Spaces:
Restarting
on
L40S
Restarting
on
L40S
Commit
Β·
1f74f2f
1
Parent(s):
dec57f5
reverting
Browse files- jam_worker.py +142 -152
jam_worker.py
CHANGED
@@ -410,196 +410,186 @@ class JamWorker(threading.Thread):
|
|
410 |
self._pending_drop_intro_bars = getattr(self, "_pending_drop_intro_bars", 0) + 1
|
411 |
|
412 |
def run(self):
|
413 |
-
"""Main worker loop β
|
414 |
-
|
415 |
-
|
416 |
-
|
417 |
-
|
418 |
-
|
419 |
-
|
420 |
-
|
421 |
-
|
422 |
-
|
423 |
-
|
424 |
-
|
425 |
-
|
426 |
-
|
427 |
-
|
428 |
-
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
|
433 |
-
self._resampler = None
|
434 |
-
self._stream_out = None
|
435 |
-
self._resample_cursor_in = int(getattr(self, "_resample_cursor_in", 0)) # how many INPUT samples we fed to the resampler
|
436 |
-
|
437 |
-
if sr_out != sr_in:
|
438 |
-
try:
|
439 |
-
from utils import StreamingResampler
|
440 |
-
self._resampler = StreamingResampler(in_sr=sr_in, out_sr=sr_out, channels=ch, quality="VHQ")
|
441 |
-
self._stream_out = _np.zeros((0, ch), dtype=_np.float32)
|
442 |
-
except Exception as e:
|
443 |
-
print(f"β οΈ Could not init StreamingResampler ({e}); falling back to alias-mode (sr_out==sr_in).")
|
444 |
-
sr_out = sr_in
|
445 |
-
self.params.target_sr = sr_out
|
446 |
-
self._resampler = None
|
447 |
-
self._stream_out = _np.zeros((0, ch), dtype=_np.float32)
|
448 |
-
self._resample_cursor_in = 0
|
449 |
-
else:
|
450 |
-
self._stream_out = _np.zeros((0, ch), dtype=_np.float32)
|
451 |
-
self._resample_cursor_in = 0
|
452 |
-
|
453 |
-
# --- helper: how many more INPUT samples (stable) we need to be able to emit next client chunk ---
|
454 |
-
def _need(first_chunk_extra: bool=False) -> int:
|
455 |
-
start = int(getattr(self, "_next_emit_start", 0))
|
456 |
-
total_in = 0 if getattr(self, "_stream", None) is None else int(self._stream.shape[0])
|
457 |
-
total_in_stable = max(0, total_in - xfade_n_in) # hold back xfade tail (overlap will be replaced)
|
458 |
-
have = max(0, total_in_stable - start)
|
459 |
emit_phase = float(getattr(self, "_emit_phase", 0.0))
|
460 |
-
|
461 |
-
|
|
|
|
|
462 |
if first_chunk_extra:
|
463 |
-
|
|
|
|
|
|
|
464 |
return max(0, want - have)
|
465 |
|
466 |
-
|
467 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
468 |
|
469 |
-
# --- main loop ---
|
470 |
while not self._stop_event.is_set():
|
471 |
-
# 0) Backpressure: don't run too far ahead
|
472 |
if not self._should_generate_next_chunk():
|
473 |
-
time.sleep(0.
