Spaces:
Sleeping
Sleeping
File size: 19,063 Bytes
b57c46e f2ee297 b57c46e 8089429 6e51e38 b57c46e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 | #!/usr/bin/env python3
"""
Gradio demo for LaM-SLidE Music Score Autoencoder.
Upload a MusicXML file (.mxl / .musicxml) or pick an example, choose a
model, and get back the reconstructed MusicXML plus per-feature accuracy
and a visual comparison of original vs. reconstructed scores.
"""
import shutil
import sys
import warnings
from pathlib import Path
# Add app root to path
sys.path.insert(0, str(Path(__file__).resolve().parent))
warnings.filterwarnings("ignore")
import os
import gradio as gr
import numpy as np
import torch
import verovio
from omegaconf import OmegaConf
VEROVIO_DATA_DIR = os.path.join(os.path.dirname(verovio.__file__), "data")
from src.model.autoencoder import create_autoencoder_from_dict
from inference import (
extract_features_from_graph,
reconstruct_from_graph,
undo_feature_shifts,
)
from reconstruct_mxl import (
reconstruct_score,
load_duration_vocabulary,
load_position_vocabulary,
)
from convert_mxl import mxl_to_graph_data, load_vocab_forward
# =============================================================================
# Paths
# =============================================================================
APP_DIR = Path(__file__).resolve().parent
VOCAB_DIR = APP_DIR / "vocabs"
MODELS_DIR = APP_DIR / "models"
EXAMPLES_DIR = APP_DIR / "examples"
TMP_DIR = APP_DIR / "tmp"
TMP_DIR.mkdir(exist_ok=True)
# Available models
MODELS = {
"Wide Large Factorized (best model)": "wide_large_factorized",
"HGT Wide Factorized (graph-aware)": "hgt_wide_factorized",
}
# Example .pt graph files (each may have a paired .mxl for original rendering)
# Sort by note count (extracted from filename pattern *_<N>notes_*) so the
# dropdown goes from smallest to largest score.
def _note_count_key(p: Path) -> int:
import re
m = re.search(r"_(\d+)notes", p.stem)
return int(m.group(1)) if m else 0
def _example_display_name(p: Path) -> str:
"""Format: '64 notes — Schubert: An die Laute (QmNcR2oAkq)'"""
import re
stem = p.stem
m = re.search(r"_(\d+)notes_(.+)$", stem)
if not m:
return stem
notes = m.group(1)
qm_hash = m.group(2)
# Everything before _<N>notes is the composer_title part
prefix = stem[:m.start()]
# Convert underscores to spaces and title-case
title = prefix.replace("_", " ").title()
return f"{notes} notes \u2014 {title} ({qm_hash})"
EXAMPLES = sorted(
(p for p in EXAMPLES_DIR.iterdir() if p.suffix == ".pt"),
key=_note_count_key,
) if EXAMPLES_DIR.exists() else []
# Map display name -> filename for dropdown
_EXAMPLE_DISPLAY = {_example_display_name(p): p.name for p in EXAMPLES}
_EXAMPLE_LOOKUP = {v: k for k, v in _EXAMPLE_DISPLAY.items()} # filename -> display
# File extensions recognised as MusicXML
MXL_EXTENSIONS = {".mxl", ".musicxml", ".xml"}
# Short display names for the accuracy table
FEATURE_SHORT_NAMES = {
"grid_position": "grid_pos",
"micro_offset": "micro",
"measure_idx": "bar",
"voice": "voice",
"pitch_step": "step",
"pitch_alter": "alter",
"pitch_octave": "oct",
"duration": "dur",
"clef": "clef",
"ts_beats": "ts_b",
"ts_beat_type": "ts_bt",
"key_fifths": "key",
"staff": "staff",
}
# =============================================================================
# Load vocabs once
# =============================================================================
duration_vocab_inv = load_duration_vocabulary(VOCAB_DIR / "duration_vocab.json")
grid_vocab_inv = load_position_vocabulary(VOCAB_DIR / "grid_vocab.json")
micro_vocab_inv = load_position_vocabulary(VOCAB_DIR / "micro_vocab.json")
# Forward vocabs for MXL → graph conversion (loaded from training data)
duration_vocab_fwd = load_vocab_forward(VOCAB_DIR / "duration_vocab.json")
grid_vocab_fwd = load_vocab_forward(VOCAB_DIR / "grid_vocab.json")
micro_vocab_fwd = load_vocab_forward(VOCAB_DIR / "micro_vocab.json")
# =============================================================================
# Model cache
# =============================================================================
_model_cache = {}
def get_model(model_key: str):
"""Load model (cached)."""
