| | import math
|
| | import torch
|
| | from torch import device, nn, einsum
|
| | import torch.nn.functional as F
|
| | from inspect import isfunction
|
| | from functools import partial
|
| | import numpy as np
|
| | from tqdm import tqdm
|
| |
|
| |
|
| | def _warmup_beta(linear_start, linear_end, n_timestep, warmup_frac):
|
| | betas = linear_end * np.ones(n_timestep, dtype=np.float64)
|
| | warmup_time = int(n_timestep * warmup_frac)
|
| | betas[:warmup_time] = np.linspace(
|
| | linear_start, linear_end, warmup_time, dtype=np.float64)
|
| | return betas
|
| |
|
| |
|
| | def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
|
| | if schedule == 'quad':
|
| | betas = np.linspace(linear_start ** 0.5, linear_end ** 0.5,
|
| | n_timestep, dtype=np.float64) ** 2
|
| | elif schedule == 'linear':
|
| | betas = np.linspace(linear_start, linear_end,
|
| | n_timestep, dtype=np.float64)
|
| | elif schedule == 'warmup10':
|
| | betas = _warmup_beta(linear_start, linear_end,
|
| | n_timestep, 0.1)
|
| | elif schedule == 'warmup50':
|
| | betas = _warmup_beta(linear_start, linear_end,
|
| | n_timestep, 0.5)
|
| | elif schedule == 'const':
|
| | betas = linear_end * np.ones(n_timestep, dtype=np.float64)
|
| | elif schedule == 'jsd':
|
| | betas = 1. / np.linspace(n_timestep,
|
| | 1, n_timestep, dtype=np.float64)
|
| | elif schedule == "cosine":
|
| | print('======================adopting cosine scheduler========================')
|
| |
|
| | timesteps = (
|
| | torch.arange(n_timestep + 1, dtype=torch.float64) /
|
| | n_timestep + cosine_s
|
| | )
|
| | alphas = timesteps / (1 + cosine_s) * math.pi / 2
|
| | alphas = torch.cos(alphas).pow(2)
|
| | alphas = alphas / alphas[0]
|
| | betas = 1 - alphas[1:] / alphas[:-1]
|
| | betas = betas.clamp(max=0.999)
|
| | else:
|
| | raise NotImplementedError(schedule)
|
| | return betas
|
| |
|
| |
|
| |
|
| |
|
| | def exists(x):
|
| | return x is not None
|
| |
|
| |
|
| | def default(val, d):
|
| | if exists(val):
|
| | return val
|
| | return d() if isfunction(d) else d
|
| |
|
| |
|
| | class GaussianDiffusion(nn.Module):
|
| | def __init__(
|
| | self,
|
| | denoise_fn,
|
| | image_size,
|
| | channels=3,
|
| | loss_type='l1',
|
| | conditional=True,
|
| | schedule_opt=None,
|
| | xT_noise_r=0.1,
|
| | seed = 1,
|
| | opt=None
|
| |
|
| | ):
|
| | super().__init__()
|
| | self.lq_noiselevel_val = schedule_opt["lq_noiselevel"]
|
| | self.opt = opt
|
| | self.channels = channels
|
| | self.image_size = image_size
|
| | self.denoise_fn = denoise_fn
|
| | self.loss_type = loss_type
|
| | self.conditional = conditional
|
| | self.ddim = schedule_opt['ddim']
|
| | self.xT_noise_r = xT_noise_r
|
| | self.seed = seed
|
| | if schedule_opt is not None:
|
| | pass
|
| |
|
| |
|
| | def set_loss(self, device):
|
| | if self.loss_type == 'l1':
|
| | self.loss_func = nn.L1Loss(reduction='sum').to(device)
|
| | elif self.loss_type == 'l2':
|
| | self.loss_func = nn.MSELoss(reduction='sum').to(device)
|
| | else:
|
| | raise NotImplementedError()
|
| |
|
| |
|
| |
|
| |
|
| | def betas_for_alpha_bar(
|
| | num_diffusion_timesteps,
|
| | max_beta=0.999,
|
| | alpha_transform_type="cosine",
|
| | ):
|
| | """
|
| | Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
|
| | (1-beta) over time from t = [0,1].
|
| |
|
| | Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
|
| | to that part of the diffusion process.
|
| |
|
| | Args:
|
| | num_diffusion_timesteps (`int`): the number of betas to produce.
|
| | max_beta (`float`): the maximum beta to use; use values lower than 1 to
|
| | prevent singularities.
