Dynamatrix/DiffBIR-OpenXLab
0
1"""SAMPLING ONLY."""2 3import torch4import numpy as np5from tqdm import tqdm6from functools import partial7 8from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like9from ldm.models.diffusion.sampling_util import norm_thresholding10 11 12class PLMSSampler(object):13 def __init__(self, model, schedule="linear", **kwargs):14 super().__init__()15 self.model = model16 self.ddpm_num_timesteps = model.num_timesteps17 self.schedule = schedule18 19 def register_buffer(self, name, attr):20 if type(attr) == torch.Tensor:21 if attr.device != torch.device("cuda"):22 attr = attr.to(torch.device("cuda"))23 setattr(self, name, attr)24 25 def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):26 if ddim_eta != 0:27 raise ValueError('ddim_eta must be 0 for PLMS')28 self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,29 num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)30 alphas_cumprod = self.model.alphas_cumprod31 assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'32 to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)33 34 self.register_buffer('betas', to_torch(self.model.betas))35 self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))36 self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))37 38 # calculations for diffusion q(x_t | x_{t-1}) and others39 self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))40 self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))41 self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))42 self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))43 self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))44 45 # ddim sampling parameters46 ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),47 ddim_timesteps=self.ddim_timesteps,48 eta=ddim_eta,verbose=verbose)49 self.register_buffer('ddim_sigmas', ddim_sigmas)50 self.register_buffer('ddim_alphas', ddim_alphas)51 self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)52 self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))53 sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(54 (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (55 1 - self.alphas_cumprod / self.alphas_cumprod_prev))56 self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)57 58 @torch.no_grad()59 def sample(self,60 S,61 batch_size,62 shape,63 conditioning=None,64 callback=None,65 normals_sequence=None,66 img_callback=None,67 quantize_x0=False,68 eta=0.,69 mask=None,70 x0=None,71 temperature=1.,72 noise_dropout=0.,73 score_corrector=None,74 corrector_kwargs=None,75 verbose=True,76 x_T=None,77 log_every_t=100,78 unconditional_guidance_scale=1.,79 unconditional_conditioning=None,80 # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...81 dynamic_threshold=None,82 **kwargs83 ):84 if conditioning is not None:85 if isinstance(conditioning, dict):86 cbs = conditioning[list(conditioning.keys())[0]].shape[0]87 if cbs != batch_size:88 print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")89 else:90 if conditioning.shape[0] != batch_size:91 print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")92 93 self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)94 # sampling95 C, H, W = shape96 size = (batch_size, C, H, W)97 print(f'Data shape for PLMS sampling is {size}')98 99 samples, intermediates = self.plms_sampling(conditioning, size,100 callback=callback,101 img_callback=img_callback,102 quantize_denoised=quantize_x0,103 mask=mask, x0=x0,104 ddim_use_original_steps=False,105 noise_dropout=noise_dropout,106 temperature=temperature,107 score_corrector=score_corrector,108 corrector_kwargs=corrector_kwargs,109 x_T=x_T,110 log_every_t=log_every_t,111 unconditional_guidance_scale=unconditional_guidance_scale,112 unconditional_conditioning=unconditional_conditioning,113 dynamic_threshold=dynamic_threshold,114 )115 return samples, intermediates116 117 @torch.no_grad()118 def plms_sampling(self, cond, shape,119 x_T=None, ddim_use_original_steps=False,120 callback=None, timesteps=None, quantize_denoised=False,121 mask=None, x0=None, img_callback=None, log_every_t=100,122 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,123 unconditional_guidance_scale=1., unconditional_conditioning=None,124 dynamic_threshold=None):125 device = self.model.betas.device126 b = shape[0]127 if x_T is None:128 img = torch.randn(shape, device=device)129 else:130 img = x_T131 132 if timesteps is None:133 timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps134 elif timesteps is not None and not ddim_use_original_steps:135 subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1136 timesteps = self.ddim_timesteps[:subset_end]137 138 intermediates = {'x_inter': [img], 'pred_x0': [img]}139 time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)140 total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]141 print(f"Running PLMS Sampling with {total_steps} timesteps")142 143 iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)144 old_eps = []145 146 for i, step in enumerate(iterator):147 index = total_steps - i - 1148 ts = torch.full((b,), step, device=device, dtype=torch.long)149 ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)150 151 if mask is not None:152 assert x0 is not None153 img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass?154 img = img_orig * mask + (1. - mask) * img155 156 outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,157 quantize_denoised=quantize_denoised, temperature=temperature,158 noise_dropout=noise_dropout, score_corrector=score_corrector,159 corrector_kwargs=corrector_kwargs,160 unconditional_guidance_scale=unconditional_guidance_scale,161 unconditional_conditioning=unconditional_conditioning,162 old_eps=old_eps, t_next=ts_next,163 dynamic_threshold=dynamic_threshold)164 img, pred_x0, e_t = outs165 old_eps.append(e_t)166 if len(old_eps) >= 4:167 old_eps.pop(0)168 if callback: callback(i)169 if img_callback: img_callback(pred_x0, i)170 171 if index % log_every_t == 0 or index == total_steps - 1:172 intermediates['x_inter'].append(img)173 intermediates['pred_x0'].append(pred_x0)174 175 return img, intermediates176 177 @torch.no_grad()178 def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,179 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,180 unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None,181 dynamic_threshold=None):182 b, *_, device = *x.shape, x.device183 184 def get_model_output(x, t):185 if unconditional_conditioning is None or unconditional_guidance_scale == 1.:186 e_t = self.model.apply_model(x, t, c)187 else:188 x_in = torch.cat([x] * 2)189 t_in = torch.cat([t] * 2)190 c_in = torch.cat([unconditional_conditioning, c])191 e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)192 e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)193 194 if score_corrector is not None:195 assert self.model.parameterization == "eps"196 e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)197 198 return e_t199 200 alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas201 alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev202 sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas203 sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas204 205 def get_x_prev_and_pred_x0(e_t, index):206 # select parameters corresponding to the currently considered timestep207 a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)208 a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)209 sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)210 sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)211 212 # current prediction for x_0213 pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()214 if quantize_denoised:215 pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)216 if dynamic_threshold is not None:217 pred_x0 = norm_thresholding(pred_x0, dynamic_threshold)218 # direction pointing to x_t219 dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t220 noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature221 if noise_dropout > 0.:222 noise = torch.nn.functional.dropout(noise, p=noise_dropout)223 x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise224 return x_prev, pred_x0225 226 e_t = get_model_output(x, t)227 if len(old_eps) == 0:228 # Pseudo Improved Euler (2nd order)229 x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)230 e_t_next = get_model_output(x_prev, t_next)231 e_t_prime = (e_t + e_t_next) / 2232 elif len(old_eps) == 1:233 # 2nd order Pseudo Linear Multistep (Adams-Bashforth)234 e_t_prime = (3 * e_t - old_eps[-1]) / 2235 elif len(old_eps) == 2:236 # 3nd order Pseudo Linear Multistep (Adams-Bashforth)237 e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12238 elif len(old_eps) >= 3:239 # 4nd order Pseudo Linear Multistep (Adams-Bashforth)240 e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24241 242 x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)243 244 return x_prev, pred_x0, e_t245