import torch import numpy as np from tqdm import tqdm from functools import partial from copy import deepcopy from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like class PLMSSampler(object): def __init__(self, diffusion, model, schedule="linear", alpha_generator_func=None, set_alpha_scale=None): super().__init__() self.diffusion = diffusion self.model = model self.device = diffusion.betas.device self.ddpm_num_timesteps = diffusion.num_timesteps self.schedule = schedule self.alpha_generator_func = alpha_generator_func self.set_alpha_scale = set_alpha_scale def register_buffer(self, name, attr): if type(attr) == torch.Tensor: attr = attr.to(self.device) setattr(self, name, attr) def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=False): if ddim_eta != 0: raise ValueError('ddim_eta must be 0 for PLMS') self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) alphas_cumprod = self.diffusion.alphas_cumprod assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.device) self.register_buffer('betas', to_torch(self.diffusion.betas)) self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) self.register_buffer('alphas_cumprod_prev', to_torch(self.diffusion.alphas_cumprod_prev)) # calculations for diffusion q(x_t | x_{t-1}) and others self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) # ddim sampling parameters ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), ddim_timesteps=self.ddim_timesteps, eta=ddim_eta,verbose=verbose) self.register_buffer('ddim_sigmas', ddim_sigmas) self.register_buffer('ddim_alphas', ddim_alphas) self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) @torch.no_grad() def sample(self, S, shape, input, uc=None, guidance_scale=1, mask=None, x0=None): self.make_schedule(ddim_num_steps=S) return self.plms_sampling(shape, input, uc, guidance_scale, mask=mask, x0=x0) @torch.no_grad() def plms_sampling(self, shape, input, uc=None, guidance_scale=1, mask=None, x0=None): b = shape[0] img = input["x"] if img == None: img = torch.randn(shape, device=self.device) input["x"] = img time_range = np.flip(self.ddim_timesteps) total_steps = self.ddim_timesteps.shape[0] old_eps = [] if self.alpha_generator_func != None: alphas = self.alpha_generator_func(len(time_range)) for i, step in enumerate(time_range): # set alpha if self.alpha_generator_func != None: self.set_alpha_scale(self.model, alphas[i]) # run index = total_steps - i - 1 ts = torch.full((b,), step, device=self.device, dtype=torch.long) ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=self.device, dtype=torch.long) if mask is not None: assert x0 is not None img_orig = self.diffusion.q_sample(x0, ts) img = img_orig * mask + (1. - mask) * img input["x"] = img img, pred_x0, e_t = self.p_sample_plms(input, ts, index=index, uc=uc, guidance_scale=guidance_scale, old_eps=old_eps, t_next=ts_next) input["x"] = img old_eps.append(e_t) if len(old_eps) >= 4: old_eps.pop(0) return img @torch.no_grad() def p_sample_plms(self, input, t, index, guidance_scale=1., uc=None, old_eps=None, t_next=None): x = deepcopy(input["x"]) b = x.shape[0] def get_model_output(input): e_t = self.model(input) if uc is not None and guidance_scale != 1: unconditional_input = dict(x=input["x"], timesteps=input["timesteps"], context=uc) if "inpainting_extra_input" in input: unconditional_input["inpainting_extra_input"] = input["inpainting_extra_input"] e_t_uncond = self.model( unconditional_input ) e_t = e_t_uncond + guidance_scale * (e_t - e_t_uncond) return e_t def get_x_prev_and_pred_x0(e_t, index): # select parameters corresponding to the currently considered timestep a_t = torch.full((b, 1, 1, 1), self.ddim_alphas[index], device=self.device) a_prev = torch.full((b, 1, 1, 1), self.ddim_alphas_prev[index], device=self.device) sigma_t = torch.full((b, 1, 1, 1), self.ddim_sigmas[index], device=self.device) sqrt_one_minus_at = torch.full((b, 1, 1, 1), self.ddim_sqrt_one_minus_alphas[index],device=self.device) # current prediction for x_0 pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() # direction pointing to x_t dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t noise = sigma_t * torch.randn_like(x) x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise return x_prev, pred_x0 input["timesteps"] = t e_t = get_model_output(input) if len(old_eps) == 0: # Pseudo Improved Euler (2nd order) x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index) input["x"] = x_prev input["timesteps"] = t_next e_t_next = get_model_output(input) e_t_prime = (e_t + e_t_next) / 2 elif len(old_eps) == 1: # 2nd order Pseudo Linear Multistep (Adams-Bashforth) e_t_prime = (3 * e_t - old_eps[-1]) / 2 elif len(old_eps) == 2: # 3nd order Pseudo Linear Multistep (Adams-Bashforth) e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12 elif len(old_eps) >= 3: # 4nd order Pseudo Linear Multistep (Adams-Bashforth) e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24 x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index) return x_prev, pred_x0, e_t