loss.py

import torch
import torch.nn.functional as F

def total_variation_loss(x):
    """Total variation regularization"""
    batch_size = x.size(0)
    h_tv = torch.abs(x[:, :, 1:, :] - x[:, :, :-1, :]).sum()
    w_tv = torch.abs(x[:, :, :, 1:] - x[:, :, :, :-1]).sum()
    return (h_tv + w_tv) / batch_size

def gradient_loss(x):
    """Sobel gradient loss"""
    sobel_x = torch.tensor([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], dtype=torch.float32, device=x.device).view(1, 1, 3, 3)
    sobel_y = torch.tensor([[-1, -2, -1], [0, 0, 0], [1, 2, 1]], dtype=torch.float32, device=x.device).view(1, 1, 3, 3)
    
    grad_x = F.conv2d(x, sobel_x.repeat(x.size(1), 1, 1, 1), padding=1, groups=x.size(1))
    grad_y = F.conv2d(x, sobel_y.repeat(x.size(1), 1, 1, 1), padding=1, groups=x.size(1))
    
    return torch.mean(grad_x**2 + grad_y**2)

def diffusion_loss(model, x0, t, noise_scheduler, config):
    xt, noise = noise_scheduler.apply_noise(x0, t)  # Get both noisy image and noise
    pred_noise = model(xt, t)
    
    # MSE loss between predicted noise and actual noise
    mse_loss = F.mse_loss(pred_noise, noise)
    
    # Re-enable regularization with very small weights since base training is stable
    tv_loss = total_variation_loss(xt)
    grad_loss = gradient_loss(xt)
    
    # Very small regularization weights to preserve the good training dynamics
    total_loss = mse_loss + config.tv_weight * tv_loss + 0.001 * grad_loss
    
    # Debug: check for extreme values
    if torch.isnan(total_loss) or total_loss > 1e6:
        print(f"WARNING: Extreme loss detected!")
        print(f"MSE: {mse_loss.item():.4f}, TV: {tv_loss.item():.4f}, Grad: {grad_loss.item():.4f}")
        print(f"Noise range: [{noise.min().item():.4f}, {noise.max().item():.4f}]")
        print(f"Pred range: [{pred_noise.min().item():.4f}, {pred_noise.max().item():.4f}]")
    
    return total_loss