lrf/train_v2.py

"""
LRF v2 Training on CIFAR-10 with pre-trained TAESD VAE.

This script:
1. Loads TAESD (pre-trained, frozen) as the image encoder/decoder
2. Pre-computes all CIFAR-10 latents (fast, ~30s)
3. Trains the RecursiveLatentCore denoiser on real latents
4. Generates real images and saves them
"""

import os
import sys
import time
import json
import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader, TensorDataset
import torchvision
import torchvision.transforms as T
import numpy as np
from pathlib import Path

sys.path.insert(0, '/app')
from lrf.model_v2 import LRFv2, RectifiedFlowScheduler


def load_taesd(device='cpu'):
    """Load pre-trained TAESD VAE."""
    from diffusers import AutoencoderTiny
    vae = AutoencoderTiny.from_pretrained('madebyollin/taesd', torch_dtype=torch.float32)
    vae.eval()
    vae.to(device)
    for p in vae.parameters():
        p.requires_grad_(False)
    return vae


def precompute_latents(vae, dataset, batch_size=64, device='cpu'):
    """Pre-compute all latent representations. Much faster than encoding on-the-fly."""
    loader = DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=0)
    
    all_latents = []
    all_labels = []
    
    total = len(loader)
    print(f"Pre-computing latents for {len(dataset)} images ({total} batches)...", flush=True)
    t0 = time.time()
    
    with torch.no_grad():
        for batch_idx, (images, labels) in enumerate(loader):
            images = images.to(device)
            latents = vae.encode(images).latents
            all_latents.append(latents.cpu())
            all_labels.append(labels)
            if (batch_idx + 1) % 50 == 0 or batch_idx == 0:
                elapsed = time.time() - t0
                print(f"  Batch {batch_idx+1}/{total} ({elapsed:.0f}s)", flush=True)
    
    all_latents = torch.cat(all_latents, dim=0)
    all_labels = torch.cat(all_labels, dim=0)
    
    dt = time.time() - t0
    print(f"Done in {dt:.1f}s. Latent shape: {all_latents.shape}", flush=True)
    print(f"Latent stats: mean={all_latents.mean():.4f}, std={all_latents.std():.4f}, "
          f"min={all_latents.min():.4f}, max={all_latents.max():.4f}", flush=True)
    
    return all_latents, all_labels


def train_denoiser(
    config=None,
    num_epochs=50,
    batch_size=128,
    lr=2e-4,
    device='cpu',
    output_dir='/app/lrf_v2_output',
    save_every=10,
):
    """Train the LRF denoiser on CIFAR-10 latents."""
    os.makedirs(output_dir, exist_ok=True)
    
    print("=" * 60)
    print("LRF v2 - Training on CIFAR-10")
    print("=" * 60)
    
    # 1. Load TAESD
    print("\n[Step 1] Loading TAESD VAE...")
    vae = load_taesd(device)
    print(f"  TAESD loaded: {sum(p.numel() for p in vae.parameters()):,} params (frozen)")
    
    # 2. Load CIFAR-10
    print("\n[Step 2] Loading CIFAR-10...")
    transform = T.Compose([
        T.ToTensor(),
        T.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),  # -> [-1, 1]
    ])
    
    # CIFAR-10 is 32x32, TAESD expects multiples of 8 -> resize to 32
    # Actually TAESD works on any size, 32x32 -> 4x4 latent (f=8)
    trainset = torchvision.datasets.CIFAR10(
        root='/app/data', train=True, download=True, transform=transform,
    )
    testset = torchvision.datasets.CIFAR10(
        root='/app/data', train=False, download=True, transform=transform,
    )
    print(f"  Train: {len(trainset)}, Test: {len(testset)}")
    print(f"  Image size: {trainset[0][0].shape}")
    
    # 3. Pre-compute latents (or load from cache)
    print("\n[Step 3] Pre-computing latents...", flush=True)
    cache_path = os.path.join(output_dir, 'latent_cache.pt')
    if os.path.exists(cache_path):
        print("  Loading cached latents...", flush=True)
        cache = torch.load(cache_path, weights_only=True)
        train_latents = cache['train_latents']
        train_labels = cache['train_labels']
        test_latents = cache['test_latents']
        test_labels = cache['test_labels']
        print(f"  Loaded from cache. Train: {train_latents.shape}, Test: {test_latents.shape}", flush=True)
    else:
        train_latents, train_labels = precompute_latents(vae, trainset, batch_size=256, device=device)
        test_latents, test_labels = precompute_latents(vae, testset, batch_size=256, device=device)
        torch.save({
            'train_latents': train_latents, 'train_labels': train_labels,
            'test_latents': test_latents, 'test_labels': test_labels,
        }, cache_path)
        print(f"  Cached latents to {cache_path}", flush=True)
    
