|
|
import math |
|
|
import os |
|
|
|
|
|
|
|
|
import matplotlib.pyplot as plt |
|
|
import numpy as np |
|
|
import torch |
|
|
from tqdm.auto import tqdm |
|
|
|
|
|
import random |
|
|
|
|
|
import pyiqa |
|
|
|
|
|
from util.img_utils import clear_color |
|
|
from .posterior_mean_variance import get_mean_processor, get_var_processor |
|
|
|
|
|
|
|
|
def set_seed(seed): |
|
|
torch.manual_seed(seed) |
|
|
np.random.seed(seed) |
|
|
random.seed(seed) |
|
|
torch.cuda.manual_seed_all(seed) |
|
|
|
|
|
|
|
|
|
|
|
__SAMPLER__ = {} |
|
|
|
|
|
def register_sampler(name: str): |
|
|
def wrapper(cls): |
|
|
if __SAMPLER__.get(name, None): |
|
|
raise NameError(f"Name {name} is already registered!") |
|
|
__SAMPLER__[name] = cls |
|
|
return cls |
|
|
return wrapper |
|
|
|
|
|
|
|
|
def get_sampler(name: str): |
|
|
if __SAMPLER__.get(name, None) is None: |
|
|
raise NameError(f"Name {name} is not defined!") |
|
|
return __SAMPLER__[name] |
|
|
|
|
|
|
|
|
def create_sampler(sampler, |
|
|
steps, |
|
|
noise_schedule, |
|
|
model_mean_type, |
|
|
model_var_type, |
|
|
dynamic_threshold, |
|
|
clip_denoised, |
|
|
rescale_timesteps, |
|
|
timestep_respacing=""): |
|
|
|
|
|
sampler = get_sampler(name=sampler) |
|
|
|
|
|
betas = get_named_beta_schedule(noise_schedule, steps) |
|
|
if not timestep_respacing: |
|
|
timestep_respacing = [steps] |
|
|
|
|
|
return sampler(use_timesteps=space_timesteps(steps, timestep_respacing), |
|
|
betas=betas, |
|
|
model_mean_type=model_mean_type, |
|
|
model_var_type=model_var_type, |
|
|
dynamic_threshold=dynamic_threshold, |
|
|
clip_denoised=clip_denoised, |
|
|
rescale_timesteps=rescale_timesteps) |
|
|
|
|
|
def compute_psnr(img1, img2): |
|
|
""" |
|
|
Computes the Peak Signal-to-Noise Ratio (PSNR) between two images. |
|
|
The images should have pixel values in the range [-1, 1]. |
|
|
|
|
|
Args: |
|
|
img1 (torch.Tensor): The first image tensor (e.g., reference image). |
|
|
Shape: (N, C, H, W) or (C, H, W). |
|
|
img2 (torch.Tensor): The second image tensor (e.g., generated image). |
|
|
Shape: same as img1. |
|
|
|
|
|
Returns: |
|
|
psnr (float): The computed PSNR value in decibels (dB). |
|
|
""" |
|
|
|
|
|
assert img1.shape == img2.shape, "Input images must have the same shape" |
|
|
|
|
|
|
|
|
mse = torch.mean((img1 - img2) ** 2) |
|
|
|
|
|
|
|
|
if mse == 0: |
|
|
return float('inf') |
|
|
|
|
|
|
|
|
max_pixel_value = 2.0 |
|
|
|
|
|
|
|
|
psnr = 20 * torch.log10(max_pixel_value / torch.sqrt(mse)) |
|
|
|
|
|
return psnr.item() |
|
|
|
|
|
class GaussianDiffusion: |
|
|
def __init__(self, |
|
|
betas, |
|
|
model_mean_type, |
|
|
model_var_type, |
|
|
dynamic_threshold, |
|
|
clip_denoised, |
|
|
rescale_timesteps |
|
|
): |
|
|
|
|
|
|
|
|
betas = np.array(betas, dtype=np.float64) |
|
|
self.betas = betas |
|
|
assert self.betas.ndim == 1, "betas must be 1-D" |
|
|
assert (0 < self.betas).all() and (self.betas <=1).all(), "betas must be in (0..1]" |
|
|
|
|
|
self.num_timesteps = int(self.betas.shape[0]) |
|
|
self.rescale_timesteps = rescale_timesteps |
|
|
|
|
|
alphas = 1.0 - self.betas |
|
|
self.alphas = alphas |
|
|
self.alphas_cumprod = np.cumprod(alphas, axis=0) |
|
|
self.alphas_cumprod_prev = np.append(1.0, self.alphas_cumprod[:-1]) |
|
|
self.alphas_cumprod_next = np.append(self.alphas_cumprod[1:], 0.0) |
|
|
assert self.alphas_cumprod_prev.shape == (self.num_timesteps,) |
|
|
|
|
|
|
|
|
self.sqrt_alphas_cumprod = np.sqrt(self.alphas_cumprod) |
|
|
self.sqrt_one_minus_alphas_cumprod = np.sqrt(1.0 - self.alphas_cumprod) |
|
|
self.log_one_minus_alphas_cumprod = np.log(1.0 - self.alphas_cumprod) |
|
|
self.sqrt_recip_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod) |
|
|
self.sqrt_recipm1_alphas_cumprod = np.sqrt(1.0 / self.alphas_cumprod - 1) |
|
|