if model_key in _model_cache:
return _model_cache[model_key]
model_dir = MODELS_DIR / model_key
config_path = list(model_dir.glob("*.yaml"))[0]
checkpoint_path = model_dir / "best_model.pt"
cfg = OmegaConf.load(config_path)
config_dict = OmegaConf.to_container(cfg.model, resolve=True)
checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
model = create_autoencoder_from_dict(config_dict)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()
model.cpu()
_model_cache[model_key] = model
return model
# =============================================================================
# Score rendering with verovio
# =============================================================================
def render_score_to_svg_pages(
mxl_path: str, prefix: str,
page_width: int = 2100, scale: int = 35,
) -> list[str]:
"""Render a MusicXML file to one SVG file per page. Returns list of paths."""
tk = verovio.toolkit(False)
tk.setResourcePath(VEROVIO_DATA_DIR)
tk.loadFile(mxl_path)
tk.setOptions({
"pageWidth": page_width,
"scale": scale,
"adjustPageHeight": True,
"footer": "none",
"header": "none",
})
tk.redoLayout()
paths = []
for page in range(1, tk.getPageCount() + 1):
svg_path = f"{prefix}_p{page}.svg"
Path(svg_path).write_text(tk.renderToSVG(page), encoding="utf-8")
paths.append(svg_path)
return paths
# =============================================================================
# Core logic
# =============================================================================
FEATURE_KEY_MAP = {
"grid_position": "position_grid_token",
"micro_offset": "position_micro_token",
"duration": "duration_token",
"pitch_step": "pitch_step",
"pitch_alter": "pitch_alter",
"pitch_octave": "pitch_octave",
"measure_idx": "measure_idx",
"voice": "voice",
"staff": "staff",
"clef": "clef",
"ts_beats": "ts_beats",
"ts_beat_type": "ts_beat_type",
"key_fifths": "key_fifths",
}
def build_recon_features(gt_raw: dict, raw_predictions: dict) -> dict:
"""Build feature dict for reconstruct_score from raw ground-truth and predictions."""
recon_features = {}
for raw_key, tensor in gt_raw.items():
recon_features[raw_key] = tensor.numpy()
for model_key, output_key in FEATURE_KEY_MAP.items():
if model_key in raw_predictions:
val = raw_predictions[model_key]
recon_features[output_key] = val.numpy() if isinstance(val, torch.Tensor) else val
return recon_features
def score_to_mxl(score, path: Path) -> str:
"""Write a music21 score to MusicXML, return the path string."""
with warnings.catch_warnings():
warnings.simplefilter("ignore")
score.write("musicxml", fp=str(path))
return str(path)
def compute_per_feature_accuracy(
predictions: dict, ground_truth: dict, feature_names: list
) -> dict:
"""Compute per-feature token accuracy."""
accs = {}
for name in feature_names:
if name in predictions and name in ground_truth:
pred = predictions[name] if isinstance(predictions[name], torch.Tensor) else torch.tensor(predictions[name])
gt = ground_truth[name] if isinstance(ground_truth[name], torch.Tensor) else torch.tensor(ground_truth[name])
accs[name] = (pred == gt).float().mean().item()
return accs
def run_reconstruction(input_file, model_name: str):
"""
Main pipeline: load/convert input -> run model -> reconstruct MusicXML.
Accepts either a pre-processed score graph (.pt) or a MusicXML file
(.mxl / .musicxml / .xml). MusicXML files are converted on-the-fly.