|
474 |
continue
|
475 |
|
476 |
-
# 1)
|
477 |
need = _need(first_chunk_extra=(self.idx == 0))
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
|
483 |
-
|
484 |
-
|
485 |
wav, self.state = self.mrt.generate_chunk(state=self.state, style=style_vec)
|
486 |
-
|
487 |
-
|
488 |
-
|
489 |
-
|
490 |
-
|
491 |
-
|
492 |
-
|
493 |
-
|
494 |
-
|
495 |
-
|
496 |
-
|
497 |
-
|
498 |
-
|
499 |
-
|
500 |
-
|
501 |
-
|
502 |
-
|
503 |
-
|
504 |
-
continue
|
505 |
-
|
506 |
-
# 2) Optional, one-shot bar realign (occurs on first slice or reseed)
|
507 |
-
if getattr(self, "_needs_bar_realign", False):
|
508 |
-
self._realign_emit_pointer_to_bar(sr_in)
|
509 |
-
self._emit_phase = 0.0 # restart fractional phase at clean bar
|
510 |
self._needs_bar_realign = False
|
511 |
self._reseed_ref_loop = None
|
512 |
|
513 |
-
# 3)
|
514 |
-
|
515 |
-
|
516 |
-
|
517 |
-
|
518 |
-
|
519 |
-
|
520 |
-
|
521 |
-
step_total_out = chunk_step_out_f + self._emit_phase_out
|
522 |
-
step_int_out = int(floor(step_total_out))
|
523 |
-
new_phase_out = float(step_total_out - step_int_out)
|
524 |
-
end_out = start_out + step_int_out
|
525 |
-
|
526 |
-
# 4) Guards β do we actually have enough ready in both domains?
|
527 |
-
total_in_stable = 0
|
528 |
-
if getattr(self, "_stream", None) is not None:
|
529 |
-
total_in_stable = max(0, int(self._stream.shape[0]) - xfade_n_in)
|
530 |
-
total_out_ready = 0 if self._stream_out is None else int(self._stream_out.shape[0])
|
531 |
-
|
532 |
-
if end_in > total_in_stable or end_out > total_out_ready:
|
533 |
-
time.sleep(0.005)
|
534 |
continue
|
535 |
|
536 |
-
|
537 |
-
|
538 |
-
|
539 |
-
# Advance pointers + phases atomically
|
540 |
-
self._next_emit_start = end_in
|
541 |
-
self._emit_phase = new_phase_in
|
542 |
-
self._next_emit_start_out = end_out
|
543 |
-
self._emit_phase_out = new_phase_out
|
544 |
|
545 |
-
|
546 |
-
y = au.Waveform(slice_out.astype(_np.float32, copy=False), sr_out).as_stereo()
|
547 |
|
548 |
-
#
|
549 |
if self.idx == 0 and self.params.ref_loop is not None:
|
550 |
y, _ = match_loudness_to_reference(
|
551 |
self.params.ref_loop, y,
|
552 |
method=self.params.loudness_mode,
|
553 |
headroom_db=self.params.headroom_db
|
554 |
)
|
555 |
-
|
|
|
556 |
|
557 |
-
#
|
558 |
-
b64,
|
559 |
-
y
|
560 |
)
|
|
|
561 |
|
562 |
-
|
563 |
-
"bpm": int(round(self.params.bpm)),
|
564 |
-
"bars": int(self.params.bars_per_chunk),
|
565 |
-
"beats_per_bar": int(self.params.beats_per_bar),
|
566 |
-
"sample_rate": int(sr_out),
|
567 |
-
"channels": int(channels),
|
568 |
-
"total_samples": int(total_samples),
|
569 |
-
"seconds_per_bar": float(spb),
|
570 |
-
"loop_duration_seconds": float(self.params.bars_per_chunk) * float(spb),
|
571 |
-
"guidance_weight": float(self.params.guidance_weight),
|
572 |
-
"temperature": float(self.params.temperature),
|
573 |
-
"topk": int(self.params.topk),
|
574 |
-
"xfade_seconds": float(xfade_s),
|
575 |
-
}
|
576 |
-
|
577 |
with self._lock:
|
578 |
self.idx += 1
|
579 |
self.outbox.append(JamChunk(index=self.idx, audio_base64=b64, metadata=meta))
|
580 |
-
# prune outbox
|
581 |
if len(self.outbox) > 10:
|
582 |
cutoff = self._last_delivered_index - 5
|
583 |
self.outbox = [ch for ch in self.outbox if ch.index > cutoff]
|
584 |
|
585 |
-
#
|
586 |
-
if
|
587 |
pkg = self._pending_reseed
|
588 |
self._pending_reseed = None
|
589 |
-
|
590 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
591 |
self._needs_bar_realign = True
|
592 |
-
self._reseed_ref_loop = pkg.get("ref") if isinstance(pkg, dict) else None
|
593 |
|
594 |
-
|
595 |
|
596 |
-
|
597 |
-
try:
|
598 |
-
if self._resampler is not None:
|
599 |
-
tail = self._resampler.flush()
|
600 |
-
if tail.size:
|
601 |
-
self._stream_out = tail if self._stream_out.size == 0 else _np.vstack([self._stream_out, tail])
|
602 |
-
except Exception as e:
|
603 |
-
print(f"β οΈ Resampler flush error: {e}")
|
604 |
|
605 |
print("π JamWorker stopped")
|
|
|
|
410 |
self._pending_drop_intro_bars = getattr(self, "_pending_drop_intro_bars", 0) + 1
|
411 |
|
412 |
def run(self):
|
413 |
+
"""Main worker loop β generate into a continuous stream, then emit bar-aligned slices."""
