File size: 10,707 Bytes
a3b520a | 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 | """Training script for the latent diffusion model."""
import copy
import csv
import sys
from pathlib import Path
import torch
from torch.utils.data import DataLoader, Dataset, random_split
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
sys.path.insert(0, str(Path(__file__).parent.parent))
from models.autoencoder import KickVAE
from models.diffusion import LatentUNet, NoiseScheduler
from models.text_encoder import KeywordEncoder, build_vocab
from training.config import DiffusionConfig
# ---------------------------------------------------------------------------
# Dataset
# ---------------------------------------------------------------------------
class LatentDataset(Dataset):
"""Dataset of pre-encoded VAE latents with keyword token IDs."""
def __init__(
self,
latents_dir: Path,
metadata_csv: Path,
vocab: list[str],
) -> None:
self.latent_files = sorted(
f for f in latents_dir.glob("*.pt") if not f.name.startswith("._")
)
if not self.latent_files:
raise FileNotFoundError(f"No .pt files in {latents_dir}")
# Build keyword lookup: filename_stem -> list of keyword strings
self.kw_to_idx = {kw: i for i, kw in enumerate(vocab)}
self.keywords: dict[str, list[int]] = {}
with open(metadata_csv) as f:
reader = csv.DictReader(f)
for row in reader:
stem = Path(row["filename"]).stem
ids = []
for kw in row["keywords"].split(","):
kw = kw.strip().lower()
if kw in self.kw_to_idx:
ids.append(self.kw_to_idx[kw])
self.keywords[stem] = ids
def __len__(self) -> int:
return len(self.latent_files)
def __getitem__(self, idx: int) -> tuple[torch.Tensor, list[int]]:
path = self.latent_files[idx]
latent = torch.load(path, weights_only=False)
# latent filename matches mel filename stem
stem = path.stem
token_ids = self.keywords.get(stem, [])
return latent, token_ids
def collate_fn(
batch: list[tuple[torch.Tensor, list[int]]],
) -> tuple[torch.Tensor, list[list[int]]]:
"""Custom collate to handle variable-length keyword lists."""
latents = torch.stack([b[0] for b in batch])
token_ids = [b[1] for b in batch]
return latents, token_ids
# ---------------------------------------------------------------------------
# Pre-encode latents
# ---------------------------------------------------------------------------
def pre_encode_latents(cfg: DiffusionConfig) -> None:
"""Encode all mel spectrograms to latents using frozen VAE."""
cfg.latents_dir.mkdir(parents=True, exist_ok=True)
# Check if already done
existing = list(cfg.latents_dir.glob("*.pt"))
if len(existing) > 100:
print(f"Latents dir already has {len(existing)} files, skipping encoding.")
return
device = torch.device(
"cuda" if torch.cuda.is_available()
else "mps" if torch.backends.mps.is_available()
else "cpu"
)
# Load VAE
checkpoint = torch.load(cfg.vae_checkpoint, weights_only=False)
vae = KickVAE(latent_dim=cfg.latent_dim).to(device)
vae.load_state_dict(checkpoint["model_state_dict"])
vae.eval()
mel_files = sorted(
f for f in cfg.data_dir.glob("*.pt") if not f.name.startswith("._")
)
print(f"Encoding {len(mel_files)} mel spectrograms to latents...")
with torch.no_grad():
for f in tqdm(mel_files):
out_path = cfg.latents_dir / f.name
if out_path.exists():
continue
mel = torch.load(f, weights_only=False).unsqueeze(0).to(device)
latent = vae.encode(mel).squeeze(0).cpu()
torch.save(latent, out_path)
print("Latent encoding complete.")
# ---------------------------------------------------------------------------
# EMA
# ---------------------------------------------------------------------------
class EMA:
"""Exponential moving average of model parameters."""
