sgm/modules/diffusionmodules/sampling.py

"""
    Partially ported from https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py
"""


from typing import Dict, Union

import torch
from omegaconf import ListConfig, OmegaConf
from tqdm import tqdm

from ...modules.diffusionmodules.sampling_utils import (
    get_ancestral_step,
    linear_multistep_coeff,
    to_d,
    to_neg_log_sigma,
    to_sigma,
)
from ...util import append_dims, default, instantiate_from_config

DEFAULT_GUIDER = {"target": "sgm.modules.diffusionmodules.guiders.IdentityGuider"}


class BaseDiffusionSampler:
    def __init__(
        self,
        discretization_config: Union[Dict, ListConfig, OmegaConf],
        num_steps: Union[int, None] = None,
        guider_config: Union[Dict, ListConfig, OmegaConf, None] = None,
        verbose: bool = False,
        device: str = "cuda",
    ):
        self.num_steps = num_steps
        self.discretization = instantiate_from_config(discretization_config)
        self.guider = instantiate_from_config(
            default(
                guider_config,
                DEFAULT_GUIDER,
            )
        )
        self.verbose = verbose
        self.device = device

    def prepare_sampling_loop(self, x, cond, uc=None, num_steps=None):
        sigmas = self.discretization(
            self.num_steps if num_steps is None else num_steps, device=self.device
        )
        uc = default(uc, cond)

        x *= torch.sqrt(1.0 + sigmas[0] ** 2.0)
        num_sigmas = len(sigmas)

        s_in = x.new_ones([x.shape[0]])

        return x, s_in, sigmas, num_sigmas, cond, uc

    def denoise(self, x, denoiser, sigma, cond, uc):
        denoised, _, _, rgb_list = denoiser(*self.guider.prepare_inputs(x, sigma, cond, uc))
        denoised = self.guider(denoised, sigma)
        return denoised, rgb_list

    def get_sigma_gen(self, num_sigmas):
        sigma_generator = range(num_sigmas - 1)
        if self.verbose:
            print("#" * 30, " Sampling setting ", "#" * 30)
            print(f"Sampler: {self.__class__.__name__}")
            print(f"Discretization: {self.discretization.__class__.__name__}")
            print(f"Guider: {self.guider.__class__.__name__}")
            sigma_generator = tqdm(
                sigma_generator,
                total=num_sigmas,
                desc=f"Sampling with {self.__class__.__name__} for {num_sigmas} steps",
            )
        return sigma_generator


class SingleStepDiffusionSampler(BaseDiffusionSampler):
    def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc, *args, **kwargs):
        raise NotImplementedError

    def euler_step(self, x, d, dt):
        return x + dt * d


class EDMSampler(SingleStepDiffusionSampler):
    def __init__(
        self, s_churn=0.0, s_tmin=0.0, s_tmax=float("inf"), s_noise=1.0, *args, **kwargs
    ):
        super().__init__(*args, **kwargs)

        self.s_churn = s_churn
        self.s_tmin = s_tmin
        self.s_tmax = s_tmax
        self.s_noise = s_noise

    def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, gamma=0.0):
        sigma_hat = sigma * (gamma + 1.0)
        if gamma > 0:
            eps = torch.randn_like(x) * self.s_noise
            x = x + eps * append_dims(sigma_hat**2 - sigma**2, x.ndim) ** 0.5

        denoised, rgb_list = self.denoise(x, denoiser, sigma_hat, cond, uc)
        d = to_d(x, sigma_hat, denoised)
        dt = append_dims(next_sigma - sigma_hat, x.ndim)

        euler_step = self.euler_step(x, d, dt)
        x = self.possible_correction_step(
            euler_step, x, d, dt, next_sigma, denoiser, cond, uc
        )
        return x, rgb_list

    def __call__(self, denoiser, x, cond, uc=None, num_steps=None, mask=None, init_im=None):
        return self.forward(denoiser, x, cond, uc=uc, num_steps=num_steps, mask=mask, init_im=init_im)