|
|
|
|
|
|
self.posterior_variance = ( |
|
|
betas * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod) |
|
|
) |
|
|
|
|
|
|
|
|
self.posterior_log_variance_clipped = np.log( |
|
|
np.append(self.posterior_variance[1], self.posterior_variance[1:]) |
|
|
) |
|
|
self.posterior_mean_coef1 = ( |
|
|
betas * np.sqrt(self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod) |
|
|
) |
|
|
self.posterior_mean_coef2 = ( |
|
|
(1.0 - self.alphas_cumprod_prev) |
|
|
* np.sqrt(alphas) |
|
|
/ (1.0 - self.alphas_cumprod) |
|
|
) |
|
|
|
|
|
self.mean_processor = get_mean_processor(model_mean_type, |
|
|
betas=betas, |
|
|
dynamic_threshold=dynamic_threshold, |
|
|
clip_denoised=clip_denoised) |
|
|
|
|
|
self.var_processor = get_var_processor(model_var_type, |
|
|
betas=betas) |
|
|
|
|
|
def q_mean_variance(self, x_start, t): |
|
|
""" |
|
|
Get the distribution q(x_t | x_0). |
|
|
|
|
|
:param x_start: the [N x C x ...] tensor of noiseless inputs. |
|
|
:param t: the number of diffusion steps (minus 1). Here, 0 means one step. |
|
|
:return: A tuple (mean, variance, log_variance), all of x_start's shape. |
|
|
""" |
|
|
|
|
|
mean = extract_and_expand(self.sqrt_alphas_cumprod, t, x_start) * x_start |
|
|
variance = extract_and_expand(1.0 - self.alphas_cumprod, t, x_start) |
|
|
log_variance = extract_and_expand(self.log_one_minus_alphas_cumprod, t, x_start) |
|
|
|
|
|
return mean, variance, log_variance |
|
|
|
|
|
def q_sample(self, x_start, t): |
|
|
""" |
|
|
Diffuse the data for a given number of diffusion steps. |
|
|
|
|
|
In other words, sample from q(x_t | x_0). |
|
|
|
|
|
:param x_start: the initial data batch. |
|
|
:param t: the number of diffusion steps (minus 1). Here, 0 means one step. |
|
|
:param noise: if specified, the split-out normal noise. |
|
|
:return: A noisy version of x_start. |
|
|
""" |
|
|
noise = torch.randn_like(x_start) |
|
|
assert noise.shape == x_start.shape |
|
|
|
|
|
coef1 = extract_and_expand(self.sqrt_alphas_cumprod, t, x_start) |
|
|
coef2 = extract_and_expand(self.sqrt_one_minus_alphas_cumprod, t, x_start) |
|
|
|
|
|
return coef1 * x_start + coef2 * noise |
|
|
|
|
|
def q_posterior_mean_variance(self, x_start, x_t, t): |
|
|
""" |
|
|
Compute the mean and variance of the diffusion posterior: |
|
|
|
|
|
q(x_{t-1} | x_t, x_0) |
|
|
|
|
|
""" |
|
|
assert x_start.shape == x_t.shape |
|
|
coef1 = extract_and_expand(self.posterior_mean_coef1, t, x_start) |
|
|
coef2 = extract_and_expand(self.posterior_mean_coef2, t, x_t) |
|
|
posterior_mean = coef1 * x_start + coef2 * x_t |
|
|
posterior_variance = extract_and_expand(self.posterior_variance, t, x_t) |
|
|
posterior_log_variance_clipped = extract_and_expand(self.posterior_log_variance_clipped, t, x_t) |
|
|
|
|
|
assert ( |
|
|
posterior_mean.shape[0] |
|
|
== posterior_variance.shape[0] |
|
|
== posterior_log_variance_clipped.shape[0] |
|
|
== x_start.shape[0] |
|
|
) |
|
|
return posterior_mean, posterior_variance, posterior_log_variance_clipped |
|
|
|
|
|
torch.no_grad() |
|
|
def p_sample_loop_compression(self, |
|
|
model, |
|
|
x_start, |
|
|
ref_img, |
|
|
record, |
|
|
save_root, |
|
|
num_opt_noises, |
|
|
num_random_noises, |
|
|
loss_type, |
|
|
decode_residual_gap, |
|
|
fname, |
|
|
eta, |
|
|
num_best_opt_noises, |
|
|
num_pursuit_noises, |
|
|
num_pursuit_coef_bits, |
|
|
random_opt_mse_noises): |
|
|
""" |
|
|
The function used for sampling from noise. |
|
|
""" |
|
|
assert num_best_opt_noises + num_random_noises > 0 |
|
|
|
|
|
|
|
|
|
|
|
set_seed(100000) |
|
|
device = x_start.device |
|
|
img = torch.randn(1 + random_opt_mse_noises, *x_start.shape[1:], device=device) |
|
|
|
|
|
plt.imsave(os.path.join(save_root, f"progress/img_to_compress.png"), clear_color(ref_img)) |
|
|
best_indices_list = [] |
|
|
