mirror of https://github.com/vladmandic/automatic
345 lines
15 KiB
Python
345 lines
15 KiB
Python
|
|
import numpy as np
|
|
import torch
|
|
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
|
from diffusers.schedulers.scheduling_utils import SchedulerMixin, SchedulerOutput
|
|
|
|
|
|
# pylint: disable=no-member
|
|
class SpecializedRKScheduler(SchedulerMixin, ConfigMixin):
|
|
"""
|
|
SpecializedRKScheduler: High-order and specialized Runge-Kutta integrators.
|
|
Supports SSPRK, TSI_7S, Ralston 4s, and Bogacki-Shampine 4s.
|
|
Adapted from the RES4LYF repository.
|
|
"""
|
|
|
|
order = 1
|
|
|
|
@register_to_config
|
|
def __init__(
|
|
self,
|
|
num_train_timesteps: int = 1000,
|
|
beta_start: float = 0.00085,
|
|
beta_end: float = 0.012,
|
|
beta_schedule: str = "linear",
|
|
trained_betas: np.ndarray | list[float] | None = None,
|
|
prediction_type: str = "epsilon",
|
|
variant: str = "ssprk3_3s", # ssprk3_3s, ssprk4_4s, tsi_7s, ralston_4s, bogacki-shampine_4s
|
|
use_karras_sigmas: bool = False,
|
|
use_exponential_sigmas: bool = False,
|
|
use_beta_sigmas: bool = False,
|
|
use_flow_sigmas: bool = False,
|
|
sigma_min: float | None = None,
|
|
sigma_max: float | None = None,
|
|
rho: float = 7.0,
|
|
shift: float | None = None,
|
|
base_shift: float = 0.5,
|
|
max_shift: float = 1.15,
|
|
use_dynamic_shifting: bool = False,
|
|
timestep_spacing: str = "linspace",
|
|
clip_sample: bool = False,
|
|
sample_max_value: float = 1.0,
|
|
set_alpha_to_one: bool = False,
|
|
skip_prk_steps: bool = False,
|
|
interpolation_type: str = "linear",
|
|
steps_offset: int = 0,
|
|
timestep_type: str = "discrete",
|
|
rescale_betas_zero_snr: bool = False,
|
|
final_sigmas_type: str = "zero",
|
|
):
|
|
if trained_betas is not None:
|
|
self.betas = torch.tensor(trained_betas, dtype=torch.float32)
|
|
elif beta_schedule == "linear":
|
|
self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
|
|
elif beta_schedule == "scaled_linear":
|
|
self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
|
|
else:
|
|
raise NotImplementedError(f"{beta_schedule} is not implemented")
|
|
|
|
self.alphas = 1.0 - self.betas
|
|
self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
|
|
|
|
self.num_inference_steps = None
|
|
self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy())
|
|
self.sigmas = None
|
|
self.init_noise_sigma = 1.0
|
|
|
|
# Internal state
|
|
self.model_outputs = []
|
|
self.sample_at_start_of_step = None
|
|
self._step_index = None
|
|
|
|
def _get_tableau(self):
|
|
v = self.config.variant
|
|
if v == "ssprk3_3s":
|
|
a, b, c = [[1], [1 / 4, 1 / 4]], [1 / 6, 1 / 6, 2 / 3], [0, 1, 1 / 2]
|
|
elif v == "ssprk4_4s":
|
|
a, b, c = [[1 / 2], [1 / 2, 1 / 2], [1 / 6, 1 / 6, 1 / 6]], [1 / 6, 1 / 6, 1 / 6, 1 / 2], [0, 1 / 2, 1, 1 / 2]
|
|
elif v == "ralston_4s":
|
|
r5 = 5**0.5
|
|
a = [[2 / 5], [(-2889 + 1428 * r5) / 1024, (3785 - 1620 * r5) / 1024], [(-3365 + 2094 * r5) / 6040, (-975 - 3046 * r5) / 2552, (467040 + 203968 * r5) / 240845]]
|
|
b = [(263 + 24 * r5) / 1812, (125 - 1000 * r5) / 3828, (3426304 + 1661952 * r5) / 5924787, (30 - 4 * r5) / 123]
|
|
c = [0, 2 / 5, (14 - 3 * r5) / 16, 1]
|
|
elif v == "bogacki-shampine_4s":
|
|
a, b, c = [[1 / 2], [0, 3 / 4], [2 / 9, 1 / 3, 4 / 9]], [2 / 9, 1 / 3, 4 / 9, 0], [0, 1 / 2, 3 / 4, 1]
|
|
elif v == "tsi_7s":
|
|
a = [
|
|
[0.161],
|
|
[-0.008480655492356989, 0.335480655492357],
|
|
[2.8971530571054935, -6.359448489975075, 4.3622954328695815],
|
|
[5.325864828439257, -11.748883564062828, 7.4955393428898365, -0.09249506636175525],
|
|
[5.86145544294642, -12.92096931784711, 8.159367898576159, -0.071584973281401, -0.02826905039406838],
|
|