Returns
-------
(gt_first, recon_first, gt_orig_pages, gt_recon_pages, recon_pages,
has_original_mxl, num_pages, mxl_path, accuracy_md, info_md)
"""
empty = (None, None, [], [], [], False, 1, None, "", "")
if input_file is None:
return empty
# Clean tmp directory
if TMP_DIR.exists():
shutil.rmtree(TMP_DIR)
TMP_DIR.mkdir(exist_ok=True)
model_key = MODELS[model_name]
model = get_model(model_key)
# Load or convert input
input_path = Path(input_file) if isinstance(input_file, str) else Path(input_file.name)
is_mxl = input_path.suffix.lower() in MXL_EXTENSIONS
original_mxl_path = str(input_path) if is_mxl else None
# For .pt graphs, check for a paired original .mxl alongside
if not is_mxl:
paired_mxl = input_path.with_suffix(".mxl")
if paired_mxl.exists():
original_mxl_path = str(paired_mxl)
if is_mxl:
try:
graph_data = mxl_to_graph_data(
str(input_path), duration_vocab_fwd,
grid_vocab_fwd, micro_vocab_fwd,
)
except Exception as e:
return (*empty[:8], f"**Conversion error**: {e}", "")
else:
graph_data = torch.load(input_path, map_location="cpu", weights_only=False)
num_notes = graph_data["num_notes"]
source = graph_data.get("source_file", "unknown")
# Clean source path: keep only from /mxl/ onwards, prefix with pdmx
if "/mxl/" in source:
source = "pdmx" + source[source.index("/mxl/"):]
feature_names = [f.name for f in model.config.input_features]
# Extract ground truth features
gt_features, entity_ids, gt_raw = extract_features_from_graph(
graph_data, feature_names,
identifier_pool_size=model.config.identifier_pool_size,
id_assignment="sequential",
)
# Run model
features_batch = {k: v.unsqueeze(0) for k, v in gt_features.items()}
entity_ids_batch = entity_ids.unsqueeze(0)
mask = torch.ones(1, num_notes, dtype=torch.bool)
kwargs = {}
if getattr(model.config, "use_hgt", False):
from src.model.note_hgt import NoteHGT
edge_dict = NoteHGT.extract_edge_dict(graph_data["graph"])
kwargs["edge_dicts"] = [edge_dict]
with torch.no_grad():
logits = model(features_batch, entity_ids_batch, mask=mask, **kwargs)
predictions = {
name: logits[name][0].argmax(dim=-1).cpu()
for name in logits.keys()
}
# ---- Accuracy ----
accs = compute_per_feature_accuracy(predictions, gt_features, feature_names)
core_names = [
"grid_position", "micro_offset", "measure_idx",
"pitch_step", "pitch_alter", "pitch_octave",
"duration", "staff", "voice",
]
core_correct = None
for name in core_names:
if name in predictions and name in gt_features:
match = (predictions[name] == gt_features[name])
core_correct = match if core_correct is None else (core_correct & match)
core_joint = core_correct.float().mean().item() if core_correct is not None else 0.0
all_correct = None
for name in feature_names:
if name in predictions and name in gt_features:
match = (predictions[name] == gt_features[name])
all_correct = match if all_correct is None else (all_correct & match)
all_joint = all_correct.float().mean().item() if all_correct is not None else 0.0
# Horizontal accuracy table
header_cells = " | ".join(
FEATURE_SHORT_NAMES.get(n, n) for n in feature_names if n in accs
)
value_cells = " | ".join(
f"{accs[n]*100:.1f}%" for n in feature_names if n in accs
)
accuracy_md = (
f"| {header_cells} | **core** | **all** |\n"
f"|{'---|' * (sum(1 for n in feature_names if n in accs) + 2)}\n"
f"| {value_cells} | **{core_joint*100:.1f}%** | **{all_joint*100:.1f}%** |"
)
# ---- Reconstruct MusicXML (predicted) ----
raw_predictions = undo_feature_shifts(predictions)
recon_features = build_recon_features(gt_raw, raw_predictions)
mxl_path = None