x_hat_0_list = [] |
|
|
|
|
|
pbar = tqdm(list(range(self.num_timesteps))[::-1]) |
|
|
num_noises_total = 0 |
|
|
num_steps_total = 0 |
|
|
for idx in pbar: |
|
|
set_seed(idx) |
|
|
time = torch.tensor([idx] * img.shape[0], device=device) |
|
|
if len(x_hat_0_list) >= 2: |
|
|
x_hat_0_list = x_hat_0_list[-decode_residual_gap:] |
|
|
x_hat_0_list_tensor = torch.stack(x_hat_0_list, dim=0) |
|
|
|
|
|
|
|
|
probs = torch.linspace(0, 1, len(x_hat_0_list) - 1, device=device) |
|
|
probs /= torch.sum(probs) |
|
|
|
|
|
residual = torch.sum(probs.view(-1, 1) * (x_hat_0_list_tensor[1:] - x_hat_0_list_tensor[:-1]).view(len(x_hat_0_list) - 1, -1), dim=0) |
|
|
|
|
|
new_noise = torch.randn(num_opt_noises, *img.shape[1:], device=device) |
|
|
similarity = torch.matmul(new_noise.view(num_opt_noises, -1), |
|
|
residual.view(-1, 1)).squeeze(1) |
|
|
sorted_similarity, sorted_indices = torch.sort(similarity, descending=False) |
|
|
|
|
|
noise = new_noise[sorted_indices][:num_best_opt_noises] |
|
|
if num_random_noises > 0: |
|
|
noise = torch.cat((noise, torch.randn(num_random_noises, *img.shape[1:], device=device)), dim=0) |
|
|
|
|
|
else: |
|
|
noise = torch.randn(num_best_opt_noises + num_random_noises, *img.shape[1:], device=device) |
|
|
num_noises_total += noise.shape[0] |
|
|
num_steps_total += 1 |
|
|
|
|
|
out = self.p_sample(x=img, |
|
|
t=time, |
|
|
model=model, |
|
|
noise=noise, |
|
|
ref=ref_img, |
|
|
loss_type=loss_type, |
|
|
random_opt_mse_noises=random_opt_mse_noises, |
|
|
eta=eta, |
|
|
num_pursuit_noises=num_pursuit_noises, |
|
|
num_pursuit_coef_bits=num_pursuit_coef_bits) |
|
|
best_idx = out['best_idx'] |
|
|
best_indices_list.append(best_idx.cpu().numpy()) |
|
|
|
|
|
|
|
|
img = out['sample'] |
|
|
x_0_hat = out['pred_xstart'] |
|
|
x_hat_0_list.append(x_0_hat[0].unsqueeze(0)) |
|
|
|
|
|
|
|
|
|
|
|
if record: |
|
|
if idx % 50 == 0: |
|
|
plt.imsave(os.path.join(save_root, f"progress/x_0_hat_{str(idx).zfill(4)}.png"), clear_color(x_0_hat[0].unsqueeze(0).clip(-1, 1))) |
|
|
plt.imsave(os.path.join(save_root, f"progress/x_t_{str(idx).zfill(4)}.png"), clear_color(img[0].unsqueeze(0).clip(-1, 1))) |
|
|
plt.imsave(os.path.join(save_root, f"progress/noise_t_{str(idx).zfill(4)}.png"), clear_color(noise[0].unsqueeze(0).clip(-1, 1))) |
|
|
plt.imsave(os.path.join(save_root, f"progress/err_t_{str(idx).zfill(4)}.png"), clear_color((ref_img - x_0_hat)[0].unsqueeze(0))) |
|
|
del noise |
|
|
|
|
|
|
|
|
|
|
|
plt.imsave(os.path.join(save_root, |
|
|
f"progress/x_0_hat_final_psnr={compute_psnr(img[0].unsqueeze(0), ref_img)}_bpp={np.log2(num_noises_total / num_steps_total)}.png"), |
|
|
clear_color(img[0].unsqueeze(0))) |
|
|
indices_save_folder = os.path.join(save_root, 'best_indices') |
|
|
os.makedirs(indices_save_folder, exist_ok=True) |
|
|
np.save(os.path.join(indices_save_folder, os.path.splitext(os.path.basename(fname))[0] + '.bestindices'), np.array(best_indices_list)) |
|
|
|
|
|
return img |
|
|
|
|
|
@torch.no_grad() |
|
|
def p_sample_loop_blind_restoration(self, |
|
|
model, |
|
|
x_start, |
|
|
mmse_img, |
|
|
num_opt_noises, |
|
|
iqa_metric, |
|
|
iqa_coef, |
|
|
eta, |
|
|
loaded_indices): |
|
|
|
|
|
assert iqa_metric == 'niqe' or iqa_metric == 'clipiqa+' or iqa_metric == 'topiq_nr-face' |
|
|
iqa = pyiqa.create_metric(iqa_metric, device=x_start.device) |
|
|
device = x_start.device |
|
|
|
|
|
set_seed(100000) |
|
|
img = torch.randn(2, *x_start.shape[1:], device=device) |
|
|
|
|
|
pbar = tqdm(list(range(self.num_timesteps))[::-1]) |
|
|
next_idx = np.array([0, 1]) |
|
|
if loaded_indices is not None: |
|
|
indices = loaded_indices |
|
|
loaded_indices = torch.cat((loaded_indices, torch.tensor([0], device=device, dtype=loaded_indices.dtype)), dim=0) |
|
|
else: |
|
|
indices = [] |
|
|
for i, idx in enumerate(pbar): |
