File size: 14,826 Bytes
c4858e4 | 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 | """
LiquidGen Training Pipeline
Flow Matching training objective (velocity prediction):
- Forward: x_t = (1 - t) * x_0 + t * ε (linear interpolation)
- Target: v = ε - x_0 (velocity)
- Loss: MSE(model(x_t, t), v)
At inference: solve ODE from t=1 (noise) to t=0 (clean) using Euler steps.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset
from torch.amp import autocast, GradScaler
import math
import os
import json
import time
from pathlib import Path
from typing import Optional, Dict, Any
from dataclasses import dataclass, field, asdict
@dataclass
class TrainConfig:
"""Training configuration with sensible defaults for Colab free tier."""
# Model
model_size: str = "small"
num_classes: int = 0
class_drop_prob: float = 0.1
# Data
image_size: int = 256
dataset_name: str = "huggan/wikiart"
dataset_config: str = ""
image_column: str = "image"
label_column: str = ""
# VAE
vae_id: str = "black-forest-labs/FLUX.1-schnell"
vae_subfolder: str = "vae"
vae_dtype: str = "float16"
vae_scaling_factor: float = 0.3611
vae_shift_factor: float = 0.1159
# Training
batch_size: int = 8
gradient_accumulation_steps: int = 4
learning_rate: float = 1e-4
weight_decay: float = 0.01
max_grad_norm: float = 2.0
num_epochs: int = 100
warmup_steps: int = 1000
ema_decay: float = 0.9999
mixed_precision: bool = True
# Flow matching
min_timestep: float = 0.001
max_timestep: float = 0.999
# Saving
output_dir: str = "./outputs"
save_every_n_steps: int = 5000
sample_every_n_steps: int = 1000
log_every_n_steps: int = 50
# Sampling
num_sample_steps: int = 50
cfg_scale: float = 1.5
num_samples: int = 4
# System
seed: int = 42
num_workers: int = 2
pin_memory: bool = True
compile_model: bool = False
# Hub
push_to_hub: bool = False
hub_model_id: str = ""
def get_model_config(size: str, num_classes: int = 0, class_drop_prob: float = 0.1) -> dict:
"""Get model kwargs for a given size preset."""
configs = {
"small": dict(embed_dim=512, depth=12, spatial_kernel=7, scan_kernel=31,
expand_ratio=2.0, mlp_ratio=3.0),
"base": dict(embed_dim=640, depth=18, spatial_kernel=7, scan_kernel=31,
expand_ratio=2.0, mlp_ratio=4.0),
"large": dict(embed_dim=768, depth=24, spatial_kernel=7, scan_kernel=31,
expand_ratio=2.5, mlp_ratio=4.0),
}
cfg = configs[size]
cfg["num_classes"] = num_classes
cfg["class_drop_prob"] = class_drop_prob
cfg["use_zigzag"] = True
return cfg
class EMAModel:
"""Exponential Moving Average of model parameters."""
def __init__(self, model: nn.Module, decay: float = 0.9999):
self.decay = decay
self.shadow = {name: p.clone().detach() for name, p in model.named_parameters() if p.requires_grad}
@torch.no_grad()
def update(self, model: nn.Module):
for name, p in model.named_parameters():
if p.requires_grad and name in self.shadow:
self.shadow[name].mul_(self.decay).add_(p.data, alpha=1 - self.decay)
def apply(self, model: nn.Module):
self.backup = {name: p.data.clone() for name, p in model.named_parameters() if p.requires_grad}
for name, p in model.named_parameters():
if p.requires_grad and name in self.shadow:
p.data.copy_(self.shadow[name])
def restore(self, model: nn.Module):
for name, p in model.named_parameters():
if p.requires_grad and name in self.backup:
p.data.copy_(self.backup[name])
self.backup = {}
def state_dict(self):
return self.shadow
def load_state_dict(self, state_dict):
self.shadow = state_dict
class FlowMatchingScheduler:
"""
Flow Matching scheduler for training and sampling.