recon_pages: list[str] = []
try:
recon_score = reconstruct_score(
recon_features, grid_vocab_inv, micro_vocab_inv,
duration_vocab_inv, verbose=False,
)
mxl_path = score_to_mxl(recon_score, TMP_DIR / "reconstructed.musicxml")
recon_pages = render_score_to_svg_pages(
mxl_path, str(TMP_DIR / "reconstructed"),
)
except Exception as e:
accuracy_md += f"\n\n**Reconstruction error**: {e}"
# ---- Render ground truth: original MXL (if available) ----
gt_orig_pages: list[str] = []
if original_mxl_path is not None:
try:
gt_orig_pages = render_score_to_svg_pages(
original_mxl_path, str(TMP_DIR / "gt_original"),
)
except Exception:
pass
# ---- Render ground truth: reconstructed from features (always) ----
gt_recon_pages: list[str] = []
gt_raw_np = {k: v.numpy() if isinstance(v, torch.Tensor) else v for k, v in gt_raw.items()}
try:
gt_score = reconstruct_score(
gt_raw_np, grid_vocab_inv, micro_vocab_inv,
duration_vocab_inv, verbose=False,
)
gt_mxl_path = score_to_mxl(gt_score, TMP_DIR / "gt_reconstructed.musicxml")
gt_recon_pages = render_score_to_svg_pages(
gt_mxl_path, str(TMP_DIR / "gt_reconstructed"),
)
except Exception:
pass # ground-truth rendering is best-effort
has_original_mxl = bool(gt_orig_pages)
# Default: show original MXL if available, otherwise features-reconstructed
gt_pages = gt_orig_pages if has_original_mxl else gt_recon_pages
info_md = (
f"**Source**: {source} | "
f"**Notes**: {num_notes} | "
f"**Model**: {model_name}"
)
if is_mxl and graph_data.get("truncated", False):
info_md += (
f" | **Truncated** to {num_notes} notes "
f"({graph_data['total_bars']} bars) to fit model limit"
)
# Page count = max of the *active* GT variant and the reconstructed pages
num_pages = max(len(gt_pages), len(recon_pages), 1)
gt_first = gt_pages[0] if gt_pages else None
recon_first = recon_pages[0] if recon_pages else None
return (
gt_first, recon_first,
gt_orig_pages, gt_recon_pages, recon_pages,
has_original_mxl, num_pages,
mxl_path, accuracy_md, info_md,
)
# =============================================================================
# Gradio UI
# =============================================================================
def build_demo():
with gr.Blocks(
title="A Fixed-size Latent Space Autoencoder for Music Scores",
theme=gr.themes.Soft(),
) as demo:
gr.Markdown(
"# A Fixed-size Latent Space Autoencoder for Music Scores\n"
"Upload a MusicXML file (`.mxl` / `.musicxml`) or select an example from [PDMX](https://zenodo.org/records/15571083), "
"choose a model, and reconstruct a MusicXML file. Scores are rendered "
"with [Verovio](https://www.verovio.org).\n\n"
"*Companion demo for: A Fixed-size Latent Space Autoencoder for Music Scores "
"(Hendrik Roth, Emmanouil Karystinaios & Gerhard Widmer, JKU Linz) — "
"[GitHub](https://github.com/hendrik-roth/score-ae)*"
)
# Hidden state for page lists
gt_orig_pages_state = gr.State([])
gt_recon_pages_state = gr.State([])
recon_pages_state = gr.State([])
has_orig_mxl_state = gr.State(False)
with gr.Row(equal_height=False):
# ---- Left column: controls ----
with gr.Column(scale=1, min_width=280):
model_dropdown = gr.Dropdown(
choices=list(MODELS.keys()),
value=list(MODELS.keys())[0],
label="Model",
)
graph_input = gr.File(
label="Upload MusicXML (.mxl / .musicxml / .xml)",
file_types=[".mxl", ".musicxml", ".xml", ".pt"],
)
example_dropdown = gr.Dropdown(
choices=["(none)"] + list(_EXAMPLE_DISPLAY.keys()),
value="(none)",
label="Or pick an example from our PDMX test set (deduplicated, no licence conflict subset)",
)