|
|
set_seed(idx) |
|
|
|
|
|
|
|
|
noise = torch.randn(num_opt_noises, *img.shape[1:], device=device) |
|
|
if loaded_indices is None: |
|
|
time = torch.tensor([idx] * img.shape[0], device=device) |
|
|
out = self.p_sample(x=img, |
|
|
t=time, |
|
|
model=model, |
|
|
noise=noise, |
|
|
ref=mmse_img, |
|
|
loss_type='dot_prod', |
|
|
optimize_iqa=True, |
|
|
eta=eta, |
|
|
iqa=iqa, |
|
|
iqa_coef=iqa_coef) |
|
|
img = out['sample'] |
|
|
best_perceptual_idx_cur = out['best_perceptual_idx'] |
|
|
indices.append(next_idx[best_perceptual_idx_cur]) |
|
|
next_idx = out['best_idx'] |
|
|
else: |
|
|
time = torch.tensor([idx], device=device) |
|
|
if i == 0: |
|
|
img = img[loaded_indices[0]].unsqueeze(0) |
|
|
out = self.p_sample(x=img, |
|
|
t=time, |
|
|
model=model, |
|
|
noise=noise[loaded_indices[i+1]].unsqueeze(0), |
|
|
ref=img, |
|
|
loss_type='dot_prod', |
|
|
optimize_iqa=False, |
|
|
eta=eta, |
|
|
iqa='niqe', |
|
|
iqa_coef=0.0) |
|
|
img = out['sample'] |
|
|
|
|
|
|
|
|
if type(indices) is list: |
|
|
indices = torch.tensor(indices).flatten() |
|
|
return img[0].unsqueeze(0), indices |
|
|
|
|
|
|
|
|
@torch.no_grad() |
|
|
def p_sample_loop_linear_restoration(self, |
|
|
model, |
|
|
x_start, |
|
|
ref_img, |
|
|
linear_operator, |
|
|
y_n, |
|
|
num_pursuit_noises, |
|
|
num_pursuit_coef_bits, |
|
|
record, |
|
|
save_root, |
|
|
num_opt_noises, |
|
|
fname, |
|
|
eta): |
|
|
""" |
|
|
The function used for sampling from noise. |
|
|
""" |
|
|
|
|
|
set_seed(100000) |
|
|
device = x_start.device |
|
|
img = torch.randn(1, *x_start.shape[1:], device=device) |
|
|
|
|
|
|
|
|
pbar = tqdm(list(range(self.num_timesteps))[::-1]) |
|
|
for idx in pbar: |
|
|
set_seed(idx) |
|
|
time = torch.tensor([idx] * img.shape[0], device=device) |
|
|
|
|
|
noise = torch.randn(num_opt_noises, *img.shape[1:], device=device) |
|
|
|
|
|
out = self.p_sample(x=img, |
|
|
t=time, |
|
|
model=model, |
|
|
noise=noise, |
|
|
ref=ref_img, |
|
|
loss_type='mse', |
|
|
eta=eta, |
|
|
y_n=y_n, |
|
|
linear_operator=linear_operator, |
|
|
num_pursuit_noises=num_pursuit_noises, |
|
|
num_pursuit_coef_bits=num_pursuit_coef_bits, |
|
|
optimize_iqa=False, |
|
|
iqa=None, |
|
|
iqa_coef=None) |
|
|
x_0_hat = out['pred_xstart'] |
|
|
img = out['sample'] |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if record: |
|
|
if idx % 50 == 0: |
|
|
plt.imsave(os.path.join(save_root, f"progress/x_0_hat_{str(idx).zfill(4)}.png"), clear_color(x_0_hat[0].unsqueeze(0).clip(-1, 1))) |
|
|
plt.imsave(os.path.join(save_root, f"progress/x_t_{str(idx).zfill(4)}.png"), clear_color(img[0].unsqueeze(0).clip(-1, 1))) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
return img |
|
|
def p_sample(self, model, x, t, noise, ref, loss_type, eta=None): |
|
|
raise NotImplementedError |
|
|
|
|
|
def p_mean_variance(self, model, x, t): |
|
|
model_output = model(x, self._scale_timesteps(t)) |
|
|
|
|
|
|
|
|
if model_output.shape[1] == 2 * x.shape[1]: |
|
|
model_output, model_var_values = torch.split(model_output, x.shape[1], dim=1) |
|
|
else: |
|
|
|
|
|
|
|
|
|
|
|
model_var_values = model_output |
|
|
|
|
|
model_mean, pred_xstart = self.mean_processor.get_mean_and_xstart(x, t, model_output) |
|
|
model_variance, model_log_variance = self.var_processor.get_variance(model_var_values, t) |
|
|
|
|
|
assert model_mean.shape == model_log_variance.shape == pred_xstart.shape == x.shape |
|
|
|
|
|
return {'mean': model_mean, |
|
|
'variance': model_variance, |
|
|
'log_variance': model_log_variance, |
|
|
'pred_xstart': pred_xstart} |
|
|
|
|
|
|
|
|
def _scale_timesteps(self, t): |
|
|
if self.rescale_timesteps: |
|
|
return t.float() * (1000.0 / self.num_timesteps) |
|
|
return t |
|
|
|
|
|
def space_timesteps(num_timesteps, section_counts): |
|
|