|
414 |
+
spb = self._seconds_per_bar() # seconds per bar
|
415 |
+
chunk_secs = self.params.bars_per_chunk * spb
|
416 |
+
xfade = float(self.mrt.config.crossfade_length) # seconds
|
417 |
+
sr = int(self.mrt.sample_rate)
|
418 |
+
chunk_step_f = chunk_secs * sr # float samples per chunk
|
419 |
+
self._emit_phase = getattr(self, "_emit_phase", 0.0)
|
420 |
+
|
421 |
+
def _need(first_chunk_extra: bool = False) -> int:
|
422 |
+
"""
|
423 |
+
How many more samples we still need in the stream to emit the next slice.
|
424 |
+
Uses the fractional step (chunk_step_f) + current _emit_phase to compute
|
425 |
+
the *integer* number of samples required for the next chunk, without
|
426 |
+
mutating _emit_phase here.
|
427 |
+
"""
|
428 |
+
start = getattr(self, "_next_emit_start", 0)
|
429 |
+
total = 0 if getattr(self, "_stream", None) is None else self._stream.shape[0]
|
430 |
+
have = max(0, total - start)
|
431 |
+
|
432 |
+
# Compute the integer step we'd use for the next emit, non-mutating.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
433 |
emit_phase = float(getattr(self, "_emit_phase", 0.0))
|
434 |
+
step_int = int(floor(chunk_step_f + emit_phase)) # matches the logic used when advancing
|
435 |
+
|
436 |
+
# How much we want available beyond 'start' for this emit.
|
437 |
+
want = step_int
|
438 |
if first_chunk_extra:
|
439 |
+
# Reserve two extra bars so the first-chunk onset alignment has material.
|
440 |
+
# Use ceil to be conservative so we don't under-request.
|
441 |
+
want += int(ceil(2.0 * spb * sr))
|
442 |
+
|
443 |
return max(0, want - have)
|
444 |
|
445 |
+
def _mono_env(x: np.ndarray, sr: int, win_ms: float = 10.0) -> np.ndarray:
|
446 |
+
if x.ndim == 2: x = x.mean(axis=1)
|
447 |
+
x = np.abs(x).astype(np.float32)
|
448 |
+
w = max(1, int(round(win_ms * 1e-3 * sr)))
|
449 |
+
if w > 1:
|
450 |
+
kern = np.ones(w, dtype=np.float32) / float(w)
|
451 |
+
x = np.convolve(x, kern, mode="same")
|
452 |
+
d = np.diff(x, prepend=x[:1])
|
453 |
+
d[d < 0] = 0.0
|
454 |
+
return d
|
455 |
+
|
456 |
+
def _estimate_first_offset_samples(ref_loop_wav, gen_head_wav, sr: int, spb: float) -> int:
|
457 |
+
"""Tempo-aware first-downbeat offset (positive => model late)."""