[0.09646076681806523, 0.01, 0.4798896504144996, 1.379008574103742, -3.290069515436081, 2.324710524099774],
|
|
]
|
|
b = [0.09646076681806523, 0.01, 0.4798896504144996, 1.379008574103742, -3.290069515436081, 2.324710524099774, 0.0]
|
|
c = [0.0, 0.161, 0.327, 0.9, 0.9800255409045097, 1.0, 1.0]
|
|
else:
|
|
raise ValueError(f"Unknown variant: {v}")
|
|
|
|
stages = len(c)
|
|
full_a = np.zeros((stages, stages))
|
|
for i, row in enumerate(a):
|
|
full_a[i + 1, : len(row)] = row
|
|
|
|
return full_a, np.array(b), np.array(c)
|
|
|
|
def set_timesteps(
|
|
self,
|
|
num_inference_steps: int,
|
|
device: str | torch.device = None,
|
|
mu: float | None = None, dtype: torch.dtype = torch.float32):
|
|
self.num_inference_steps = num_inference_steps
|
|
|
|
# 1. Spacing
|
|
if self.config.timestep_spacing == "linspace":
|
|
timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=float)[::-1].copy()
|
|
elif self.config.timestep_spacing == "leading":
|
|
step_ratio = self.config.num_train_timesteps // num_inference_steps
|
|
timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(float)
|
|
elif self.config.timestep_spacing == "trailing":
|
|
step_ratio = self.config.num_train_timesteps / num_inference_steps
|
|
timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(float)
|
|
timesteps -= 1
|
|
else:
|
|
raise ValueError(f"timestep_spacing must be one of 'linspace', 'leading', or 'trailing', got {self.config.timestep_spacing}")
|
|
|
|
sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
|
|
if self.config.interpolation_type == "linear":
|
|
sigmas = np.interp(timesteps, np.arange(len(sigmas)), sigmas)
|
|
elif self.config.interpolation_type == "log_linear":
|
|
sigmas = np.exp(np.interp(timesteps, np.arange(len(sigmas)), np.log(sigmas)))
|
|
else:
|
|
raise ValueError(f"interpolation_type must be one of 'linear' or 'log_linear', got {self.config.interpolation_type}")
|
|
|
|
# 2. Sigma Schedule
|
|
if self.config.use_karras_sigmas:
|
|
sigma_min = self.config.sigma_min if self.config.sigma_min is not None else sigmas[-1]
|
|
sigma_max = self.config.sigma_max if self.config.sigma_max is not None else sigmas[0]
|
|
rho = self.config.rho
|
|
ramp = np.linspace(0, 1, num_inference_steps)
|
|
sigmas = (sigma_max ** (1 / rho) + ramp * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))) ** rho
|
|
elif self.config.use_exponential_sigmas:
|
|
sigma_min = self.config.sigma_min if self.config.sigma_min is not None else sigmas[-1]
|
|
sigma_max = self.config.sigma_max if self.config.sigma_max is not None else sigmas[0]
|
|
sigmas = np.exp(np.linspace(np.log(sigma_max), np.log(sigma_min), num_inference_steps))
|
|
elif self.config.use_beta_sigmas:
|
|
sigma_min = self.config.sigma_min if self.config.sigma_min is not None else sigmas[-1]
|
|
sigma_max = self.config.sigma_max if self.config.sigma_max is not None else sigmas[0]
|
|
alpha, beta = 0.6, 0.6
|
|
ramp = np.linspace(0, 1, num_inference_steps)
|
|
try:
|
|
import torch.distributions as dist
|
|
|
|
b = dist.Beta(alpha, beta)
|
|
ramp = b.sort().values.numpy() # assume single batch sample for schedule
|
|
except Exception:
|
|
pass
|
|
sigmas = sigma_max * (1 - ramp) + sigma_min * ramp
|
|
elif self.config.use_flow_sigmas:
|
|
sigmas = np.linspace(1.0, 1 / 1000, num_inference_steps)
|
|
|
|
# 3. Shifting
|
|
if self.config.use_dynamic_shifting and mu is not None:
|
|
sigmas = mu * sigmas / (1 + (mu - 1) * sigmas)
|
|
elif self.config.shift is not None:
|
|
sigmas = self.config.shift * sigmas / (1 + (self.config.shift - 1) * sigmas)
|
|
|
|
# We handle multi-history expansion
|
|
_a_mat, _b_vec, c_vec = self._get_tableau()
|
|
len(c_vec)
|
|
|
|
sigmas_expanded = []