    def forward(self, denoiser, x, cond, uc=None, num_steps=None, mask=None, init_im=None):
        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
            x, cond, uc, num_steps
        )
        for i in self.get_sigma_gen(num_sigmas):
            gamma = (
                min(self.s_churn / (num_sigmas - 1), 2**0.5 - 1)
                if self.s_tmin <= sigmas[i] <= self.s_tmax
                else 0.0
            )
            x_new, rgb_list = self.sampler_step(
                s_in * sigmas[i],
                s_in * sigmas[i + 1],
                denoiser,
                x,
                cond,
                uc,
                gamma,
            )
            x = x_new

        return x, rgb_list


def get_views(panorama_height, panorama_width, window_size=64, stride=48):
    # panorama_height /= 8
    # panorama_width /= 8
    num_blocks_height = (panorama_height - window_size) // stride + 1
    num_blocks_width = (panorama_width - window_size) // stride + 1
    total_num_blocks = int(num_blocks_height * num_blocks_width)
    views = []
    for i in range(total_num_blocks):
        h_start = int((i // num_blocks_width) * stride)
        h_end = h_start + window_size
        w_start = int((i % num_blocks_width) * stride)
        w_end = w_start + window_size
        views.append((h_start, h_end, w_start, w_end))
    return views


class EDMMultidiffusionSampler(SingleStepDiffusionSampler):
    def __init__(
        self, s_churn=0.0, s_tmin=0.0, s_tmax=float("inf"), s_noise=1.0, *args, **kwargs
    ):
        super().__init__(*args, **kwargs)

        self.s_churn = s_churn
        self.s_tmin = s_tmin
        self.s_tmax = s_tmax
        self.s_noise = s_noise

    def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, gamma=0.0):
        sigma_hat = sigma * (gamma + 1.0)
        if gamma > 0:
            eps = torch.randn_like(x) * self.s_noise
            x = x + eps * append_dims(sigma_hat**2 - sigma**2, x.ndim) ** 0.5

        denoised, rgb_list = self.denoise(x, denoiser, sigma_hat, cond, uc)
        d = to_d(x, sigma_hat, denoised)
        dt = append_dims(next_sigma - sigma_hat, x.ndim)

        euler_step = self.euler_step(x, d, dt)
        x = self.possible_correction_step(
            euler_step, x, d, dt, next_sigma, denoiser, cond, uc
        )
        return x, rgb_list

    def __call__(self, denoiser, model, x, cond, uc=None, num_steps=None, multikwargs=None):
        return self.forward(denoiser, model, x, cond, uc=uc, num_steps=num_steps, multikwargs=multikwargs)

    def forward(self, denoiser, model, x, cond, uc=None, num_steps=None, multikwargs=None):
        views = get_views(x.shape[-2], 48*(len(multikwargs)+1))
        shape = x.shape
        x = torch.randn(shape[0], shape[1], shape[2], 48*(len(multikwargs)+1)).to(x.device)
        count = torch.zeros_like(x, device=x.device)
        value = torch.zeros_like(x, device=x.device)

        x, s_in, sigmas, num_sigmas, cond_, uc = self.prepare_sampling_loop(
                x, cond[0], uc[0], num_steps
            )

        for i in self.get_sigma_gen(num_sigmas):
            gamma = (
                min(self.s_churn / (num_sigmas - 1), 2**0.5 - 1)
                if self.s_tmin <= sigmas[i] <= self.s_tmax
                else 0.0
            )
            count.zero_()
            value.zero_()

            for j, (h_start, h_end, w_start, w_end) in enumerate(views):
                # TODO we can support batches, and pass multiple views at once to the unet
                latent_view = x[:, :, h_start:h_end, w_start:w_end]
                # expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
                kwargs = {'pose': multikwargs[j]['pose'], 'mask_ref':None, 'drop_im':j}
                x_new, rgb_list = self.sampler_step(
                                    s_in * sigmas[i],
                                    s_in * sigmas[i + 1],
                                    lambda input, sigma, c: denoiser(
                                        model, input, sigma, c, **kwargs
                                    ),
                                    latent_view,
                                    cond[j],
                                    uc,
                                    gamma,
                                )
                # compute the denoising step with the reference model
                value[:, :, h_start:h_end, w_start:w_end] += x_new
                count[:, :, h_start:h_end, w_start:w_end] += 1