|
| | alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
|
| | Choose from `cosine` or `exp`
|
| |
|
| | Returns:
|
| | betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
|
| | """
|
| | if alpha_transform_type == "cosine":
|
| | def alpha_bar_fn(t):
|
| | return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
|
| | elif alpha_transform_type == "exp":
|
| |
|
| | def alpha_bar_fn(t):
|
| | return math.exp(t * -12.0)
|
| | else:
|
| | raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
|
| |
|
| | betas = []
|
| | for i in range(num_diffusion_timesteps):
|
| | t1 = i / num_diffusion_timesteps
|
| | t2 = (i + 1) / num_diffusion_timesteps
|
| | betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
|
| | return torch.tensor(betas, dtype=torch.float32)
|
| |
|
| |
|
| |
|
| | def set_new_noise_schedule(self, schedule_opt, device, num_train_timesteps=1000):
|
| | self.ddim = schedule_opt['ddim']
|
| | self.num_train_timesteps = num_train_timesteps
|
| | to_torch = partial(torch.tensor, dtype=torch.float32, device=device)
|
| |
|
| | betas = make_beta_schedule(
|
| | schedule=schedule_opt['schedule'],
|
| | n_timestep=num_train_timesteps,
|
| | linear_start=schedule_opt['linear_start'],
|
| | linear_end=schedule_opt['linear_end'])
|
| | betas = betas.detach().cpu().numpy() if isinstance(
|
| | betas, torch.Tensor) else betas
|
| | alphas = 1. - betas
|
| | alphas_cumprod = np.cumprod(alphas, axis=0)
|
| | alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
|
| | self.sqrt_alphas_cumprod_prev = np.sqrt(
|
| | np.append(1., alphas_cumprod))
|
| |
|
| | timesteps, = betas.shape
|
| | self.num_timesteps = int(timesteps)
|
| | self.register_buffer('betas', to_torch(betas))
|
| | self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
| | self.register_buffer('alphas_cumprod_prev',
|
| | to_torch(alphas_cumprod_prev))
|
| |
|
| |
|
| | self.register_buffer('sqrt_alphas_cumprod',
|
| | to_torch(np.sqrt(alphas_cumprod)))
|
| | self.register_buffer('sqrt_one_minus_alphas_cumprod',
|
| | to_torch(np.sqrt(1. - alphas_cumprod)))
|
| | self.register_buffer('log_one_minus_alphas_cumprod',
|
| | to_torch(np.log(1. - alphas_cumprod)))
|
| | self.register_buffer('sqrt_recip_alphas_cumprod',
|
| | to_torch(np.sqrt(1. / alphas_cumprod)))
|
| | self.register_buffer('sqrt_recipm1_alphas_cumprod',
|
| | to_torch(np.sqrt(1. / alphas_cumprod - 1)))
|
| |
|
| |
|
| | posterior_variance = betas * \
|
| | (1. - alphas_cumprod_prev) / (1. - alphas_cumprod)
|
| |
|
| | self.register_buffer('posterior_variance',
|
| | to_torch(posterior_variance))
|
| |
|
| | self.register_buffer('posterior_log_variance_clipped', to_torch(
|
| | np.log(np.maximum(posterior_variance, 1e-20))))
|
| | self.register_buffer('posterior_mean_coef1', to_torch(
|
| | betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
|
| | self.register_buffer('posterior_mean_coef2', to_torch(
|
| | (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
|
| |
|
| | self.schedule_type = schedule_opt['schedule']
|
| | if self.ddim>0:
|
| | print('================ddim scheduler is adopted===================')
|
| | self.ddim_num_steps = schedule_opt['n_timestep']
|
| | print('==========ddim sampling steps: {}==========='.format(self.ddim_num_steps))
|
| |
|
| |
|
| |
|
| |
|
| | def predict_start_from_noise(self, x_t, t, noise):
|
| | return self.sqrt_recip_alphas_cumprod[t] * x_t - \
|
| | self.sqrt_recipm1_alphas_cumprod[t] * noise
|
| |
|
| | def q_posterior(self, x_start, x_t, t):
|
| | posterior_mean = self.posterior_mean_coef1[t] * \
|
| | x_start + self.posterior_mean_coef2[t] * x_t
|
| | posterior_log_variance_clipped = self.posterior_log_variance_clipped[t]
|
| | return posterior_mean, posterior_log_variance_clipped
|
| |
|
| | def p_mean_variance(self, x, t, clip_denoised: bool, condition_x=None):
|
| | batch_size = x.shape[0]
|
| | noise_level = torch.FloatTensor(
|
| | [self.sqrt_alphas_cumprod_prev[t+1]]).repeat(batch_size, 1).to(x.device)
|
| | if condition_x is not None:
|
| | x_recon = self.predict_start_from_noise(
|
| | x, t=t, noise=self.denoise_fn(torch.cat([condition_x, x], dim=1), noise_level, t))
|
| | else:
|
| | x_recon = self.predict_start_from_noise(
|
| | x, t=t, noise=self.denoise_fn(x, noise_level))
|
| |
|
| | if clip_denoised:
|
| | x_recon.clamp_(-1., 1.)