""" |
|
|
Create a list of timesteps to use from an original diffusion process, |
|
|
given the number of timesteps we want to take from equally-sized portions |
|
|
of the original process. |
|
|
For example, if there's 300 timesteps and the section counts are [10,15,20] |
|
|
then the first 100 timesteps are strided to be 10 timesteps, the second 100 |
|
|
are strided to be 15 timesteps, and the final 100 are strided to be 20. |
|
|
If the stride is a string starting with "ddim", then the fixed striding |
|
|
from the DDIM paper is used, and only one section is allowed. |
|
|
:param num_timesteps: the number of diffusion steps in the original |
|
|
process to divide up. |
|
|
:param section_counts: either a list of numbers, or a string containing |
|
|
comma-separated numbers, indicating the step count |
|
|
per section. As a special case, use "ddimN" where N |
|
|
is a number of steps to use the striding from the |
|
|
DDIM paper. |
|
|
:return: a set of diffusion steps from the original process to use. |
|
|
""" |
|
|
if isinstance(section_counts, str): |
|
|
if section_counts.startswith("ddim"): |
|
|
desired_count = int(section_counts[len("ddim") :]) |
|
|
for i in range(1, num_timesteps): |
|
|
if len(range(0, num_timesteps, i)) == desired_count: |
|
|
return set(range(0, num_timesteps, i)) |
|
|
raise ValueError( |
|
|
f"cannot create exactly {num_timesteps} steps with an integer stride" |
|
|
) |
|
|
section_counts = [int(x) for x in section_counts.split(",")] |
|
|
elif isinstance(section_counts, int): |
|
|
section_counts = [section_counts] |
|
|
|
|
|
size_per = num_timesteps // len(section_counts) |
|
|
extra = num_timesteps % len(section_counts) |
|
|
start_idx = 0 |
|
|
all_steps = [] |
|
|
for i, section_count in enumerate(section_counts): |
|
|
size = size_per + (1 if i < extra else 0) |
|
|
if size < section_count: |
|
|
raise ValueError( |
|
|
f"cannot divide section of {size} steps into {section_count}" |
|
|
) |
|
|
if section_count <= 1: |
|
|
frac_stride = 1 |
|
|
else: |
|
|
frac_stride = (size - 1) / (section_count - 1) |
|
|
cur_idx = 0.0 |
|
|
taken_steps = [] |
|
|
for _ in range(section_count): |
|
|
taken_steps.append(start_idx + round(cur_idx)) |
|
|
cur_idx += frac_stride |
|
|
all_steps += taken_steps |
|
|
start_idx += size |
|
|
return set(all_steps) |
|
|
|
|
|
|
|
|
class SpacedDiffusion(GaussianDiffusion): |
|
|
""" |
|
|
A diffusion process which can skip steps in a base diffusion process. |
|
|
:param use_timesteps: a collection (sequence or set) of timesteps from the |
|
|
original diffusion process to retain. |
|
|
:param kwargs: the kwargs to create the base diffusion process. |
|
|
""" |
|
|
|
|
|
def __init__(self, use_timesteps, **kwargs): |
|
|
self.use_timesteps = set(use_timesteps) |
|
|
self.timestep_map = [] |
|
|
self.original_num_steps = len(kwargs["betas"]) |
|
|
|
|
|
base_diffusion = GaussianDiffusion(**kwargs) |
|
|
last_alpha_cumprod = 1.0 |
|
|
new_betas = [] |
|
|
for i, alpha_cumprod in enumerate(base_diffusion.alphas_cumprod): |
|
|
if i in self.use_timesteps: |
|
|
new_betas.append(1 - alpha_cumprod / last_alpha_cumprod) |
|
|
last_alpha_cumprod = alpha_cumprod |
|
|
self.timestep_map.append(i) |
|
|
kwargs["betas"] = np.array(new_betas) |
|
|
super().__init__(**kwargs) |
|
|
|
|
|
def p_mean_variance( |
|
|
self, model, *args, **kwargs |
|
|
): |
|
|
return super().p_mean_variance(self._wrap_model(model), *args, **kwargs) |
|
|
|
|
|
def training_losses( |
|
|
self, model, *args, **kwargs |
|
|
): |
|
|
return super().training_losses(self._wrap_model(model), *args, **kwargs) |
|
|
|
|
|
def condition_mean(self, cond_fn, *args, **kwargs): |
|
|
return super().condition_mean(self._wrap_model(cond_fn), *args, **kwargs) |
|
|
|
|
|
def condition_score(self, cond_fn, *args, **kwargs): |