    # Verify VAE reconstruction works
    print("\n[Step 3b] Verifying VAE reconstruction...")
    with torch.no_grad():
        sample_imgs = torch.stack([trainset[i][0] for i in range(8)]).to(device)
        sample_lats = vae.encode(sample_imgs).latents
        sample_recs = vae.decode(sample_lats).sample
        recon_mse = F.mse_loss(sample_recs, sample_imgs).item()
        print(f"  VAE reconstruction MSE on real images: {recon_mse:.4f}")
        
        # Save reconstruction grid
        save_image_grid(
            torch.cat([sample_imgs[:4], sample_recs[:4]], dim=0),
            os.path.join(output_dir, 'vae_reconstruction.png'),
            nrow=4, title='Top: Original, Bottom: TAESD Reconstruction'
        )
        print(f"  Saved VAE reconstruction grid to {output_dir}/vae_reconstruction.png")
    
    # Normalize latents for better training
    lat_mean = train_latents.mean()
    lat_std = train_latents.std()
    print(f"\n  Latent mean: {lat_mean:.4f}, std: {lat_std:.4f}")
    # Scale latents to roughly unit variance
    latent_scale = lat_std.item()
    
    # Create dataset of (latent, label)
    train_ds = TensorDataset(train_latents, train_labels)
    train_loader = DataLoader(train_ds, batch_size=batch_size, shuffle=True, 
                              num_workers=0, drop_last=True)
    
    # 4. Create model
    print("\n[Step 4] Creating LRF denoiser...")
    config = config or LRFv2.small_config()
    config['latent_ch'] = train_latents.shape[1]  # Should be 4
    model = LRFv2(config).to(device)
    params = model.count_params()
    print(f"  Config: dim={config['dim']}, blocks={config['num_blocks']}, "
          f"T_inner={config['T_inner']}, T_outer={config['T_outer']}")
    print(f"  Parameters: {params['total']:,} total, {params['core']:,} core")
    print(f"  Effective depth: {config['T_outer'] * config['T_inner'] * config['num_blocks']} layers "
          f"from {config['num_blocks']} blocks")
    
    # 5. Training
    print(f"\n[Step 5] Training for {num_epochs} epochs...")
    scheduler = RectifiedFlowScheduler()
    optimizer = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=0.01, betas=(0.9, 0.95))
    
    # Cosine annealing
    lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
        optimizer, T_max=num_epochs * len(train_loader), eta_min=lr * 0.01
    )
    
    # EMA for stable sampling
    ema_decay = 0.999
    ema_params = {name: p.clone().detach() for name, p in model.named_parameters()}
    
    loss_history = []
    best_loss = float('inf')
    
    for epoch in range(num_epochs):
        model.train()
        epoch_loss = 0.0
        num_batches = 0
        
        for latents, labels in train_loader:
            latents = latents.to(device)
            labels = labels.to(device)
            B = latents.shape[0]
            
            # Sample timesteps and noise
            t = scheduler.sample_timesteps(B, device)
            noise = torch.randn_like(latents)
            
            # Create noisy latent
            z_t = scheduler.add_noise(latents, noise, t)
            
            # Predict velocity (with 10% CFG dropout)
            v_pred = model.predict_velocity(z_t, t, labels, cfg_dropout=0.1)
            
            # Velocity target
            v_target = scheduler.get_velocity_target(latents, noise)
            
            # MSE loss with min-SNR weighting
            loss_per_sample = (v_pred - v_target).pow(2).mean(dim=[1, 2, 3])
            
            # SNR weighting: upweight middle timesteps
            w = 1.0 / (t * (1 - t) + 0.01)
            w = w / w.mean()
            loss = (loss_per_sample * w).mean()
            
            optimizer.zero_grad()
            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
            optimizer.step()
            lr_scheduler.step()
            
            # EMA update
            with torch.no_grad():
                for name, p in model.named_parameters():
                    ema_params[name].mul_(ema_decay).add_(p, alpha=1 - ema_decay)
            
            epoch_loss += loss.item()
            num_batches += 1
        
        avg_loss = epoch_loss / num_batches
        loss_history.append(avg_loss)
        
        if avg_loss < best_loss:
            best_loss = avg_loss
        
        if (epoch + 1) % 5 == 0 or epoch == 0:
            current_lr = optimizer.param_groups[0]['lr']
            print(f"  Epoch {epoch+1:3d}/{num_epochs}: loss={avg_loss:.4f}, "
                  f"best={best_loss:.4f}, lr={current_lr:.2e}", flush=True)
        