|
| |
|
| | model_mean, posterior_log_variance = self.q_posterior(
|
| | x_start=x_recon, x_t=x, t=t)
|
| | return model_mean, posterior_log_variance, x_recon
|
| |
|
| |
|
| |
|
| |
|
| |
|
| | def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
|
| | """
|
| | "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
|
| | prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
|
| | s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
|
| | pixels from saturation at each step. We find that dynamic thresholding results in significantly better
|
| | photorealism as well as better image-text alignment, especially when using very large guidance weights."
|
| |
|
| | https://arxiv.org/abs/2205.11487
|
| | """
|
| | dtype = sample.dtype
|
| | batch_size, channels, *remaining_dims = sample.shape
|
| |
|
| | if dtype not in (torch.float32, torch.float64):
|
| | sample = sample.float()
|
| |
|
| |
|
| | sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
|
| |
|
| | abs_sample = sample.abs()
|
| |
|
| | s = torch.quantile(abs_sample, 0.995, dim=1)
|
| | s = torch.clamp(s, min=1, max=1.0)
|
| | s = s.unsqueeze(1)
|
| | sample = torch.clamp(sample, -s, s) / s
|
| |
|
| | sample = sample.reshape(batch_size, channels, *remaining_dims)
|
| | sample = sample.to(dtype)
|
| |
|
| | return sample
|
| |
|
| | def ddim_sample(self, condition_x, img_or_shape, device, seed=1, img_s1=None):
|
| |
|
| |
|
| |
|
| | if self.schedule_type=='linear':
|
| | self.ddim_sampling_eta = 0.8
|
| | simple_var=False
|
| | threshold_x = False
|
| | elif self.schedule_type=='cosine':
|
| | self.ddim_sampling_eta = 0.8
|
| | simple_var=False
|
| |
|
| | threshold_x = False
|
| |
|
| |
|
| | batch, total_timesteps, sampling_timesteps, eta= \
|
| | img_or_shape[0], self.num_train_timesteps, \
|
| | self.ddim_num_steps, self.ddim_sampling_eta
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| | noisy_img_s1 = None
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| | if simple_var:
|
| | eta = 1
|
| | ts = torch.linspace(total_timesteps, 0, (sampling_timesteps + 1)).to(device).to(torch.long)
|
| |
|
| | x = torch.randn(img_or_shape).to(device)
|
| | batch_size = x.shape[0]
|
| |
|
| | imgs = [x]
|
| | img_onestep = [condition_x[:,:self.channels,...]]