|
|
return super().condition_score(self._wrap_model(cond_fn), *args, **kwargs) |
|
|
|
|
|
def _wrap_model(self, model): |
|
|
if isinstance(model, _WrappedModel): |
|
|
return model |
|
|
return _WrappedModel( |
|
|
model, self.timestep_map, self.rescale_timesteps, self.original_num_steps |
|
|
) |
|
|
|
|
|
def _scale_timesteps(self, t): |
|
|
|
|
|
return t |
|
|
|
|
|
|
|
|
class _WrappedModel: |
|
|
def __init__(self, model, timestep_map, rescale_timesteps, original_num_steps): |
|
|
self.model = model |
|
|
self.timestep_map = timestep_map |
|
|
self.rescale_timesteps = rescale_timesteps |
|
|
self.original_num_steps = original_num_steps |
|
|
|
|
|
def __call__(self, x, ts, **kwargs): |
|
|
map_tensor = torch.tensor(self.timestep_map, device=ts.device, dtype=ts.dtype) |
|
|
new_ts = map_tensor[ts] |
|
|
if self.rescale_timesteps: |
|
|
new_ts = new_ts.float() * (1000.0 / self.original_num_steps) |
|
|
return self.model(x, new_ts, **kwargs) |
|
|
|
|
|
|
|
|
@register_sampler(name='ddpm') |
|
|
class DDPM(SpacedDiffusion): |
|
|
def __init__(self, *args, **kwargs): |
|
|
super().__init__(*args, **kwargs) |
|
|
|
|
|
def p_sample(self, model, x, t, noise, ref, perceptual_loss_weight, loss_type='mse', eta=None): |
|
|
out = self.p_mean_variance(model, x, t) |
|
|
pred_xstart = out['pred_xstart'] |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if loss_type == 'dot_prod': |
|
|
loss = torch.matmul(noise.view(noise.shape[0], -1), (ref - pred_xstart).view(pred_xstart.shape[0], -1).transpose(0, 1)) |
|
|
elif loss_type == 'mse': |
|
|
|
|
|
sqrt_recip_alphas_cumprod = extract_and_expand(self.sqrt_recip_alphas_cumprod, t-1 if t[0] > 0 else torch.zeros_like(t), noise) |
|
|
loss = - ((pred_xstart + sqrt_recip_alphas_cumprod * torch.exp(0.5 * out['log_variance']) * noise - ref).view(noise.shape[0], -1) ** 2).mean(1) |
|
|
elif loss_type == 'l1': |
|
|
sqrt_recip_alphas_cumprod = extract_and_expand(self.sqrt_recip_alphas_cumprod, t-1 if t[0] > 0 else torch.zeros_like(t), noise) |
|
|
loss = - torch.abs(pred_xstart + sqrt_recip_alphas_cumprod * torch.exp(0.5 * out['log_variance']) * noise - ref).view(noise.shape[0], -1).mean(1) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
else: |
|
|
raise NotImplementedError() |
|
|
|
|
|
best_idx = torch.argmax(loss) |
|
|
samples = out['mean'] + torch.exp(0.5 * out['log_variance']) * noise[best_idx].unsqueeze(0) |
|
|
|
|
|
return {'sample': samples if t[0] > 0 else pred_xstart, |
|
|
'pred_xstart': pred_xstart, |
|
|
'mse': loss[best_idx].item(), |
|
|
'best_idx': best_idx} |
|
|
|
|
|
|
|
|
@register_sampler(name='ddim') |
|
|
class DDIM(SpacedDiffusion): |
|
|
@torch.no_grad() |
|
|
def p_sample(self, model, x, t, noise, ref, loss_type='mse', eta=0.0, iqa=None, iqa_coef=1.0, |
|
|
optimize_iqa=False, linear_operator=None, y_n=None, random_opt_mse_noises=0, |
|
|
num_pursuit_noises=1, num_pursuit_coef_bits=1, |
|
|
cond_fn=None, |
|
|
cls=None |
|
|
): |
|
|
|
|
|
out = self.p_mean_variance(model, x, t) |
|
|
pred_xstart = out['pred_xstart'] |
|
|
best_perceptual_idx = None |
|
|
if optimize_iqa: |
|
|
assert not random_opt_mse_noises |
|
|
coef_sign = 1 if iqa.lower_better else -1 |
|
|
if iqa.metric_name == 'topiq_nr-face': |
|
|
assert not iqa.lower_better |
|
|
|
|
|
scores = [] |
|
|
for elem in pred_xstart: |
|
|
try: |
|
|
scores.append(iqa((elem.unsqueeze(0) * 0.5 + 0.5).clip(0, 1)).squeeze().view(1)) |
|
|
except AssertionError: |
|
|
|
|
|
scores.append(torch.zeros(1, device=x.device)) |
|
|
scores = torch.stack(scores, dim=0).squeeze() |
|
|
loss = (((ref - pred_xstart) ** 2).view(pred_xstart.shape[0], -1).mean(1) + coef_sign * iqa_coef * scores) |
|
|
else: |
|
|
loss = (((ref - pred_xstart) ** 2).view(pred_xstart.shape[0], -1).mean(1) + coef_sign * iqa_coef * iqa((pred_xstart * 0.5 + 0.5).clip(0, 1)).squeeze()) |
