File size: 11,274 Bytes
eca55dc | 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 | import torch
import functools
import torch.nn as nn
import torch.nn.functional as F
from typing import Any, Dict, Optional, Tuple
from src.models.audio_jepa_module import AudioJEPAModule
from src.models.components.random_projection_quantizer import RandomProjectionQuantizer
class RQAJEPAModule(AudioJEPAModule):
"""
RQA-JEPA Lightning Module.
Extends AudioJEPAModule with Random Projection Quantization loss.
Args:
optimizer (torch.optim.Optimizer): Optimizer configuration.
net (Dict[str, Any]): Configuration for sub-modules.
warmup_pct (float): Percentage of total steps for warmup.
final_lr_ratio (float): Ratio of final learning rate to initial learning rate.
ema_decay (float): Initial EMA decay rate.
ema_end_decay (float): Final EMA decay rate.
ema_anneal_end_step (int): Step at which EMA decay reaches ema_end_decay.
spectrogram_adjustment_mode (str): 'pad' or 'truncate' for spectrogram time dimension.
jepa_criterion (torch.nn.Module): Loss function for JEPA (defaults to MSELoss).
rq_criterion (torch.nn.Module): Loss function for RQ (defaults to CrossEntropyLoss).
rq_lambda (float): Weight for JEPA loss (1 - rq_lambda is used for RQ loss).
codebook_dim (int): Codebook dimension for RandomProjectionQuantizer.
vocab_size (int): Vocabulary size for RandomProjectionQuantizer.
rq_input_type (str): 'teacher' or 'spectrogram'. Source for quantization targets.
"""
def __init__(
self,
optimizer: torch.optim.Optimizer,
net: Dict[str, Any],
warmup_pct: float = 0.1,
final_lr_ratio: float = 0.001,
ema_decay: float = 0.996,
ema_end_decay: float = 1.0,
ema_anneal_end_step: Optional[int] = None,
spectrogram_adjustment_mode: str = "pad",
jepa_criterion: Optional[torch.nn.Module] = None,
rq_criterion: Optional[torch.nn.Module] = None,
rq_lambda: float = 0.5,
codebook_dim: int = 16,
vocab_size: int = 8192,
rq_input_type: str = "teacher",
):
super().__init__(
optimizer=optimizer,
net=net,
warmup_pct=warmup_pct,
final_lr_ratio=final_lr_ratio,
ema_decay=ema_decay,
ema_end_decay=ema_end_decay,
ema_anneal_end_step=ema_anneal_end_step,
spectrogram_adjustment_mode=spectrogram_adjustment_mode,
criterion=jepa_criterion, # Pass jepa_criterion as criterion to base class
)
self.save_hyperparameters(
logger=False, ignore=["jepa_criterion", "rq_criterion", "net", "optimizer"]
)
self.rq_lambda = rq_lambda
# Store rq_criterion separately
if rq_criterion is not None:
self.rq_criterion = (
rq_criterion()
if isinstance(rq_criterion, (type, functools.partial))
or callable(rq_criterion)
and not isinstance(rq_criterion, nn.Module)
else rq_criterion
)
else:
self.rq_criterion = nn.CrossEntropyLoss()
self.rq_input_type = rq_input_type
if self.rq_input_type not in ["teacher", "spectrogram"]:
raise ValueError(
f"rq_input_type must be 'teacher' or 'spectrogram', got {self.rq_input_type}"
)
# Random Projection Quantizer
# Determine input dimension for quantizer
if self.rq_input_type == "teacher":
# Input to quantizer is teacher output which has encoder_dim
quantizer_input_dim = net.get("encoder", {}).get("embed_dim", 768)
else: # spectrogram
# Input is raw patches
# patch_embed is locally available on self
patch_size = self.patch_embed.patch_size
in_chans = self.patch_embed.in_chans
quantizer_input_dim = patch_size[0] * patch_size[1] * in_chans
self.quantizer = RandomProjectionQuantizer(
input_dim=quantizer_input_dim, cb_dim=codebook_dim, cb_vocab=vocab_size
)
# Freeze quantizer (it is random and fixed)
for p in self.quantizer.parameters():
p.requires_grad = False
# Projection head for RQ prediction
# Takes predictor output (pred_dim) and predicts code indices (vocab_size)
predictor_config = net.get("predictor", {})
predictor_embed_dim = predictor_config.get("embed_dim", 768)
self.rq_proj = nn.Linear(predictor_embed_dim, vocab_size)
def _calculate_combined_loss(
self,
predictions_raw: torch.Tensor,
teacher_targets: torch.Tensor,
rq_logits: torch.Tensor,
rq_targets: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Calculates both JEPA and RQ losses and combines them.
"""
# --- JEPA Loss ---
# Project back to encoder dimension for JEPA loss
predictions_jepa = self.predictor_output_proj(
predictions_raw
) # [B, N_mask, encoder_dim]
jepa_loss = self.criterion(
predictions_jepa, teacher_targets
) # Uses self.criterion (mapped from jepa_criterion)
# --- RQ Loss ---
# Calculate Scale RQ Loss
# Flatten for loss calculation
rq_loss = self.rq_criterion(
rq_logits.reshape(-1, self.hparams.vocab_size), rq_targets.reshape(-1)
)
# --- Combine ---
total_loss = self.rq_lambda * jepa_loss + (1 - self.rq_lambda) * rq_loss
return total_loss, jepa_loss, rq_loss
def _get_raw_patches(self, spec: torch.Tensor) -> torch.Tensor:
"""
Extract raw key-value patches from spectrogram.
