Diffusion 公式推导和代码实现

预测噪声

Diffusion model(一): 公式推导详解

生成噪声的公式：

预测噪声

1、训练代码

：噪声预测公式及代码

    def q_sample(self, x_start, t, noise = None):
        noise = default(noise, lambda: torch.randn_like(x_start))

        return (
            extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
            extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
        )

2、基础定义的公式

        alphas = 1. - betas
        alphas_cumprod = torch.cumprod(alphas, axis = 0)
        alphas_cumprod_prev = F.pad(alphas_cumprod[:-1], (1, 0), value = 1.)

        timesteps, = betas.shape
        self.num_timesteps = int(timesteps)

        if loss_type == 'l1':
            loss_fn = F.l1_loss
        elif loss_type == 'l2':
            loss_fn = F.mse_loss
        elif loss_type == 'huber':
            loss_fn = F.smooth_l1_loss
        else:
            raise NotImplementedError()

        self.loss_type = loss_type
        self.loss_fn = loss_fn

        # register buffer helper function to cast double back to float

        register_buffer = lambda name, val: self.register_buffer(name, val.to(torch.float32))

        register_buffer('betas', betas)
        register_buffer('alphas_cumprod', alphas_cumprod)
        register_buffer('alphas_cumprod_prev', alphas_cumprod_prev)

        # calculations for diffusion q(x_t | x_{t-1}) and others

        register_buffer('sqrt_alphas_cumprod', torch.sqrt(alphas_cumprod))
        register_buffer('sqrt_one_minus_alphas_cumprod', torch.sqrt(1. - alphas_cumprod))
        register_buffer('log_one_minus_alphas_cumprod', torch.log(1. - alphas_cumprod))
        register_buffer('sqrt_recip_alphas_cumprod', torch.sqrt(1. / alphas_cumprod))
        register_buffer('sqrt_recipm1_alphas_cumprod', torch.sqrt(1. / alphas_cumprod - 1))

        # calculations for posterior q(x_{t-1} | x_t, x_0)

        posterior_variance = betas * (1. - alphas_cumprod_prev) / (1. - alphas_cumprod)

        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)

        register_buffer('posterior_variance', posterior_variance)

        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain

        register_buffer('posterior_log_variance_clipped', torch.log(posterior_variance.clamp(min =1e-20)))
        register_buffer('posterior_mean_coef1', betas * torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
        register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev) * torch.sqrt(alphas) / (1. - alphas_cumprod))

        # p2 loss reweighting

        self.has_p2_loss_reweighting = p2_loss_weight_gamma > 0.
        register_buffer('p2_loss_weight', (p2_loss_weight_k + alphas_cumprod / (1 - alphas_cumprod)) ** -p2_loss_weight_gamma)

3、噪声采样：生成代码

pred = model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise

生成t-1的图像

4、噪声损失函数