|
|
best_perceptual_idx = torch.argmin(loss) |
|
|
out['pred_xstart'] = out['pred_xstart'][best_perceptual_idx].unsqueeze(0) |
|
|
pred_xstart = pred_xstart[best_perceptual_idx].unsqueeze(0) |
|
|
t = t[best_perceptual_idx] |
|
|
x = x[best_perceptual_idx].unsqueeze(0) |
|
|
elif random_opt_mse_noises > 0: |
|
|
loss = (((ref - pred_xstart) ** 2).view(pred_xstart.shape[0], -1).mean(1)) |
|
|
best_mse_idx = torch.argmin(loss) |
|
|
out['pred_xstart'] = out['pred_xstart'][best_mse_idx].unsqueeze(0) |
|
|
pred_xstart = pred_xstart[best_mse_idx].unsqueeze(0) |
|
|
t = t[best_mse_idx] |
|
|
x = x[best_mse_idx].unsqueeze(0) |
|
|
|
|
|
eps = self.predict_eps_from_x_start(x, t, out['pred_xstart']) |
|
|
alpha_bar = extract_and_expand(self.alphas_cumprod, t, x) |
|
|
alpha_bar_prev = extract_and_expand(self.alphas_cumprod_prev, t, x) |
|
|
sigma = ( |
|
|
eta |
|
|
* torch.sqrt((1 - alpha_bar_prev) / (1 - alpha_bar)) |
|
|
* torch.sqrt(1 - alpha_bar / alpha_bar_prev) |
|
|
) |
|
|
mean_pred = ( |
|
|
out["pred_xstart"] * torch.sqrt(alpha_bar_prev) |
|
|
+ torch.sqrt(1 - alpha_bar_prev - sigma ** 2) * eps |
|
|
) |
|
|
sample = mean_pred |
|
|
|
|
|
if y_n is not None: |
|
|
assert linear_operator is not None |
|
|
y_n = ref if y_n is None else y_n |
|
|
|
|
|
if not optimize_iqa and random_opt_mse_noises <= 0 and cond_fn is None: |
|
|
if loss_type == 'dot_prod': |
|
|
if linear_operator is None: |
|
|
compute_loss = lambda noise_cur: torch.matmul(noise_cur.view(noise_cur.shape[0], -1), (ref - pred_xstart).view(pred_xstart.shape[0], -1).transpose(0, 1)) |
|
|
else: |
|
|
compute_loss = lambda noise_cur: torch.matmul(linear_operator.forward(noise_cur).reshape(noise_cur.shape[0], -1), (y_n - linear_operator.forward(pred_xstart)).reshape(pred_xstart.shape[0], -1).transpose(0, 1)) |
|
|
elif loss_type == 'mse': |
|
|
if linear_operator is None: |
|
|
compute_loss = lambda noise_cur: - (((sigma / torch.sqrt(alpha_bar_prev)) * noise_cur + pred_xstart - y_n) ** 2).mean((1, 2, 3)) |
|
|
else: |
|
|
compute_loss = lambda noise_cur: - (((sigma / torch.sqrt(alpha_bar_prev))[:, :, :y_n.shape[2], :y_n.shape[3]] * linear_operator.forward(noise_cur) + linear_operator.forward(pred_xstart) - y_n) ** 2).mean((1, 2, 3)) |
|
|
else: |
|
|
raise NotImplementedError() |
|
|
|
|
|
loss = compute_loss(noise) |
|
|
best_idx = torch.argmax(loss) |
|
|
best_noise = noise[best_idx] |
|
|
best_loss = loss[best_idx] |
|
|
|
|
|
if num_pursuit_noises > 1: |
|
|
pursuit_coefs = np.linspace(0, 1, 2 ** num_pursuit_coef_bits + 1)[1:] |
|
|
|
|
|
for _ in range(num_pursuit_noises - 1): |
|
|
next_best_noise = best_noise |
|
|
for pursuit_coef in pursuit_coefs: |
|
|
new_noise = best_noise.unsqueeze(0) * np.sqrt(pursuit_coef) + noise * np.sqrt(1 - pursuit_coef) |
|
|
new_noise /= new_noise.view(noise.shape[0], -1).std(1).view(noise.shape[0], 1, 1, 1) |
|
|
cur_loss = compute_loss(new_noise) |
|
|
cur_best_idx = torch.argmax(cur_loss) |
|
|
cur_best_loss = cur_loss[cur_best_idx] |
|
|
|
|
|
if cur_best_loss > best_loss: |
|
|
next_best_noise = new_noise[cur_best_idx] |
|
|
best_loss = cur_best_loss |
|
|
|
|
|
best_noise = next_best_noise |
|
|
|
|
|
if t != 0: |
|
|
sample += sigma * best_noise.unsqueeze(0) |
|
|
|
|
|
return {'sample': sample if t[0] > 0 else pred_xstart, |
|
|
'pred_xstart': pred_xstart, |
|
|
'mse': loss[best_idx].item(), |
|
|
'best_idx': best_idx} |
|
|
else: |
|
|
if random_opt_mse_noises > 0 and not optimize_iqa: |
|
|
num_rand_indices = random_opt_mse_noises |
|
|
elif optimize_iqa and random_opt_mse_noises <= 0: |
|
|
num_rand_indices = 1 |
|
|
elif cond_fn is not None: |
|
|
num_rand_indices = 2 |
|
|
else: |
|
|
raise NotImplementedError() |
|
|