Training: x_t = (1-t)*x_0 + t*ε, v_target = ε - x_0
Sampling: Euler ODE from t=1 (noise) to t=0 (clean)
"""
def __init__(self, min_t: float = 0.001, max_t: float = 0.999):
self.min_t = min_t
self.max_t = max_t
def sample_timesteps(self, batch_size: int, device: torch.device) -> torch.Tensor:
return torch.rand(batch_size, device=device) * (self.max_t - self.min_t) + self.min_t
def add_noise(self, x0: torch.Tensor, noise: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
t_expand = t.view(-1, 1, 1, 1)
return (1 - t_expand) * x0 + t_expand * noise
def get_velocity_target(self, x0: torch.Tensor, noise: torch.Tensor) -> torch.Tensor:
return noise - x0
@torch.no_grad()
def sample(
self, model: nn.Module, shape: tuple, device: torch.device,
num_steps: int = 50, class_labels: Optional[torch.Tensor] = None,
cfg_scale: float = 1.0, dtype: torch.dtype = torch.float32,
) -> torch.Tensor:
model.eval()
x = torch.randn(shape, device=device, dtype=dtype)
dt = 1.0 / num_steps
times = torch.linspace(1.0, dt, num_steps, device=device)
for t_val in times:
t = torch.full((shape[0],), t_val.item(), device=device, dtype=dtype)
if cfg_scale > 1.0 and class_labels is not None:
with torch.amp.autocast('cuda', enabled=(dtype != torch.float32)):
v_cond = model(x, t, class_labels)
v_uncond = model(x, t, torch.zeros_like(class_labels))
v = v_uncond + cfg_scale * (v_cond - v_uncond)
else:
with torch.amp.autocast('cuda', enabled=(dtype != torch.float32)):
v = model(x, t, class_labels)
x = x - dt * v
return x
def get_cosine_schedule_with_warmup(optimizer, warmup_steps, total_steps):
"""Cosine LR schedule with linear warmup."""
def lr_lambda(current_step):
if current_step < warmup_steps:
return float(current_step) / float(max(1, warmup_steps))
progress = float(current_step - warmup_steps) / float(max(1, total_steps - warmup_steps))
return max(0.0, 0.5 * (1.0 + math.cos(math.pi * progress)))
return torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)
@torch.no_grad()
def encode_images_with_vae(images, vae, scaling_factor, shift_factor):
"""Encode pixel images to VAE latents."""
images = images * 2.0 - 1.0
latents = vae.encode(images).latent_dist.sample()
latents = (latents - shift_factor) * scaling_factor
return latents
@torch.no_grad()
def decode_latents_with_vae(latents, vae, scaling_factor, shift_factor):
"""Decode VAE latents to pixel images."""
latents = latents / scaling_factor + shift_factor
images = vae.decode(latents).sample
images = (images + 1.0) / 2.0
return images.clamp(0, 1)
def train(config: TrainConfig):
"""Main training loop."""
from model import LiquidGen
torch.manual_seed(config.seed)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Device: {device}")
os.makedirs(config.output_dir, exist_ok=True)
os.makedirs(os.path.join(config.output_dir, "samples"), exist_ok=True)
os.makedirs(os.path.join(config.output_dir, "checkpoints"), exist_ok=True)
with open(os.path.join(config.output_dir, "config.json"), "w") as f:
json.dump(asdict(config), f, indent=2)
# Load VAE
print("Loading VAE...")
from diffusers import AutoencoderKL
vae_dtype = torch.float16 if config.vae_dtype == "float16" else torch.bfloat16
vae = AutoencoderKL.from_pretrained(
config.vae_id, subfolder=config.vae_subfolder, torch_dtype=vae_dtype
).to(device).eval()
for p in vae.parameters():
p.requires_grad_(False)
# Load Dataset
print(f"Loading dataset: {config.dataset_name}")
from datasets import load_dataset
from torchvision import transforms
ds_kwargs = {}
if config.dataset_config:
ds_kwargs["name"] = config.dataset_config
dataset = load_dataset(config.dataset_name, split="train", **ds_kwargs)
transform = transforms.Compose([
transforms.Resize(config.image_size, interpolation=transforms.InterpolationMode.LANCZOS),
transforms.CenterCrop(config.image_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
])
class ImageDataset(Dataset):
def __init__(self, hf_dataset, transform, image_col, label_col=""):
self.dataset = hf_dataset
self.transform = transform
self.image_col = image_col
self.label_col = label_col
def __len__(self):
return len(self.dataset)
def __getitem__(self, idx):
item = self.dataset[idx]
img = item[self.image_col]
if img.mode != "RGB":
img = img.convert("RGB")