Args:
spec (torch.Tensor): Adjusted spectrogram [B, C, F, T].
Returns:
torch.Tensor: Flattened patches [B, N, patch_dim]
"""
patch_size = self.patch_embed.patch_size # (H, W)
# Using kernel_size=patch_size, stride=patch_size ensures non-overlapping patches
# F.unfold returns [B, C*pH*pW, L]
patches = F.unfold(spec, kernel_size=patch_size, stride=patch_size) # [B, D, N]
patches = patches.transpose(1, 2) # [B, N, D]
return patches
def _get_rq_targets_input(
self, spec: torch.Tensor, teacher_full: torch.Tensor, mask_indices: torch.Tensor
) -> torch.Tensor:
"""
Helper to get the input for the RQ quantizer (either teacher embeddings or raw patches).
Only returns the targets for the MASKED locations.
"""
if self.rq_input_type == "teacher":
# Teacher targets at masked locations
return teacher_full[:, mask_indices, :] # [B, N_mask, encoder_dim]
else:
# Raw patches at masked locations
# Check if spec is None, which implies logic error in caller
if spec is None:
raise ValueError(
"Spectrogram cannot be None when rq_input_type is 'spectrogram'"
)
raw_patches = self._get_raw_patches(spec) # [B, N, patch_dim]
return raw_patches[:, mask_indices, :] # [B, N_mask, patch_dim]
def training_step(self, batch: Dict[str, Any], batch_idx: int) -> torch.Tensor:
waveform = batch["waveform"]
patches, current_grid_size = self._process_audio(waveform)
B, N, D = patches.shape
mask = self.mask_generator(1, device=self.device, grid_size=current_grid_size)
mask = mask.expand(B, -1)
student_out = self.compute_student(patches, mask, current_grid_size)
predictions_raw = self.compute_predictor(student_out, mask, current_grid_size)
self._update_teacher()
with torch.no_grad():
teacher_full = self.teacher(patches, grid_size=current_grid_size)
# Prepare targets and logits for RQA-JEPA
m = mask[0]
mask_indices = torch.nonzero(m).flatten()
# Teacher targets always needed for JEPA loss
teacher_targets = teacher_full[:, mask_indices, :] # [B, N_mask, encoder_dim]
# RQ Targets (Quantized)
with torch.no_grad():
# Need spec for 'spectrogram' mode
spec = None
if self.rq_input_type == "spectrogram":
# Re-compute spectrogram as we don't have it exposed from _process_audio
spec = self.spectrogram(waveform)
spec = self._adjust_spectrogram(spec)
rq_targets_input = self._get_rq_targets_input(
spec, teacher_full, mask_indices
)
rq_targets = self.quantizer(rq_targets_input) # [B, N_mask]
# RQ Logits
rq_logits = self.rq_proj(predictions_raw) # [B, N_mask, vocab_size]
loss, jepa_loss, rq_loss = self._calculate_combined_loss(
predictions_raw, teacher_targets, rq_logits, rq_targets
)
self.log(
"train/loss", loss, on_step=True, on_epoch=True, prog_bar=True, batch_size=B
)
self.log(
"train/jepa_loss", jepa_loss, on_step=True, on_epoch=True, batch_size=B
)
self.log("train/rq_loss", rq_loss, on_step=True, on_epoch=True, batch_size=B)
return loss
def validation_step(self, batch: Dict[str, Any], batch_idx: int) -> torch.Tensor:
waveform = batch["waveform"]
patches, current_grid_size = self._process_audio(waveform)
B, N, D = patches.shape
mask = self.mask_generator(1, device=self.device, grid_size=current_grid_size)
mask = mask.expand(B, -1)
student_out = self.compute_student(patches, mask, current_grid_size)
predictions_raw = self.compute_predictor(student_out, mask, current_grid_size)
with torch.no_grad():
teacher_full = self.teacher(patches, grid_size=current_grid_size)
# Prepare targets and logits for RQA-JEPA
m = mask[0]
mask_indices = torch.nonzero(m).flatten()
# Teacher targets at masked locations
teacher_targets = teacher_full[
:, mask_indices, :
] # [B, N_mask, encoder_dim]
# RQ Targets (Quantized)
spec = None
if self.rq_input_type == "spectrogram":
spec = self.spectrogram(waveform)
spec = self._adjust_spectrogram(spec)
rq_targets_input = self._get_rq_targets_input(
spec, teacher_full, mask_indices
)
rq_targets = self.quantizer(rq_targets_input) # [B, N_mask]
# RQ Logits
rq_logits = self.rq_proj(predictions_raw) # [B, N_mask, vocab_size]
loss, jepa_loss, rq_loss = self._calculate_combined_loss(
predictions_raw, teacher_targets, rq_logits, rq_targets
)
self.log(
"val/loss", loss, on_step=False, on_epoch=True, prog_bar=True, batch_size=B
)
self.log("val/jepa_loss", jepa_loss, on_step=False, on_epoch=True, batch_size=B)
self.log("val/rq_loss", rq_loss, on_step=False, on_epoch=True, batch_size=B)
return loss
|