run_btn = gr.Button("Reconstruct", variant="primary", size="lg")
info_output = gr.Markdown()
# ---- Right column: download, scores, accuracy ----
with gr.Column(scale=3):
with gr.Row():
mxl_output = gr.File(label="Download reconstructed MusicXML", scale=1, min_width=200)
gr.Column(scale=2) # spacer
with gr.Row():
page_slider = gr.Slider(
minimum=1, maximum=1, step=1, value=1,
label="Page", visible=False, scale=1,
)
gr.Column(scale=2) # spacer
gr.Markdown(
"*Note: The **Ground Truth** column shows the score rendered from "
"our discrete feature representation. When the original MusicXML is "
"available, tick \"Raw MXL engraving\" to see the original "
"publisher engraving instead.*",
)
with gr.Row():
with gr.Column():
with gr.Row():
gr.Markdown("#### Ground Truth")
show_orig_mxl = gr.Checkbox(
label="Raw MXL engraving",
value=True,
visible=False,
scale=0,
)
gt_image = gr.Image(
type="filepath",
show_label=False,
)
with gr.Column():
gr.Markdown("#### Reconstructed")
recon_image = gr.Image(
type="filepath",
show_label=False,
)
gr.Markdown("#### Note-level accuracy")
accuracy_output = gr.Markdown()
# -- Callbacks --
def on_run(graph_file, example_name, model_name):
if graph_file is None and example_name != "(none)":
# Resolve display name back to filename
filename = _EXAMPLE_DISPLAY.get(example_name, example_name)
graph_file = str(EXAMPLES_DIR / filename)
(
gt_first, recon_first,
gt_orig_pgs, gt_recon_pgs, recon_pgs,
has_orig, num_pages,
mxl, acc_md, info_md,
) = run_reconstruction(graph_file, model_name)
slider_update = gr.Slider(
minimum=1, maximum=num_pages, step=1, value=1,
label="Page", visible=(num_pages > 1),
)
checkbox_update = gr.Checkbox(
label="Raw MXL engraving",
value=True,
visible=has_orig,
)
return (
gt_first, recon_first,
gt_orig_pgs, gt_recon_pgs, recon_pgs,
has_orig,
slider_update, checkbox_update,
mxl, acc_md, info_md,
)
def on_page_change(page_num, show_orig, gt_orig_pgs, gt_recon_pgs, recon_pgs, has_orig):
idx = int(page_num) - 1
gt_pgs = gt_orig_pgs if (show_orig and has_orig) else gt_recon_pgs
gt_img = gt_pgs[idx] if idx < len(gt_pgs) else None
recon_img = recon_pgs[idx] if idx < len(recon_pgs) else None
return gt_img, recon_img
def on_toggle_orig(show_orig, page_num, gt_orig_pgs, gt_recon_pgs, recon_pgs, has_orig):
gt_pgs = gt_orig_pgs if (show_orig and has_orig) else gt_recon_pgs
# Recompute page count from both active GT and reconstructed
num_pages = max(len(gt_pgs), len(recon_pgs), 1)
clamped_page = min(int(page_num), num_pages)
idx = clamped_page - 1
gt_img = gt_pgs[idx] if idx < len(gt_pgs) else None
recon_img = recon_pgs[idx] if idx < len(recon_pgs) else None
slider_update = gr.Slider(
minimum=1, maximum=num_pages,
step=1, value=clamped_page,
label="Page", visible=(num_pages > 1),
)
return gt_img, recon_img, slider_update
run_btn.click(
fn=on_run,
inputs=[graph_input, example_dropdown, model_dropdown],
outputs=[
gt_image, recon_image,
gt_orig_pages_state, gt_recon_pages_state, recon_pages_state,
has_orig_mxl_state,
page_slider, show_orig_mxl,
mxl_output, accuracy_output, info_output,
],
)
page_slider.change(
fn=on_page_change,
inputs=[page_slider, show_orig_mxl, gt_orig_pages_state, gt_recon_pages_state, recon_pages_state, has_orig_mxl_state],
outputs=[gt_image, recon_image],
)
show_orig_mxl.change(
fn=on_toggle_orig,
inputs=[show_orig_mxl, page_slider, gt_orig_pages_state, gt_recon_pages_state, recon_pages_state, has_orig_mxl_state],
outputs=[gt_image, recon_image, page_slider],
)
return demo
if __name__ == "__main__":
demo = build_demo()
demo.launch()
|