Commit
·
2ab3869
1
Parent(s):
946680d
final reseed splice fix
Browse files- jam_worker.py +100 -24
jam_worker.py
CHANGED
@@ -174,6 +174,91 @@ class JamWorker(threading.Thread):
|
|
174 |
|
175 |
# ---------- context / reseed ----------
|
176 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
177 |
def _encode_exact_context_tokens(self, loop: au.Waveform) -> np.ndarray:
|
178 |
"""Build *exactly* context_length_frames worth of tokens (e.g., 250 @ 25fps),
|
179 |
while ensuring the *end* of the audio lands on a bar boundary.
|
@@ -262,44 +347,35 @@ class JamWorker(threading.Thread):
|
|
262 |
(e.g., 250 frames), then splice only the *tail* corresponding to
|
263 |
`anchor_bars` so generation continues smoothly without resetting state.
|
264 |
"""
|
265 |
-
new_ctx = self._encode_exact_context_tokens(recent_wav) # (F,D)
|
266 |
-
F =
|
267 |
-
D = int(self.mrt.config.decoder_codec_rvq_depth)
|
268 |
-
assert new_ctx.shape == (F, D), f"expected {(F, D)}, got {new_ctx.shape}"
|
269 |
|
270 |
# how many frames correspond to the requested anchor bars
|
271 |
spb = self._bar_clock.seconds_per_bar()
|
272 |
-
frames_per_bar = int(round(self._codec_fps * spb))
|
273 |
-
splice_frames = int(round(max(1, anchor_bars) * frames_per_bar))
|
274 |
-
splice_frames = max(1, min(splice_frames, F))
|
275 |
|
276 |
with self._lock:
|
277 |
# snapshot current context
|
278 |
cur = getattr(self.state, "context_tokens", None)
|
279 |
if cur is None:
|
280 |
-
#
|
281 |
self._pending_reseed = {"ctx": new_ctx}
|
282 |
return
|
283 |
-
|
284 |
-
# safety: coerce by trim/pad
|
285 |
-
if cur.shape[0] > F:
|
286 |
-
cur = cur[-F:, :]
|
287 |
-
elif cur.shape[0] < F:
|
288 |
-
pad = np.repeat(cur[0:1, :], F - cur.shape[0], axis=0)
|
289 |
-
cur = np.concatenate([pad, cur], axis=0)
|
290 |
-
if cur.shape[1] != D:
|
291 |
-
cur = cur[:, :D]
|
292 |
|
293 |
# build the spliced tensor: keep left (F - splice) from cur, take right (splice) from new
|
294 |
left = cur[:F - splice_frames, :]
|
295 |
right = new_ctx[F - splice_frames:, :]
|
296 |
spliced = np.concatenate([left, right], axis=0)
|
|
|
297 |
|
298 |
# queue for install at the *next bar boundary* right after emission
|
299 |
self._pending_token_splice = {
|
300 |
"tokens": spliced,
|
301 |
"debug": {"F": F, "D": D, "splice_frames": splice_frames, "frames_per_bar": frames_per_bar}
|
302 |
}
|
|
|
303 |
|
304 |
|
305 |
def reseed_from_waveform(self, wav: au.Waveform):
|
@@ -420,22 +496,22 @@ class JamWorker(threading.Thread):
|
|
420 |
with self._lock:
|
421 |
# Prefer seamless token splice when available
|
422 |
if self._pending_token_splice is not None:
|
|
|
423 |
try:
|
424 |
-
|
425 |
-
self.state.context_tokens = spliced
|
426 |
self._pending_token_splice = None
|
427 |
-
|
428 |
-
#
|
429 |
-
except Exception as e:
|
430 |
-
# fallback: full reseed if setter rejects
|
431 |
new_state = self.mrt.init_state()
|
432 |
new_state.context_tokens = spliced
|
433 |
self.state = new_state
|
434 |
self._model_stream = None
|
435 |
self._pending_token_splice = None
|
436 |
elif self._pending_reseed is not None:
|
|
|
437 |
new_state = self.mrt.init_state()
|
438 |
-
new_state.context_tokens =
|
439 |
self.state = new_state
|
440 |
self._model_stream = None
|
441 |
self._pending_reseed = None
|
|
|
174 |
|
175 |
# ---------- context / reseed ----------
|
176 |
|
177 |
+
def _expected_token_shape(self) -> Tuple[int, int]:
|
178 |
+
F = int(self._ctx_frames)
|
179 |
+
D = int(self.mrt.config.decoder_codec_rvq_depth)
|
180 |
+
return F, D
|
181 |
+
|
182 |
+
def _coerce_tokens(self, toks: np.ndarray) -> np.ndarray:
|
183 |
+
"""Force tokens to (context_length_frames, rvq_depth), padding/trimming as needed.
