File size: 15,332 Bytes
41a06ac | 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 | """
ZPCodec: full codec model combining encoder, RVQ, optional repair, and decoder.
Data flow:
waveform [B, 1, T]
-> ZPEncoder -> latent z [B, D, T']
-> ResidualVQ -> quantized z_q [B, D, T'], indices, commit_loss
-> (GE simulator) -> frame_mask [B, T'] (training only, if use_repair=True)
-> LatentRepairTransformer -> z_q_post [B, D, T'] (missing frames concealed)
-> ZPDecoder -> waveform [B, 1, T_out]
T' = T / hop_length (hop_length = prod(ratios) = 240 for ratios=[8,5,3,2] -> 15ms/frame)
The repair module is optional (use_repair=False for stage 1 codec-only training).
The GE simulator is optional too: if no GilbertElliottConfig is provided, no
packet loss is simulated and frame_mask is never generated automatically.
"""
import typing as tp
from contextlib import contextmanager
import numpy as np
import torch
import torch.nn as nn
from vector_quantize_pytorch import ResidualVQ
from .components import ZPEncoder, ZPDecoder
from .repair import LatentRepairTransformer
from .GilbertElliot import GilbertElliottConfig, GilbertElliottSimulator
@contextmanager
def temporarily_set(obj, attr: str, value):
"""Context manager that sets obj.attr = value for the duration of the block,
then restores the original value. Used to toggle quantize_dropout per-batch."""
original = getattr(obj, attr)
setattr(obj, attr, value)
try:
yield
finally:
setattr(obj, attr, original)
class ZPCodec(nn.Module):
"""
Full codec: encoder -> RVQ -> (repair) -> decoder.
Stage 1 (codec pre-training): use_repair=False, no GilbertElliottConfig.
Stage 2 (repair training): use_repair=True, GilbertElliottConfig provided.
Stage 3 (joint fine-tuning): use_repair=True, GE curriculum via set_gilbert_elliott_config().
"""
def __init__(
self,
channels: int = 1,
dimension: int = 128,
n_filters: int = 32,
ratios: tp.List[int] = [8, 5, 3, 2],
norm: str = 'weight_norm',
causal: bool = True,
num_quantizers: int = 9,
codebook_size: int = 1024,
sample_rate: int = 16000,
# --- Repair module ---
use_repair: bool = False,
repair_hidden_dim: int = 256,
repair_num_layers: int = 4,
repair_num_heads: int = 4,
repair_ffn_mult: int = 2,
repair_past: int = 8,
repair_future: int = 2,
repair_two_pass: bool = True,
# --- Packet loss simulation ---
gilbert_elliott_config: tp.Optional[GilbertElliottConfig] = None,
):
super().__init__()
self.encoder = ZPEncoder(
channels=channels,
dimension=dimension,
n_filters=n_filters,
ratios=ratios,
norm=norm,
causal=causal,
)
self.rvq = ResidualVQ(
dim=dimension,
num_quantizers=num_quantizers,
codebook_size=codebook_size,
kmeans_init=True,
kmeans_iters=10,
use_cosine_sim=True, # prop to improved RVQGAN's paper
threshold_ema_dead_code=2,
quantize_dropout=True,
quantize_dropout_cutoff_index=5, # first 5 quantizers are always active -
# theoretically with 5 quant active we can switch to 3kbps. But this was not my focus for that project...
quantize_dropout_multiple_of=1,
)
self.decoder = ZPDecoder(
channels=channels,
dimension=dimension,
n_filters=n_filters,
ratios=ratios,
norm=norm,
causal=causal,
)
self.sample_rate = sample_rate
self.hop_length = int(np.prod(ratios)) # 240 for ratios=[8,5,3,2]
self.use_repair = use_repair
self.repair_two_pass = repair_two_pass
if use_repair:
self.repair = LatentRepairTransformer(
latent_dim=dimension,
hidden_dim=repair_hidden_dim,
num_layers=repair_num_layers,
num_heads=repair_num_heads,
ffn_mult=repair_ffn_mult,
past=repair_past,
future=repair_future,
)
else:
self.repair = None
self.ge_simulator: tp.Optional[GilbertElliottSimulator] = None
if gilbert_elliott_config is not None:
self.set_gilbert_elliott_config(gilbert_elliott_config)
# Runtime configuration of the packet-loss simulator
def set_gilbert_elliott_config(self, config: GilbertElliottConfig) -> None:
"""Replace the GE simulator at runtime. Called between training stages
to apply a harder loss curriculum without reloading the model."""
self.ge_simulator = GilbertElliottSimulator(
config=config,
sample_rate=self.sample_rate,
hop_length=self.hop_length,
)
def sample_frame_mask(
self,
batch_size: int,
num_frames: int,
device: tp.Optional[torch.device] = None,
seed: tp.Optional[int] = None,
) -> torch.Tensor:
"""Expose the GE simulator directly. Useful when the same mask needs to
be reused across multiple points (e.g. logging, loss weighting)."""
assert self.ge_simulator is not None, (
"No GilbertElliottConfig configured. Call set_gilbert_elliott_config() first."