img = self.transform(img)
label = item[self.label_col] if self.label_col and self.label_col in item else -1
return img, label
train_dataset = ImageDataset(dataset, transform, config.image_column, config.label_column)
train_loader = DataLoader(
train_dataset, batch_size=config.batch_size, shuffle=True,
num_workers=config.num_workers, pin_memory=config.pin_memory, drop_last=True,
)
# Create Model
model_kwargs = get_model_config(config.model_size, config.num_classes, config.class_drop_prob)
model = LiquidGen(**model_kwargs).to(device)
print(f"LiquidGen-{config.model_size}: {model.count_params() / 1e6:.1f}M params")
if config.compile_model and hasattr(torch, "compile"):
model = torch.compile(model)
optimizer = torch.optim.AdamW(model.parameters(), lr=config.learning_rate,
weight_decay=config.weight_decay, betas=(0.9, 0.999))
total_steps = len(train_loader) * config.num_epochs // config.gradient_accumulation_steps
scheduler = get_cosine_schedule_with_warmup(optimizer, config.warmup_steps, total_steps)
ema = EMAModel(model, decay=config.ema_decay)
scaler = GradScaler('cuda', enabled=config.mixed_precision)
fm = FlowMatchingScheduler(min_t=config.min_timestep, max_t=config.max_timestep)
print(f"\nTraining: {total_steps} steps, effective batch {config.batch_size * config.gradient_accumulation_steps}")
global_step = 0
loss_accum = 0.0
for epoch in range(config.num_epochs):
model.train()
t_start = time.time()
for batch_idx, (images, labels) in enumerate(train_loader):
images = images.to(device)
labels = labels.to(device) if config.num_classes > 0 else None
with torch.no_grad():
latents = encode_images_with_vae(
images.to(vae_dtype), vae, config.vae_scaling_factor, config.vae_shift_factor
).float()
t = fm.sample_timesteps(latents.shape[0], device)
noise = torch.randn_like(latents)
x_t = fm.add_noise(latents, noise, t)
v_target = fm.get_velocity_target(latents, noise)
with autocast('cuda', enabled=config.mixed_precision):
v_pred = model(x_t, t, labels)
loss = F.mse_loss(v_pred, v_target) / config.gradient_accumulation_steps
scaler.scale(loss).backward()
loss_accum += loss.item()
if (batch_idx + 1) % config.gradient_accumulation_steps == 0:
scaler.unscale_(optimizer)
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), config.max_grad_norm)
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
scheduler.step()
ema.update(model)
global_step += 1
if global_step % config.log_every_n_steps == 0:
avg_loss = loss_accum / config.log_every_n_steps
lr = optimizer.param_groups[0]["lr"]
print(f"step={global_step} | epoch={epoch} | loss={avg_loss:.4f} | "
f"grad_norm={grad_norm:.2f} | lr={lr:.2e}")
loss_accum = 0.0
if math.isnan(avg_loss) or avg_loss > 100:
print("⚠️ Training diverged!")
return
if global_step % config.sample_every_n_steps == 0:
ema.apply(model)
model.eval()
latent_size = config.image_size // 8
sample_labels = None
if config.num_classes > 0:
sample_labels = torch.randint(0, config.num_classes, (config.num_samples,), device=device)
sampled = fm.sample(model, (config.num_samples, 16, latent_size, latent_size),
device, config.num_sample_steps, sample_labels, config.cfg_scale)
sample_imgs = decode_latents_with_vae(sampled.to(vae_dtype), vae,
config.vae_scaling_factor, config.vae_shift_factor).float()
from torchvision.utils import save_image
save_image(sample_imgs, os.path.join(config.output_dir, "samples", f"step_{global_step:07d}.png"), nrow=2)
ema.restore(model)
model.train()
if global_step % config.save_every_n_steps == 0:
torch.save({
"model": model.state_dict(), "ema": ema.state_dict(),
"optimizer": optimizer.state_dict(), "scheduler": scheduler.state_dict(),
"global_step": global_step, "epoch": epoch, "config": asdict(config),
}, os.path.join(config.output_dir, "checkpoints", f"step_{global_step:07d}.pt"))
print(f"Epoch {epoch} complete | time={time.time()-t_start:.0f}s")
torch.save({"model": model.state_dict(), "ema": ema.state_dict(), "config": asdict(config),
"global_step": global_step}, os.path.join(config.output_dir, "checkpoints", "final.pt"))
print(f"Training complete! Final model saved.")
if __name__ == "__main__":
config = TrainConfig(model_size="small", image_size=256, batch_size=4, num_epochs=2)
train(config)
|