|
|
for i in range(len(sigmas) - 1):
|
|
s_curr = sigmas[i]
|
|
s_next = sigmas[i + 1]
|
|
for c_val in c_vec:
|
|
sigmas_expanded.append(s_curr + c_val * (s_next - s_curr))
|
|
sigmas_expanded.append(0.0)
|
|
|
|
sigmas_interpolated = np.array(sigmas_expanded)
|
|
# Linear remapping for Flow Matching
|
|
timesteps_expanded = sigmas_interpolated * self.config.num_train_timesteps
|
|
|
|
self.sigmas = torch.from_numpy(sigmas_interpolated).to(device=device, dtype=dtype)
|
|
self.timesteps = torch.from_numpy(timesteps_expanded + self.config.steps_offset).to(device=device, dtype=dtype)
|
|
|
|
self.init_noise_sigma = self.sigmas.max().item() if self.sigmas.numel() > 0 else 1.0
|
|
self._sigmas_cpu = self.sigmas.detach().cpu().numpy()
|
|
self._timesteps_cpu = self.timesteps.detach().cpu().numpy()
|
|
self._step_index = None
|
|
|
|
self.model_outputs = []
|
|
self.sample_at_start_of_step = None
|
|
|
|
@property
|
|
def step_index(self):
|
|
"""
|
|
The index counter for the current timestep. It will increase 1 after each scheduler step.
|
|
"""
|
|
return self._step_index
|
|
|
|
def index_for_timestep(self, timestep, schedule_timesteps=None):
|
|
from .scheduler_utils import index_for_timestep
|
|
if schedule_timesteps is None:
|
|
schedule_timesteps = self.timesteps
|
|
return index_for_timestep(timestep, schedule_timesteps)
|
|
|
|
def _init_step_index(self, timestep):
|
|
if self._step_index is None:
|
|
if isinstance(timestep, torch.Tensor):
|
|
timestep = timestep.to(self.timesteps.device)
|
|
self._step_index = self.index_for_timestep(timestep)
|
|
|
|
def scale_model_input(self, sample: torch.Tensor, timestep: float | torch.Tensor) -> torch.Tensor:
|
|
if self._step_index is None:
|
|
self._init_step_index(timestep)
|
|
if self.config.prediction_type == "flow_prediction":
|
|
return sample
|
|
sigma = self.sigmas[self._step_index]
|
|
return sample / ((sigma**2 + 1) ** 0.5)
|
|
|
|
def step(
|
|
self,
|
|
model_output: torch.Tensor,
|
|
timestep: float | torch.Tensor,
|
|
sample: torch.Tensor,
|
|
return_dict: bool = True,
|
|
) -> SchedulerOutput | tuple:
|
|
self._init_step_index(timestep)
|
|
a_mat, b_vec, c_vec = self._get_tableau()
|
|
num_stages = len(c_vec)
|
|
|
|
stage_index = self._step_index % num_stages
|
|
base_step_index = (self._step_index // num_stages) * num_stages
|
|
|
|
sigma_curr = self._sigmas_cpu[base_step_index]
|
|
sigma_next_idx = min(base_step_index + num_stages, len(self._sigmas_cpu) - 1)
|
|
sigma_next = self._sigmas_cpu[sigma_next_idx]
|
|
|
|
if sigma_next <= 0:
|
|
sigma_t = self.sigmas[self._step_index]
|
|
prediction_type = getattr(self.config, "prediction_type", "epsilon")
|
|
if prediction_type == "epsilon":
|
|
denoised = sample - sigma_t * model_output
|
|
elif prediction_type == "v_prediction":
|
|
alpha_t = 1 / (sigma_t**2 + 1) ** 0.5
|
|
sigma_actual = sigma_t * alpha_t
|
|
denoised = alpha_t * sample - sigma_actual * model_output
|
|
elif prediction_type == "flow_prediction":
|
|
denoised = sample - sigma_t * model_output
|
|
else:
|
|
denoised = model_output
|
|
|
|
if getattr(self.config, "clip_sample", False):
|
|
denoised = denoised.clamp(-self.config.sample_max_value, self.config.sample_max_value)
|
|
|
|
prev_sample = denoised
|
|
self._step_index += 1
|
|
if not return_dict:
|
|
return (prev_sample,)
|
|
return SchedulerOutput(prev_sample=prev_sample)
|
|
|
|
h = sigma_next - sigma_curr
|
|
sigma_t = self.sigmas[self._step_index]
|
|
|
|
prediction_type = getattr(self.config, "prediction_type", "epsilon")
|
|
if prediction_type == "epsilon":
|
|
denoised = sample - sigma_t * model_output
|
|
elif self.config.prediction_type == "v_prediction":