            # take the MultiDiffusion step
            x = torch.where(count > 0, value / count, value)

        return x, rgb_list

    def possible_correction_step(
            self, euler_step, x, d, dt, next_sigma, denoiser, cond, uc
    ):
        return euler_step


class AncestralSampler(SingleStepDiffusionSampler):
    def __init__(self, eta=1.0, s_noise=1.0, *args, **kwargs):
        super().__init__(*args, **kwargs)

        self.eta = eta
        self.s_noise = s_noise
        self.noise_sampler = lambda x: torch.randn_like(x)

    def ancestral_euler_step(self, x, denoised, sigma, sigma_down):
        d = to_d(x, sigma, denoised)
        dt = append_dims(sigma_down - sigma, x.ndim)

        return self.euler_step(x, d, dt)

    def ancestral_step(self, x, sigma, next_sigma, sigma_up):
        x = torch.where(
            append_dims(next_sigma, x.ndim) > 0.0,
            x + self.noise_sampler(x) * self.s_noise * append_dims(sigma_up, x.ndim),
            x,
        )
        return x

    def __call__(self, denoiser, x, cond, uc=None, num_steps=None):
        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
            x, cond, uc, num_steps
        )

        for i in self.get_sigma_gen(num_sigmas):
            x = self.sampler_step(
                s_in * sigmas[i],
                s_in * sigmas[i + 1],
                denoiser,
                x,
                cond,
                uc,
            )

        return x


class LinearMultistepSampler(BaseDiffusionSampler):
    def __init__(
        self,
        order=4,
        *args,
        **kwargs,
    ):
        super().__init__(*args, **kwargs)

        self.order = order

    def __call__(self, denoiser, x, cond, uc=None, num_steps=None, **kwargs):
        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
            x, cond, uc, num_steps
        )

        ds = []
        sigmas_cpu = sigmas.detach().cpu().numpy()
        for i in self.get_sigma_gen(num_sigmas):
            sigma = s_in * sigmas[i]
            denoised, _ = denoiser(
                *self.guider.prepare_inputs(x, sigma, cond, uc), **kwargs
            )
            denoised = self.guider(denoised, sigma)
            d = to_d(x, sigma, denoised)
            ds.append(d)
            if len(ds) > self.order:
                ds.pop(0)
            cur_order = min(i + 1, self.order)
            coeffs = [
                linear_multistep_coeff(cur_order, sigmas_cpu, i, j)
                for j in range(cur_order)
            ]
            x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))

        return x


class EulerEDMSampler(EDMSampler):
    def possible_correction_step(
        self, euler_step, x, d, dt, next_sigma, denoiser, cond, uc
    ):
        return euler_step


class HeunEDMSampler(EDMSampler):
    def possible_correction_step(
        self, euler_step, x, d, dt, next_sigma, denoiser, cond, uc
    ):
        if torch.sum(next_sigma) < 1e-14:
            # Save a network evaluation if all noise levels are 0
            return euler_step
        else:
            denoised = self.denoise(euler_step, denoiser, next_sigma, cond, uc)
            d_new = to_d(euler_step, next_sigma, denoised)
            d_prime = (d + d_new) / 2.0

            # apply correction if noise level is not 0
            x = torch.where(
                append_dims(next_sigma, x.ndim) > 0.0, x + d_prime * dt, euler_step
            )
            return x


class EulerAncestralSampler(AncestralSampler):
    def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc):
        sigma_down, sigma_up = get_ancestral_step(sigma, next_sigma, eta=self.eta)
        denoised = self.denoise(x, denoiser, sigma, cond, uc)
        x = self.ancestral_euler_step(x, denoised, sigma, sigma_down)
        x = self.ancestral_step(x, sigma, next_sigma, sigma_up)

        return x


class DPMPP2SAncestralSampler(AncestralSampler):
    def get_variables(self, sigma, sigma_down):
        t, t_next = [to_neg_log_sigma(s) for s in (sigma, sigma_down)]
        h = t_next - t
        s = t + 0.5 * h
        return h, s, t, t_next