)
return self.ge_simulator.sample_frame_mask(
batch_size, num_frames, device=device, seed=seed
)
# Encoding
def _encode_raw(self, x: torch.Tensor):
"""Encode waveform to quantized latent. Returns (z, z_q, indices, commit_loss).
quantize_dropout is randomly toggled per-call during training to teach
the decoder to handle a variable number of active quantizers (bitrate scalability)."""
z = self.encoder(x) # [B, D, T']
z_seq = z.permute(0, 2, 1) # [B, T', D] — RVQ expects (B, T, D)
use_dropout = self.training and (torch.rand(1).item() < 0.5) # dropout applied only 50% of the time, this improve the
# quality at full kbps. Citing the improved RVQGAN paper.
with temporarily_set(self.rvq, 'quantize_dropout', use_dropout):
z_q, indices, commit_loss = self.rvq(z_seq)
z_q = z_q.permute(0, 2, 1) # [B, D, T']
return z, z_q, indices, commit_loss
# Repair
def _apply_repair(
self,
z_q: torch.Tensor,
frame_mask: torch.Tensor,
) -> torch.Tensor:
"""Run the repair transformer and selectively substitute only missing frames.
z_q: [B, D, T']
frame_mask: [B, T'] 1 = received, 0 = missing
The transformer outputs a full [B, D, T'] tensor, but received frames are
kept as-is from z_q — only positions where frame_mask == 0 are replaced.
This means z_q_post == z_q on received frames by construction, which is
important for latent_repair_loss (the mask isolates the useful gradient).
Two-pass mode (repair_two_pass=True): mimics streaming deployment where
previous repair estimates are already in the buffer when estimating frame t.
See LatentRepairTransformer.forward_two_pass for the full explanation.
"""
assert self.repair is not None, "use_repair=False, repair not initialised"
z_seq = z_q.permute(0, 2, 1) # [B, T', D]
if self.repair_two_pass:
z_repaired = self.repair.forward_two_pass(z_seq, frame_mask)
else:
# Single-pass fallback
z_seq_filled = self.repair.fill_missing(z_seq, frame_mask)
z_repaired = self.repair(z_seq_filled, frame_mask)
# Selective substitution: keep received frames from z_q, replace missing ones
m = frame_mask.unsqueeze(-1).to(z_seq.dtype) # [B, T', 1]
z_out = z_seq * m + z_repaired * (1.0 - m)
return z_out.permute(0, 2, 1) # [B, D, T']
def _get_frame_mask(
self,
z_q: torch.Tensor,
frame_mask: tp.Optional[torch.Tensor],
) -> torch.Tensor:
"""Return the provided frame_mask, or sample one from the GE simulator."""
if frame_mask is not None:
return frame_mask
assert self.ge_simulator is not None, (
"use_repair=True but no GilbertElliottConfig configured. "
"Call set_gilbert_elliott_config() before training."
)
B, _, T_prime = z_q.shape
return self.ge_simulator.sample_frame_mask(B, T_prime, device=z_q.device)
# Public encode / decode API
def encode(self, x: torch.Tensor) -> tp.Tuple[torch.Tensor, torch.Tensor]:
"""Encode waveform to (z_q, indices). x: [B, 1, T]"""
_, z_q, indices, _ = self._encode_raw(x)
return z_q, indices
def decode(
self,
z_q: torch.Tensor,
frame_mask: tp.Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Decode quantized latent to waveform.
z_q: [B, D, T']
frame_mask: [B, T'] optional; if provided and use_repair=True, runs repair first.
"""
if self.use_repair and frame_mask is not None:
z_q = self._apply_repair(z_q, frame_mask)
return self.decoder(z_q)
# Training forward
def forward(
self,
x: torch.Tensor,
frame_mask: tp.Optional[torch.Tensor] = None,
return_intermediates: bool = False,
):
"""
x: [B, 1, T]
frame_mask: [B, T'] optional. If use_repair=True and None,
sampled automatically from the GE simulator.
return_intermediates: if True, also returns z_q pre/post repair and the
effective frame_mask — required by latent_repair_loss
and ZPCodecTrainer.forward_codec during training.