loss = torch.matmul(noise.view(noise.shape[0], -1), |
|
|
(ref - pred_xstart).view(pred_xstart.shape[0], -1).transpose(0, 1)).squeeze() |
|
|
best_idx = torch.argmax(loss).reshape(1) |
|
|
rand_idx = torch.randint(0, noise.shape[0], size=(num_rand_indices, ), device=best_idx.device).reshape(num_rand_indices) |
|
|
best_and_rand_idx = torch.cat((best_idx, rand_idx), dim=0).flatten() |
|
|
if t != 0: |
|
|
sample = sample + sigma * noise[best_and_rand_idx] |
|
|
return {'sample': sample, |
|
|
'pred_xstart': pred_xstart, |
|
|
'best_idx': best_and_rand_idx, |
|
|
'best_perceptual_idx': best_perceptual_idx} |
|
|
|
|
|
def predict_eps_from_x_start(self, x_t, t, pred_xstart): |
|
|
coef1 = extract_and_expand(self.sqrt_recip_alphas_cumprod, t, x_t) |
|
|
coef2 = extract_and_expand(self.sqrt_recipm1_alphas_cumprod, t, x_t) |
|
|
return (coef1 * x_t - pred_xstart) / coef2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def get_named_beta_schedule(schedule_name, num_diffusion_timesteps): |
|
|
""" |
|
|
Get a pre-defined beta schedule for the given name. |
|
|
|
|
|
The beta schedule library consists of beta schedules which remain similar |
|
|
in the limit of num_diffusion_timesteps. |
|
|
Beta schedules may be added, but should not be removed or changed once |
|
|
they are committed to maintain backwards compatibility. |
|
|
""" |
|
|
if schedule_name == "linear": |
|
|
|
|
|
|
|
|
scale = 1000 / num_diffusion_timesteps |
|
|
beta_start = scale * 0.0001 |
|
|
beta_end = scale * 0.02 |
|
|
return np.linspace( |
|
|
beta_start, beta_end, num_diffusion_timesteps, dtype=np.float64 |
|
|
) |
|
|
elif schedule_name == "cosine": |
|
|
return betas_for_alpha_bar( |
|
|
num_diffusion_timesteps, |
|
|
lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2, |
|
|
) |
|
|
else: |
|
|
raise NotImplementedError(f"unknown beta schedule: {schedule_name}") |
|
|
|
|
|
|
|
|
def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): |
|
|
""" |
|
|
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]. |
|
|
|
|
|
:param num_diffusion_timesteps: the number of betas to produce. |
|
|
:param alpha_bar: a lambda that takes an argument t from 0 to 1 and |
|
|
produces the cumulative product of (1-beta) up to that |
|
|
part of the diffusion process. |
|
|
:param max_beta: the maximum beta to use; use values lower than 1 to |
|
|
prevent singularities. |
|
|
""" |
|
|
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(t2) / alpha_bar(t1), max_beta)) |
|
|
return np.array(betas) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def extract_and_expand(array, time, target): |
|
|
array = torch.from_numpy(array).to(target.device)[time].float() |
|
|
while array.ndim < target.ndim: |
|
|
array = array.unsqueeze(-1) |
|
|
return array.expand_as(target) |
|
|
|
|
|
|
|
|
def expand_as(array, target): |
|
|
if isinstance(array, np.ndarray): |
|
|
array = torch.from_numpy(array) |
|
|
elif isinstance(array, np.float): |
|
|
array = torch.tensor([array]) |
|
|
|
|
|
while array.ndim < target.ndim: |
|
|
array = array.unsqueeze(-1) |
|
|
|
|
|
return array.expand_as(target).to(target.device) |
|
|
|
|
|
|
|
|
def _extract_into_tensor(arr, timesteps, broadcast_shape): |
|
|
""" |
|
|
Extract values from a 1-D numpy array for a batch of indices. |
|
|
|
|
|
:param arr: the 1-D numpy array. |
|
|
:param timesteps: a tensor of indices into the array to extract. |
|
|
:param broadcast_shape: a larger shape of K dimensions with the batch |
|
|
dimension equal to the length of timesteps. |
|
|
:return: a tensor of shape [batch_size, 1, ...] where the shape has K dims. |
|
|
""" |
|
|
res = torch.from_numpy(arr).to(device=timesteps.device)[timesteps].float() |
|
|
while len(res.shape) < len(broadcast_shape): |
|
|
res = res[..., None] |
|
|
return res.expand(broadcast_shape) |
|
|
|