|
184 |
+
Pads missing frames by repeating the last frame (safer than zeros for RVQ stacks)."""
|
185 |
+
F, D = self._expected_token_shape()
|
186 |
+
if toks.ndim != 2:
|
187 |
+
toks = np.atleast_2d(toks)
|
188 |
+
# depth first
|
189 |
+
if toks.shape[1] > D:
|
190 |
+
toks = toks[:, :D]
|
191 |
+
elif toks.shape[1] < D:
|
192 |
+
pad_cols = np.tile(toks[:, -1:], (1, D - toks.shape[1]))
|
193 |
+
toks = np.concatenate([toks, pad_cols], axis=1)
|
194 |
+
# frames
|
195 |
+
if toks.shape[0] < F:
|
196 |
+
if toks.shape[0] == 0:
|
197 |
+
toks = np.zeros((1, D), dtype=np.int32)
|
198 |
+
pad = np.repeat(toks[-1:, :], F - toks.shape[0], axis=0)
|
199 |
+
toks = np.concatenate([pad, toks], axis=0)
|
200 |
+
elif toks.shape[0] > F:
|
201 |
+
toks = toks[-F:, :]
|
202 |
+
if toks.dtype != np.int32:
|
203 |
+
toks = toks.astype(np.int32, copy=False)
|
204 |
+
return toks
|
205 |
+
|
206 |
+
def _encode_exact_context_tokens(self, loop: au.Waveform) -> np.ndarray:
|
207 |
+
"""Build *exactly* context_length_frames worth of tokens (e.g., 250 @ 25fps),
|
208 |
+
while ensuring the *end* of the audio lands on a bar boundary.
|
209 |
+
Strategy: take the largest integer number of bars <= ctx_seconds as the tail,
|
210 |
+
then left-fill from just before that tail (wrapping if needed) to reach exactly
|
211 |
+
ctx_seconds; finally, pad/trim to exact samples and, as a last resort, pad/trim
|
212 |
+
tokens to the expected frame count.
|
213 |
+
"""
|
214 |
+
wav = loop.as_stereo().resample(self._model_sr)
|
215 |
+
data = wav.samples.astype(np.float32, copy=False)
|
216 |
+
if data.ndim == 1:
|
217 |
+
data = data[:, None]
|
218 |
+
|
219 |
+
spb = self._bar_clock.seconds_per_bar()
|
220 |
+
ctx_sec = float(self._ctx_seconds)
|
221 |
+
sr = int(self._model_sr)
|
222 |
+
|
223 |
+
# bars that fit fully inside ctx_sec (at least 1)
|
224 |
+
bars_fit = max(1, int(ctx_sec // spb))
|
225 |
+
tail_len_samps = int(round(bars_fit * spb * sr))
|
226 |
+
|
227 |
+
# ensure we have enough source by tiling
|
228 |
+
need = int(round(ctx_sec * sr)) + tail_len_samps
|
229 |
+
if data.shape[0] == 0:
|
230 |
+
data = np.zeros((1, 2), dtype=np.float32)
|
231 |
+
reps = int(np.ceil(need / float(data.shape[0])))
|
232 |
+
tiled = np.tile(data, (reps, 1))
|
233 |
+
|
234 |
+
end = tiled.shape[0]
|
235 |
+
tail = tiled[end - tail_len_samps:end]
|
236 |
+
|
237 |
+
# left-fill to reach exact ctx samples (keeps end-of-bar alignment)
|
238 |
+
ctx_samps = int(round(ctx_sec * sr))
|
239 |
+
pad_len = ctx_samps - tail.shape[0]
|
240 |
+
if pad_len > 0:
|
241 |
+
pre = tiled[end - tail_len_samps - pad_len:end - tail_len_samps]
|
242 |
+
ctx = np.concatenate([pre, tail], axis=0)
|
243 |
+
else:
|
244 |
+
ctx = tail[-ctx_samps:]
|
245 |
+
|
246 |
+
# final snap to *exact* ctx samples
|
247 |
+
if ctx.shape[0] < ctx_samps:
|
248 |
+
pad = np.zeros((ctx_samps - ctx.shape[0], ctx.shape[1]), dtype=np.float32)
|
249 |
+
ctx = np.concatenate([pad, ctx], axis=0)
|
250 |
+
elif ctx.shape[0] > ctx_samps:
|
251 |
+
ctx = ctx[-ctx_samps:]
|
252 |
+
|
253 |
+
exact = au.Waveform(ctx, sr)
|
254 |
+
tokens_full = self.mrt.codec.encode(exact).astype(np.int32)
|
255 |
+
depth = int(self.mrt.config.decoder_codec_rvq_depth)
|
256 |
+
tokens = tokens_full[:, :depth]
|
257 |
+
|
258 |
+
# Force expected (F,D) at *return time*
|
259 |
+
tokens = self._coerce_tokens(tokens)
|
260 |
+
return tokens
|
261 |
+
|
262 |
def _encode_exact_context_tokens(self, loop: au.Waveform) -> np.ndarray:
|
263 |
"""Build *exactly* context_length_frames worth of tokens (e.g., 250 @ 25fps),
|
264 |
while ensuring the *end* of the audio lands on a bar boundary.