Returns:
return_intermediates=False: (x_hat, commit_loss)
return_intermediates=True: (x_hat, commit_loss, z_q_pre, z_q_post, frame_mask)
When use_repair=False: z_q_pre == z_q_post and frame_mask == None.
"""
_, z_q_pre, _, commit_loss = self._encode_raw(x)
commit_loss = commit_loss.mean()
if self.use_repair:
frame_mask = self._get_frame_mask(z_q_pre, frame_mask)
z_q_post = self._apply_repair(z_q_pre, frame_mask)
else:
z_q_post = z_q_pre
frame_mask = None
x_hat = self.decoder(z_q_post)
if return_intermediates:
return x_hat, commit_loss, z_q_pre, z_q_post, frame_mask
return x_hat, commit_loss
# ------------------------------------------------------------------
# from_pretrained — load from Hugging Face Hub or local path
# ------------------------------------------------------------------
@classmethod
def from_pretrained(
cls,
model_id: str,
device: str = "cpu",
filename: str = "zpcodec_weights.pt",
**hf_kwargs,
) -> "ZPCodec":
"""
Load ZPCodec from a Hugging Face Hub repo or a local file.
Args:
model_id: HF repo id (e.g. "yourname/zpcodec") OR a local path
to a .pt file OR a local directory containing filename.
device: "cpu" | "cuda" | "cuda:0" etc.
filename: name of the weights file inside the HF repo.
**hf_kwargs: forwarded to huggingface_hub.hf_hub_download
(e.g. revision="main", token="hf_...").
Returns:
ZPCodec in eval mode.
Examples:
# From Hugging Face Hub
model = ZPCodec.from_pretrained("yourname/zpcodec")
# From a local .pt file
model = ZPCodec.from_pretrained("./zpcodec_weights.pt")
# With explicit device
model = ZPCodec.from_pretrained("yourname/zpcodec", device="cuda")
"""
import os
import torch
# Resolve checkpoint path: local file, local dir, or HF Hub
if os.path.isfile(model_id):
ckpt_path = model_id
elif os.path.isdir(model_id):
ckpt_path = os.path.join(model_id, filename)
if not os.path.isfile(ckpt_path):
raise FileNotFoundError(
f"No '{filename}' found in directory '{model_id}'"
)
else:
# Treat as a Hugging Face Hub repo id
try:
from huggingface_hub import hf_hub_download
except ImportError:
raise ImportError(
"huggingface_hub is required to download from the Hub.\n"
"Install with: pip install huggingface_hub"
)
ckpt_path = hf_hub_download(
repo_id=model_id, filename=filename, **hf_kwargs
)
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
# Support both clean checkpoints (with 'config' key) and raw
# full-trainer checkpoints (with 'args' key) for backward compat
if "config" in ckpt:
cfg = ckpt["config"]
state_dict = ckpt["model_state_dict"]
elif "args" in ckpt and "trainer" in ckpt:
# Full trainer checkpoint — extract codec weights and config
args = ckpt["args"]
state_dict = {
k[len("codec."):]: v
for k, v in ckpt["trainer"].items()
if k.startswith("codec.")
}
cfg = {
"channels": 1, "dimension": args["dimension"],
"n_filters": args["n_filters"], "ratios": [8, 5, 3, 2],
"norm": "weight_norm", "causal": True,
"num_quantizers": args["num_quantizers"],
"codebook_size": args["codebook_size"], "sample_rate": 16000,
"use_repair": True,
"repair_hidden_dim": args["repair_hidden_dim"],
"repair_num_layers": args["repair_num_layers"],
"repair_num_heads": args["repair_num_heads"],
"repair_ffn_mult": args["repair_ffn_mult"],
"repair_past": args["repair_past"],
"repair_future": args["repair_future"],
"repair_two_pass": True,
}
else:
raise ValueError(
"Unrecognised checkpoint format. "
"Expected keys: 'config'+'model_state_dict' or 'args'+'trainer'."
)
model = cls(**cfg)
missing, unexpected = model.load_state_dict(state_dict, strict=True)
if missing:
raise RuntimeError(f"Missing keys: {missing[:5]}")
if unexpected:
raise RuntimeError(f"Unexpected keys: {unexpected[:5]}")
n_params = sum(p.numel() for p in model.parameters()) / 1e6
info = ckpt.get("training_info", {})
stoi = info.get("best_val_stoi", ckpt.get("best_val_metric", "?"))
print(f"✓ ZPCodec loaded — {n_params:.1f}M params | best val STOI: {stoi}")
model = model.to(device)
model.eval()
return model
|