    def get_mult(self, h, s, t, t_next):
        mult1 = to_sigma(s) / to_sigma(t)
        mult2 = (-0.5 * h).expm1()
        mult3 = to_sigma(t_next) / to_sigma(t)
        mult4 = (-h).expm1()

        return mult1, mult2, mult3, mult4

    def sampler_step(self, sigma, next_sigma, denoiser, x, cond, uc=None, **kwargs):
        sigma_down, sigma_up = get_ancestral_step(sigma, next_sigma, eta=self.eta)
        denoised = self.denoise(x, denoiser, sigma, cond, uc)
        x_euler = self.ancestral_euler_step(x, denoised, sigma, sigma_down)

        if torch.sum(sigma_down) < 1e-14:
            # Save a network evaluation if all noise levels are 0
            x = x_euler
        else:
            h, s, t, t_next = self.get_variables(sigma, sigma_down)
            mult = [
                append_dims(mult, x.ndim) for mult in self.get_mult(h, s, t, t_next)
            ]

            x2 = mult[0] * x - mult[1] * denoised
            denoised2 = self.denoise(x2, denoiser, to_sigma(s), cond, uc)
            x_dpmpp2s = mult[2] * x - mult[3] * denoised2

            # apply correction if noise level is not 0
            x = torch.where(append_dims(sigma_down, x.ndim) > 0.0, x_dpmpp2s, x_euler)

        x = self.ancestral_step(x, sigma, next_sigma, sigma_up)
        return x


class DPMPP2MSampler(BaseDiffusionSampler):
    def get_variables(self, sigma, next_sigma, previous_sigma=None):
        t, t_next = [to_neg_log_sigma(s) for s in (sigma, next_sigma)]
        h = t_next - t

        if previous_sigma is not None:
            h_last = t - to_neg_log_sigma(previous_sigma)
            r = h_last / h
            return h, r, t, t_next
        else:
            return h, None, t, t_next

    def get_mult(self, h, r, t, t_next, previous_sigma):
        mult1 = to_sigma(t_next) / to_sigma(t)
        mult2 = (-h).expm1()

        if previous_sigma is not None:
            mult3 = 1 + 1 / (2 * r)
            mult4 = 1 / (2 * r)
            return mult1, mult2, mult3, mult4
        else:
            return mult1, mult2

    def sampler_step(
        self,
        old_denoised,
        previous_sigma,
        sigma,
        next_sigma,
        denoiser,
        x,
        cond,
        uc=None,
    ):
        denoised = self.denoise(x, denoiser, sigma, cond, uc)

        h, r, t, t_next = self.get_variables(sigma, next_sigma, previous_sigma)
        mult = [
            append_dims(mult, x.ndim)
            for mult in self.get_mult(h, r, t, t_next, previous_sigma)
        ]

        x_standard = mult[0] * x - mult[1] * denoised
        if old_denoised is None or torch.sum(next_sigma) < 1e-14:
            # Save a network evaluation if all noise levels are 0 or on the first step
            return x_standard, denoised
        else:
            denoised_d = mult[2] * denoised - mult[3] * old_denoised
            x_advanced = mult[0] * x - mult[1] * denoised_d

            # apply correction if noise level is not 0 and not first step
            x = torch.where(
                append_dims(next_sigma, x.ndim) > 0.0, x_advanced, x_standard
            )

        return x, denoised

    def __call__(self, denoiser, x, cond, uc=None, num_steps=None, **kwargs):
        x, s_in, sigmas, num_sigmas, cond, uc = self.prepare_sampling_loop(
            x, cond, uc, num_steps
        )

        old_denoised = None
        for i in self.get_sigma_gen(num_sigmas):
            x, old_denoised = self.sampler_step(
                old_denoised,
                None if i == 0 else s_in * sigmas[i - 1],
                s_in * sigmas[i],
                s_in * sigmas[i + 1],
                denoiser,
                x,
                cond,
                uc=uc,
            )

        return x