|
458 |
+
try:
|
459 |
+
max_ms = int(max(160.0, min(0.25 * spb * 1000.0, 450.0)))
|
460 |
+
ref = ref_loop_wav if ref_loop_wav.sample_rate == sr else ref_loop_wav.resample(sr)
|
461 |
+
n_bar = int(round(spb * sr))
|
462 |
+
ref_tail = ref.samples[-n_bar:, :] if ref.samples.shape[0] >= n_bar else ref.samples
|
463 |
+
gen_head = gen_head_wav.samples[: int(2 * n_bar), :]
|
464 |
+
if ref_tail.size == 0 or gen_head.size == 0:
|
465 |
+
return 0
|
466 |
+
|
467 |
+
# envelopes + z-score
|
468 |
+
def _z(a):
|
469 |
+
m, s = float(a.mean()), float(a.std() or 1.0); return (a - m) / s
|
470 |
+
e_ref = _z(_mono_env(ref_tail, sr)).astype(np.float32)
|
471 |
+
e_gen = _z(_mono_env(gen_head, sr)).astype(np.float32)
|
472 |
+
|
473 |
+
# upsample x4 for finer lag
|
474 |
+
def _upsample(a, r=4):
|
475 |
+
n = len(a); grid = np.arange(n, dtype=np.float32)
|
476 |
+
fine = np.linspace(0, n - 1, num=n * r, dtype=np.float32)
|
477 |
+
return np.interp(fine, grid, a).astype(np.float32)
|
478 |
+
up = 4
|
479 |
+
e_ref_u, e_gen_u = _upsample(e_ref, up), _upsample(e_gen, up)
|
480 |
+
|
481 |
+
max_lag_u = int(round((max_ms / 1000.0) * sr * up))
|
482 |
+
seg = min(len(e_ref_u), len(e_gen_u))
|
483 |
+
e_ref_u = e_ref_u[-seg:]
|
484 |
+
pad = np.zeros(max_lag_u, dtype=np.float32)
|
485 |
+
e_gen_u_pad = np.concatenate([pad, e_gen_u, pad])
|
486 |
+
|
487 |
+
best_lag_u, best_score = 0, -1e9
|
488 |
+
for lag_u in range(-max_lag_u, max_lag_u + 1):
|
489 |
+
start = max_lag_u + lag_u
|
490 |
+
b = e_gen_u_pad[start : start + seg]
|
491 |
+
denom = (np.linalg.norm(e_ref_u) * np.linalg.norm(b)) or 1.0
|
492 |
+
score = float(np.dot(e_ref_u, b) / denom)
|
493 |
+
if score > best_score:
|
494 |
+
best_score, best_lag_u = score, lag_u
|
495 |
+
return int(round(best_lag_u / up))
|
496 |
+
except Exception:
|
497 |
+
return 0
|
498 |
+
|
499 |
+
print("π JamWorker started (bar-aligned streaming)β¦")
|
500 |
|
|
|
501 |
while not self._stop_event.is_set():
|
|
|
502 |
if not self._should_generate_next_chunk():
|
503 |
+
time.sleep(0.25)
|
504 |
continue
|
505 |
|
506 |
+
# 1) Generate until we have enough material in the stream
|
507 |
need = _need(first_chunk_extra=(self.idx == 0))
|
508 |
+
while need > 0 and not self._stop_event.is_set():
|
509 |
+
with self._lock:
|
510 |
+
style_vec = self.params.style_vec
|
511 |
+
self.mrt.guidance_weight = float(self.params.guidance_weight)
|
512 |
+
self.mrt.temperature = float(self.params.temperature)
|
513 |
+
self.mrt.topk = int(self.params.topk)
|
|
|
514 |
wav, self.state = self.mrt.generate_chunk(state=self.state, style=style_vec)
|
515 |
+
self._append_model_chunk_to_stream(wav) # equal-power xfade into a persistent stream
|
516 |
+
need = _need(first_chunk_extra=(self.idx == 0))
|
517 |
+
|
518 |
+
if self._stop_event.is_set():
|
519 |
+
break
|
520 |
+
|
521 |
+
# 2) One-time: align the emit pointer to the groove
|
522 |
+
if (self.idx == 0 and self.params.combined_loop is not None) or self._needs_bar_realign:
|
523 |
+
ref_loop = self._reseed_ref_loop or self.params.combined_loop
|
524 |
+
if ref_loop is not None:
|
525 |
+