|
| | if self.opt['stage']!=2:
|
| | tbar = tqdm(range(1, sampling_timesteps + 1),f'seed{seed} DDIM sampling ({self.schedule_type}) with eta {eta} simple_var {simple_var}')
|
| | else:
|
| | tbar = range(1, sampling_timesteps + 1)
|
| | for i in tbar:
|
| | cur_t = ts[i - 1] - 1
|
| | prev_t = ts[i] - 1
|
| | noise_level = torch.FloatTensor(
|
| |
|
| | [self.sqrt_alphas_cumprod_prev[cur_t]]).repeat(batch_size, 1).to(x.device)
|
| |
|
| |
|
| | alpha_prod_t = self.alphas_cumprod[cur_t]
|
| | alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else 1
|
| | beta_prod_t = 1 - alpha_prod_t
|
| |
|
| |
|
| |
|
| |
|
| | model_output = self.denoise_fn(torch.cat([condition_x, x], dim=1), noise_level)
|
| |
|
| | sigma_2 = eta * (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
|
| | noise = torch.randn_like(x)
|
| |
|
| |
|
| |
|
| |
|
| | pred_original_sample = (x - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
|
| |
|
| |
|
| | if threshold_x:
|
| | pred_original_sample = self._threshold_sample(pred_original_sample)
|
| | else:
|
| | pred_original_sample = pred_original_sample.clamp(-1, 1)
|
| |
|
| | pred_sample_direction = (1 - alpha_prod_t_prev - sigma_2) ** (0.5) * model_output
|
| |
|
| |
|
| |
|
| | if simple_var:
|
| | third_term = (1 - alpha_prod_t / alpha_prod_t_prev)**0.5 * noise
|
| | else:
|
| | third_term = sigma_2**0.5 * noise
|
| |
|
| | x = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + third_term
|
| | imgs.append(x)
|
| | img_onestep.append(pred_original_sample)
|
| |
|
| | imgs = torch.concat(imgs, dim = 0)
|
| | img_onestep = torch.concat(img_onestep, dim = 0)
|
| |
|
| |
|
| | return imgs, img_onestep
|
| |
|
| |
|
| | @torch.no_grad()
|
| | def p_sample(self, x, t, clip_denoised=True, condition_x=None):
|
| | model_mean, model_log_variance, x_recon = self.p_mean_variance(
|
| | x=x, t=t, clip_denoised=clip_denoised, condition_x=condition_x)
|
| | noise = torch.randn_like(x) if t > 0 else torch.zeros_like(x)
|
| | return model_mean + noise * (0.5 * model_log_variance).exp(), x_recon
|
| |
|
| | @torch.no_grad()
|
| | def p_sample_loop(self, x_in, continous=False, seed=1, img_s1=None):
|
| | device = self.betas.device
|
| |
|
| | sample_inter = 1
|
| | if not self.conditional:
|
| | shape = x_in
|
| | img = torch.randn(shape, device=device)
|
| | ret_img = img
|
| | if not self.ddim:
|
| | for i in tqdm(reversed(range(0, self.num_timesteps)), desc='sampling loop time step', total=self.num_timesteps):
|
| | img, x_recon = self.p_sample(img, i)
|
| | if i % sample_inter == 0:
|
| | ret_img = torch.cat([ret_img, img], dim=0)
|
| | else:
|
| | for i in tqdm(range(0, len(self.ddim_timesteps)), desc='sampling loop time step', total=len(self.ddim_timesteps)):
|
| | ddim_t = self.ddim_timesteps[i]
|
| | img = self.ddim_sample(img, ddim_t)
|
| | if i % sample_inter == 0:
|
| | ret_img = torch.cat([ret_img, img], dim=0)
|
| |
|
| | else:
|
| | x = x_in
|
| | shape = (x.shape[0], self.channels, x.shape[-2], x.shape[-1])
|
| |
|
| |
|
| | if self.xT_noise_r>0:
|
| |
|
| | print('adopting ddpm inversion as initial noise, ratio is {}'.format(self.xT_noise_r))
|
| | img0 = torch.randn(shape, device=device)
|
| | x_start = x_in[:, 0:1, ...]
|
| | continuous_sqrt_alpha_cumprod = torch.FloatTensor(
|
| | np.random.uniform(
|
| | self.sqrt_alphas_cumprod_prev[self.num_timesteps-1],
|
| | self.sqrt_alphas_cumprod_prev[self.num_timesteps],
|
| | size=x_start.shape[0]
|
| | )).to(x_start.device)
|
| | continuous_sqrt_alpha_cumprod = continuous_sqrt_alpha_cumprod.view(x_start.shape[0], -1)
|
| |
|
| | noise = default(x_start, lambda: torch.randn_like(x_start))
|
| | img = self.q_sample(
|
| | x_start=x_start, continuous_sqrt_alpha_cumprod=continuous_sqrt_alpha_cumprod.view(-1, 1, 1, 1), noise=noise)
|
| |
|
| |
|
| | img = self.xT_noise_r*img + (1-self.xT_noise_r)*img0