        # Save and generate samples periodically
        if (epoch + 1) % save_every == 0 or epoch == num_epochs - 1:
            # Swap to EMA for sampling
            saved_params = {}
            with torch.no_grad():
                for name, p in model.named_parameters():
                    saved_params[name] = p.clone()
                    p.copy_(ema_params[name])
            
            # Generate samples
            model.eval()
            samples = generate_samples(model, vae, scheduler, device, 
                                       num_samples=16, num_steps=10, cfg_scale=2.0)
            
            save_image_grid(
                samples, 
                os.path.join(output_dir, f'samples_epoch{epoch+1:03d}.png'),
                nrow=8, title=f'Epoch {epoch+1}, Loss={avg_loss:.4f}'
            )
            
            # Restore original params
            with torch.no_grad():
                for name, p in model.named_parameters():
                    p.copy_(saved_params[name])
            
            # Save checkpoint
            torch.save({
                'model_state': model.state_dict(),
                'ema_params': ema_params,
                'config': config,
                'epoch': epoch + 1,
                'loss': avg_loss,
                'latent_scale': latent_scale,
                'loss_history': loss_history,
            }, os.path.join(output_dir, 'checkpoint.pt'))
    
    # Final generation with EMA
    with torch.no_grad():
        for name, p in model.named_parameters():
            p.copy_(ema_params[name])
    
    model.eval()
    
    # Generate class-conditional samples
    print("\n[Step 6] Generating final samples...")
    cifar_classes = ['airplane', 'automobile', 'bird', 'cat', 'deer',
                     'dog', 'frog', 'horse', 'ship', 'truck']
    
    all_samples = []
    for cls_idx in range(10):
        samples = generate_samples(model, vae, scheduler, device,
                                   num_samples=4, num_steps=50, cfg_scale=3.0,
                                   class_label=cls_idx)
        all_samples.append(samples)
    
    all_samples = torch.cat(all_samples, dim=0)
    save_image_grid(
        all_samples,
        os.path.join(output_dir, 'final_class_conditional.png'),
        nrow=4, title='Class-conditional generation (rows: airplane, auto, bird, cat, deer, dog, frog, horse, ship, truck)'
    )
    
    # Save loss plot
    save_loss_plot(loss_history, os.path.join(output_dir, 'loss.png'))
    
    # Save config
    with open(os.path.join(output_dir, 'config.json'), 'w') as f:
        json.dump(config, f, indent=2)
    
    print(f"\n{'='*60}")
    print(f"Training complete! Best loss: {best_loss:.4f}")
    print(f"Output directory: {output_dir}")
    print(f"{'='*60}")
    
    return model, vae, loss_history


def generate_samples(model, vae, scheduler, device, num_samples=8, 
                     num_steps=20, cfg_scale=2.0, class_label=None):
    """Generate images from the model."""
    model.eval()
    
    # Latent shape for CIFAR-10: [B, 4, 4, 4] (32x32 image, f=8)
    shape = (num_samples, 4, 4, 4)
    
    if class_label is not None:
        labels = torch.full((num_samples,), class_label, dtype=torch.long, device=device)
    else:
        labels = torch.randint(0, 10, (num_samples,), device=device)
    
    z = scheduler.sample(model, shape, labels, num_steps=num_steps, 
                         cfg_scale=cfg_scale, device=device)
    
    # Decode through TAESD
    with torch.no_grad():
        images = vae.decode(z.to(device)).sample
    
    return images.clamp(-1, 1).cpu()


def save_image_grid(images, path, nrow=8, title=''):
    """Save image grid using torchvision."""
    # Convert from [-1,1] to [0,1]
    images = (images + 1) / 2
    images = images.clamp(0, 1)
    
    grid = torchvision.utils.make_grid(images, nrow=nrow, padding=2, normalize=False)
    
    # Save using PIL
    from PIL import Image
    grid_np = grid.permute(1, 2, 0).numpy()
    grid_np = (grid_np * 255).astype(np.uint8)
    img = Image.fromarray(grid_np)
    img.save(path)


def save_loss_plot(losses, path):
    """Save loss curve."""
    try:
        import matplotlib
        matplotlib.use('Agg')
        import matplotlib.pyplot as plt
        
        plt.figure(figsize=(10, 4))
        plt.plot(losses)
        plt.xlabel('Epoch')
        plt.ylabel('Loss')
        plt.title('Training Loss')
        plt.grid(True, alpha=0.3)
        plt.savefig(path, dpi=100, bbox_inches='tight')
        plt.close()
    except ImportError:
        print("matplotlib not available, skipping loss plot")


if __name__ == '__main__':
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    print(f"Device: {device}")
    
    train_denoiser(
        config=LRFv2.fast_config(),
        num_epochs=30,
        batch_size=64,
        lr=3e-4,
        device=device,
        output_dir='/app/lrf_v2_output',
        save_every=5,
    )