}")
for system, scores in system_means.items():
print(
f"[probe_completion] {system}: "
f"tonal={scores['tonal_stability']:.3f} "
f"rhythmic={scores['rhythmic_regularity']:.3f} "
f"closure={scores['phrase_closure']:.3f}"
)
if __name__ == "__main__":
main()
|