|
| |
|
| | else:
|
| | img = torch.randn(shape, device=device)
|
| |
|
| | ret_img = x
|
| | img_onestep = x
|
| |
|
| | if self.opt['stage']!=2:
|
| | if not self.ddim:
|
| | for i in tqdm(reversed(range(0, self.num_timesteps)), desc='ddpm sampling loop time step', total=self.num_timesteps):
|
| | img, x_recon = self.p_sample(img, i, condition_x=x)
|
| | if i % sample_inter == 0:
|
| | ret_img = torch.cat([ret_img[:,:self.channels,...], img], dim=0)
|
| | if i % sample_inter==0 or i==self.num_timesteps-1:
|
| | img_onestep = torch.cat([img_onestep[:,:self.channels,...], x_recon], dim=0)
|
| |
|
| | else:
|
| | ret_img, img_onestep = self.ddim_sample(condition_x=x, img_or_shape=shape, device=device, seed=seed, img_s1=img_s1)
|
| |
|
| |
|
| | if continous:
|
| | return ret_img, img_onestep
|
| | else:
|
| | return ret_img[-x_in.shape[0]:], img_onestep
|
| | else:
|
| |
|
| | self.ddim_num_steps = self.opt['ddim_steps']
|
| | ret_img, img_onestep = self.ddim_sample(condition_x=x, img_or_shape=shape, device=device, seed=seed, img_s1=img_s1)
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| | if continous:
|
| | return ret_img, img_onestep
|
| | else:
|
| | return ret_img[-x_in.shape[0]:], img_onestep
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| | @torch.no_grad()
|
| | def sample(self, batch_size=1, continous=False):
|
| | image_size = self.image_size
|
| | channels = self.channels
|
| | return self.p_sample_loop((batch_size, channels, image_size, image_size), continous)
|
| |
|
| |
|
| | @torch.no_grad()
|
| | def super_resolution(self, x_in, continous=False, seed=1, img_s1=None):
|
| |
|
| | return self.p_sample_loop(x_in, continous, seed=seed, img_s1=img_s1)
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
| | def q_sample(self, x_start, continuous_sqrt_alpha_cumprod, noise=None):
|
| | noise = default(noise, lambda: torch.randn_like(x_start))
|
| |
|
| |
|
| | return (
|
| | continuous_sqrt_alpha_cumprod * x_start +
|
| | (1 - continuous_sqrt_alpha_cumprod**2).sqrt() * noise
|
| | )
|
| |
|
| | def p_losses(self, x_in, noise=None):
|
| |
|
| | x_start = x_in['HR']
|
| |
|
| |
|
| |
|
| | [b, c, h, w] = x_start.shape
|
| | if self.opt['stage'] ==2:
|
| | t = 999
|
| | self.ddim_num_steps = self.opt['ddim_steps']
|
| | x = x_in['SR']
|
| | shape = (x.shape[0], self.channels, x.shape[-2], x.shape[-1])
|
| | ret_img, img_onestep = self.ddim_sample(condition_x=x, img_or_shape=shape, device=x.device, seed=self.seed, img_s1=x)
|
| | x_recon = ret_img[-x.shape[0]:]
|
| |
|
| |
|
| | else:
|
| | t = np.random.randint(1, self.num_timesteps + 1)
|
| |
|
| | continuous_sqrt_alpha_cumprod = torch.FloatTensor(
|
| | np.random.uniform(
|
| | self.sqrt_alphas_cumprod_prev[t-1],
|
| | self.sqrt_alphas_cumprod_prev[t],
|
| | size=b
|
| | )).to(x_start.device)
|
| | continuous_sqrt_alpha_cumprod = continuous_sqrt_alpha_cumprod.view(b, -1)
|
| |
|
| |
|
| | noise = default(noise, lambda: torch.randn_like(x_start))
|
| | x_noisy = self.q_sample(
|
| | x_start=x_start, continuous_sqrt_alpha_cumprod=continuous_sqrt_alpha_cumprod.view(-1, 1, 1, 1), noise=noise)
|
| |
|
| |
|
| |
|
| | if not self.conditional:
|
| | x_recon = self.denoise_fn(x_noisy, continuous_sqrt_alpha_cumprod)
|
| | else:
|
| |
|
| | x_recon, condition_feats = self.denoise_fn(
|
| | torch.cat([x_in['SR'], x_noisy], dim=1),
|
| | continuous_sqrt_alpha_cumprod,
|
| |
|
| |
|
| | return_condition=True
|
| | )
|
| | if self.opt['stage']==2:
|
| | l_pix = self.loss_func(x_start, x_recon)
|
| |
|
| | else:
|
| | l_pix = self.loss_func(noise, x_recon)
|
| |
|
| |
|
| | x_pred = x_recon
|
| | condition_feats=None
|
| |
|
| |
|
| | return l_pix, x_start, x_pred, condition_feats, torch.tensor(t, device=l_pix.device)
|
| |
|
| |
|
| | def forward(self, x, *args, **kwargs):
|
| | return self.p_losses(x, *args, **kwargs)
|
| |
|