head_len = min(self._stream.shape[0] - self._next_emit_start, int(round(2 * spb * sr)))
|
526 |
+
seg = self._stream[self._next_emit_start : self._next_emit_start + head_len]
|
527 |
+
gen_head = au.Waveform(seg.astype(np.float32, copy=False), sr).as_stereo()
|
528 |
+
offs = _estimate_first_offset_samples(ref_loop, gen_head, sr, spb)
|
529 |
+
if offs != 0:
|
530 |
+
self._next_emit_start = max(0, self._next_emit_start + offs)
|
531 |
+
print(f"π― Offset compensation: {offs/sr:+.3f}s")
|
532 |
+
self._realign_emit_pointer_to_bar(sr)
|
|
|
|
|
|
|
|
|
|
|
|
|
533 |
self._needs_bar_realign = False
|
534 |
self._reseed_ref_loop = None
|
535 |
|
536 |
+
# 3) Emit exactly bars_per_chunk Γ spb from the stream
|
537 |
+
start = self._next_emit_start
|
538 |
+
step_total = chunk_step_f + self._emit_phase
|
539 |
+
step_int = int(np.floor(step_total))
|
540 |
+
self._emit_phase = float(step_total - step_int)
|
541 |
+
end = start + step_int
|
542 |
+
if end > self._stream.shape[0]:
|
543 |
+
# shouldn't happen often; generate a bit more and loop
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
544 |
continue
|
545 |
|
546 |
+
slice_ = self._stream[start:end]
|
547 |
+
self._next_emit_start = end
|
|
|
|
|
|
|
|
|
|
|
|
|
548 |
|
549 |
+
y = au.Waveform(slice_.astype(np.float32, copy=False), sr).as_stereo()
|
|
|
550 |
|
551 |
+
# 4) Post-processing / loudness
|
552 |
if self.idx == 0 and self.params.ref_loop is not None:
|
553 |
y, _ = match_loudness_to_reference(
|
554 |
self.params.ref_loop, y,
|
555 |
method=self.params.loudness_mode,
|
556 |
headroom_db=self.params.headroom_db
|
557 |
)
|
558 |
+
else:
|
559 |
+
apply_micro_fades(y, 3)
|
560 |
|
561 |
+
# 5) Resample + exact-length snap + encode
|
562 |
+
b64, meta = self._snap_and_encode(
|
563 |
+
y, seconds=chunk_secs, target_sr=self.params.target_sr, bars=self.params.bars_per_chunk
|
564 |
)
|
565 |
+
meta["xfade_seconds"] = xfade
|
566 |
|
567 |
+
# 6) Publish
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
568 |
with self._lock:
|
569 |
self.idx += 1
|
570 |
self.outbox.append(JamChunk(index=self.idx, audio_base64=b64, metadata=meta))
|
|
|
571 |
if len(self.outbox) > 10:
|
572 |
cutoff = self._last_delivered_index - 5
|
573 |
self.outbox = [ch for ch in self.outbox if ch.index > cutoff]
|
574 |
|
575 |
+
# π If a reseed was requested, apply it *now*, between chunks
|
576 |
+
if self._pending_reseed is not None:
|
577 |
pkg = self._pending_reseed
|
578 |
self._pending_reseed = None
|
579 |
+
|
580 |
+
new_state = self.mrt.init_state()
|
581 |
+
new_state.context_tokens = pkg["ctx"] # exact (ctx_frames, depth)
|
582 |
+
self.state = new_state
|
583 |
+
|
584 |
+
# start a fresh stream and schedule one-time alignment
|
585 |
+
self._stream = None
|
586 |
+
self._next_emit_start = 0
|
587 |
+
self._reseed_ref_loop = pkg.get("ref") or self.params.combined_loop
|
588 |
self._needs_bar_realign = True
|
|
|
589 |
|
590 |
+
print("π Reseed installed at bar boundary; will realign before next slice")
|
591 |
|
592 |
+
print(f"β
Completed chunk {self.idx}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
593 |
|
594 |
print("π JamWorker stopped")
|
595 |
+
|