def __init__(self, model: torch.nn.Module, decay: float = 0.9999) -> None:
self.decay = decay
self.shadow = copy.deepcopy(model)
self.shadow.eval()
for p in self.shadow.parameters():
p.requires_grad_(False)
@torch.no_grad()
def update(self, model: torch.nn.Module) -> None:
for s, p in zip(self.shadow.parameters(), model.parameters()):
s.data.mul_(self.decay).add_(p.data, alpha=1 - self.decay)
# ---------------------------------------------------------------------------
# Training
# ---------------------------------------------------------------------------
def train(cfg: DiffusionConfig | None = None) -> None:
if cfg is None:
cfg = DiffusionConfig()
# Pre-encode latents
pre_encode_latents(cfg)
device = torch.device(
"cuda" if torch.cuda.is_available()
else "mps" if torch.backends.mps.is_available()
else "cpu"
)
print(f"Using device: {device}")
# Build vocab and dataset
vocab = build_vocab(cfg.metadata_csv)
print(f"Vocabulary size: {len(vocab)}")
dataset = LatentDataset(cfg.latents_dir, cfg.metadata_csv, vocab)
val_size = int(len(dataset) * cfg.val_split)
train_size = len(dataset) - val_size
train_set, val_set = random_split(
dataset, [train_size, val_size],
generator=torch.Generator().manual_seed(42),
)
train_loader = DataLoader(
train_set,
batch_size=cfg.batch_size,
shuffle=True,
num_workers=cfg.num_workers,
pin_memory=True,
collate_fn=collate_fn,
)
val_loader = DataLoader(
val_set,
batch_size=cfg.batch_size,
shuffle=False,
num_workers=cfg.num_workers,
pin_memory=True,
collate_fn=collate_fn,
)
print(f"Train: {train_size}, Val: {val_size}")
# Model
model = LatentUNet(
latent_dim=cfg.latent_dim,
base_channels=cfg.base_channels,
cond_dim=cfg.cond_dim,
).to(device)
text_enc = KeywordEncoder(
vocab_size=len(vocab),
embed_dim=cfg.text_embed_dim,
cond_dim=cfg.cond_dim,
).to(device)
scheduler = NoiseScheduler(cfg.timesteps, cfg.beta_start, cfg.beta_end).to(device)
ema = EMA(model, cfg.ema_decay)
optimizer = torch.optim.AdamW(
list(model.parameters()) + list(text_enc.parameters()),
lr=cfg.learning_rate,
)
scaler = torch.amp.GradScaler(enabled=cfg.use_amp and device.type == "cuda")
# Logging
cfg.log_dir.mkdir(parents=True, exist_ok=True)
cfg.checkpoint_dir.mkdir(parents=True, exist_ok=True)
writer = SummaryWriter(cfg.log_dir)
# Training loop (iteration-based)
global_step = 0
model.train()
text_enc.train()
print(f"Training for {cfg.iterations} iterations...")
while global_step < cfg.iterations:
for latents, token_ids in train_loader:
if global_step >= cfg.iterations:
break
latents = latents.to(device)
batch_size = latents.shape[0]
# Classifier-free guidance dropout: replace keywords with empty list
dropped_ids = []
for ids in token_ids:
if torch.rand(1).item() < cfg.cfg_dropout:
dropped_ids.append([])
else:
dropped_ids.append(ids)
# Sample timesteps and noise
t = torch.randint(0, cfg.timesteps, (batch_size,), device=device)
noise = torch.randn_like(latents)
noisy = scheduler.add_noise(latents, noise, t)
with torch.amp.autocast(
device_type=device.type,
enabled=cfg.use_amp and device.type == "cuda",
):
cond = text_enc(dropped_ids, device)
pred_noise = model(noisy, t, cond)
loss = torch.nn.functional.mse_loss(pred_noise, noise)
loss = loss / cfg.gradient_accumulation
scaler.scale(loss).backward()
if (global_step + 1) % cfg.gradient_accumulation == 0:
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
ema.update(model)
# Logging
if global_step % 50 == 0:
writer.add_scalar(
"train/loss", loss.item() * cfg.gradient_accumulation, global_step
)
if global_step % 500 == 0:
print(
f"Step {global_step}/{cfg.iterations} "
f"loss={loss.item() * cfg.gradient_accumulation:.6f}"
)
# Validation
if global_step % 1000 == 0 and global_step > 0:
model.eval()
text_enc.eval()
val_loss_sum = 0.0
val_count = 0
with torch.no_grad():
for vl, vt in val_loader:
vl = vl.to(device)
vt_step = torch.randint(
0, cfg.timesteps, (vl.shape[0],), device=device
)
vn = torch.randn_like(vl)
vnoisy = scheduler.add_noise(vl, vn, vt_step)
vcond = text_enc(vt, device)
vpred = model(vnoisy, vt_step, vcond)
val_loss_sum += torch.nn.functional.mse_loss(vpred, vn).item()
val_count += 1
avg_val = val_loss_sum / max(val_count, 1)
writer.add_scalar("val/loss", avg_val, global_step)
print(f" val_loss={avg_val:.6f}")
model.train()
text_enc.train()
# Checkpoint
if (global_step + 1) % cfg.checkpoint_every == 0:
path = cfg.checkpoint_dir / f"diffusion_step_{global_step+1}.pt"
torch.save({
"step": global_step + 1,
"model_state_dict": model.state_dict(),
"ema_state_dict": ema.shadow.state_dict(),
"text_enc_state_dict": text_enc.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"vocab": vocab,
"config": cfg,
}, path)
print(f"Saved checkpoint: {path}")
global_step += 1
writer.close()
print("Diffusion training complete.")
if __name__ == "__main__":
train()
|