|
|
|
347 |
(e.g., 250 frames), then splice only the *tail* corresponding to
|
348 |
`anchor_bars` so generation continues smoothly without resetting state.
|
349 |
"""
|
350 |
+
new_ctx = self._encode_exact_context_tokens(recent_wav) # coerce to (F,D)
|
351 |
+
F, D = self._expected_token_shape()
|
|
|
|
|
352 |
|
353 |
# how many frames correspond to the requested anchor bars
|
354 |
spb = self._bar_clock.seconds_per_bar()
|
355 |
+
frames_per_bar = max(1, int(round(self._codec_fps * spb)))
|
356 |
+
splice_frames = max(1, min(int(round(max(1.0, float(anchor_bars)) * frames_per_bar)), F))
|
|
|
357 |
|
358 |
with self._lock:
|
359 |
# snapshot current context
|
360 |
cur = getattr(self.state, "context_tokens", None)
|
361 |
if cur is None:
|
362 |
+
# fall back to full reseed (still coerced)
|
363 |
self._pending_reseed = {"ctx": new_ctx}
|
364 |
return
|
365 |
+
cur = self._coerce_tokens(cur)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
366 |
|
367 |
# build the spliced tensor: keep left (F - splice) from cur, take right (splice) from new
|
368 |
left = cur[:F - splice_frames, :]
|
369 |
right = new_ctx[F - splice_frames:, :]
|
370 |
spliced = np.concatenate([left, right], axis=0)
|
371 |
+
spliced = self._coerce_tokens(spliced)
|
372 |
|
373 |
# queue for install at the *next bar boundary* right after emission
|
374 |
self._pending_token_splice = {
|
375 |
"tokens": spliced,
|
376 |
"debug": {"F": F, "D": D, "splice_frames": splice_frames, "frames_per_bar": frames_per_bar}
|
377 |
}
|
378 |
+
}
|
379 |
|
380 |
|
381 |
def reseed_from_waveform(self, wav: au.Waveform):
|
|
|
496 |
with self._lock:
|
497 |
# Prefer seamless token splice when available
|
498 |
if self._pending_token_splice is not None:
|
499 |
+
spliced = self._coerce_tokens(self._pending_token_splice["tokens"])
|
500 |
try:
|
501 |
+
# inplace update (no reset)
|
502 |
+
self.state.context_tokens = spliced
|
503 |
self._pending_token_splice = None
|
504 |
+
except Exception:
|
505 |
+
# fallback: full reseed using spliced tokens
|
|
|
|
|
506 |
new_state = self.mrt.init_state()
|
507 |
new_state.context_tokens = spliced
|
508 |
self.state = new_state
|
509 |
self._model_stream = None
|
510 |
self._pending_token_splice = None
|
511 |
elif self._pending_reseed is not None:
|
512 |
+
ctx = self._coerce_tokens(self._pending_reseed["ctx"])
|
513 |
new_state = self.mrt.init_state()
|
514 |
+
new_state.context_tokens = ctx
|
515 |
self.state = new_state
|
516 |
self._model_stream = None
|
517 |
self._pending_reseed = None
|