|
|
alpha_t = 1 / (sigma_t**2 + 1) ** 0.5
|
|
sigma_actual = sigma_t * alpha_t
|
|
denoised = alpha_t * sample - sigma_actual * model_output
|
|
# If we want pure x-space x0 from alpha x - sigma v:
|
|
# x0 = x * (1/sqrt(1+sigma^2)) - v * (sigma/sqrt(1+sigma^2))
|
|
# which matches the above.
|
|
elif prediction_type == "flow_prediction":
|
|
denoised = sample - sigma_t * model_output
|
|
elif prediction_type == "sample":
|
|
denoised = model_output
|
|
else:
|
|
raise ValueError(f"prediction_type error: {getattr(self.config, 'prediction_type', 'epsilon')}")
|
|
|
|
if self.config.clip_sample:
|
|
denoised = denoised.clamp(-self.config.sample_max_value, self.config.sample_max_value)
|
|
|
|
# derivative = (x - x0) / sigma
|
|
derivative = (sample - denoised) / sigma_t if sigma_t > 1e-6 else torch.zeros_like(sample)
|
|
|
|
if self.sample_at_start_of_step is None:
|
|
if stage_index > 0:
|
|
# Mid-step fallback for Img2Img/Inpainting
|
|
sigma_next_t = self._sigmas_cpu[self._step_index + 1]
|
|
dt = sigma_next_t - sigma_t
|
|
prev_sample = sample + dt * derivative
|
|
self._step_index += 1
|
|
if not return_dict:
|
|
return (prev_sample,)
|
|
return SchedulerOutput(prev_sample=prev_sample)
|
|
|
|
self.sample_at_start_of_step = sample
|
|
self.model_outputs = [derivative] * stage_index
|
|
|
|
if stage_index == 0:
|
|
self.model_outputs = [derivative]
|
|
self.sample_at_start_of_step = sample
|
|
else:
|
|
self.model_outputs.append(derivative)
|
|
|
|
next_stage_idx = stage_index + 1
|
|
if next_stage_idx < num_stages:
|
|
sum_ak = 0
|
|
for j in range(len(self.model_outputs)):
|
|
sum_ak = sum_ak + a_mat[next_stage_idx][j] * self.model_outputs[j]
|
|
|
|
sigma_next_stage = self.sigmas[min(self._step_index + 1, len(self.sigmas) - 1)]
|
|
|
|
# Update x (unnormalized sample)
|
|
prev_sample = self.sample_at_start_of_step + (sigma_next_stage - sigma_curr) * sum_ak
|
|
else:
|
|
sum_bk = 0
|
|
for j in range(len(self.model_outputs)):
|
|
sum_bk = sum_bk + b_vec[j] * self.model_outputs[j]
|
|
|
|
prev_sample = self.sample_at_start_of_step + h * sum_bk
|
|
|
|
self.model_outputs = []
|
|
self.sample_at_start_of_step = None
|
|
|
|
self._step_index += 1
|
|
|
|
if not return_dict:
|
|
return (prev_sample,)
|
|
return SchedulerOutput(prev_sample=prev_sample)
|
|
|
|
def add_noise(
|
|
self,
|
|
original_samples: torch.Tensor,
|
|
noise: torch.Tensor,
|
|
timesteps: torch.Tensor,
|
|
) -> torch.Tensor:
|
|
from .scheduler_utils import add_noise_to_sample
|
|
return add_noise_to_sample(original_samples, noise, self.sigmas, timesteps, self.timesteps)
|
|
|
|
def __len__(self):
|
|
return self.config.num_train_timesteps
|