automatic/modules/res4lyf/deis_scheduler_alt.py

403 lines
18 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
from .phi_functions import Phi
def get_def_integral_2(a, b, start, end, c):
coeff = (end**3 - start**3) / 3 - (end**2 - start**2) * (a + b) / 2 + (end - start) * a * b
return coeff / ((c - a) * (c - b))
def get_def_integral_3(a, b, c, start, end, d):
coeff = (end**4 - start**4) / 4 - (end**3 - start**3) * (a + b + c) / 3 + (end**2 - start**2) * (a * b + a * c + b * c) / 2 - (end - start) * a * b * c
return coeff / ((d - a) * (d - b) * (d - c))
class RESDEISMultistepScheduler(SchedulerMixin, ConfigMixin):
"""
RESDEISMultistepScheduler: Diffusion Explicit Iterative Sampler with high-order multistep.
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",
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",
solver_order: int = 2,
use_analytic_solution: bool = True,
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.hist_samples = []
self._step_index = None
self._sigmas_cpu = None
self.all_coeffs = []
self.prev_sigmas = []
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(self.config.num_train_timesteps - 1, 0, num_inference_steps, dtype=float).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(num_inference_steps, 0, -step_ratio)).round().copy().astype(float)
timesteps -= step_ratio
else:
raise ValueError(f"timestep_spacing must be one of 'linspace', 'leading', or 'trailing', got {self.config.timestep_spacing}")
if self.config.timestep_spacing == "trailing":
timesteps = np.maximum(timesteps, 0)
sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
log_sigmas_all = np.log(np.maximum(sigmas, 1e-10))
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.sample((num_inference_steps,)).sort().values.numpy()
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)
# Map back to timesteps
if self.config.use_flow_sigmas:
timesteps = sigmas * self.config.num_train_timesteps
else:
timesteps = np.interp(np.log(np.maximum(sigmas, 1e-10)), log_sigmas_all, np.arange(len(log_sigmas_all)))
self.sigmas = torch.from_numpy(np.append(sigmas, 0.0)).to(device=device, dtype=dtype)
self.timesteps = torch.from_numpy(timesteps + 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()
# Precompute coefficients
self.all_coeffs = []
num_steps = len(timesteps)
for i in range(num_steps):
sigma_t = self._sigmas_cpu[i]
sigma_next = self._sigmas_cpu[i + 1]
if sigma_next <= 0:
coeffs = None
else:
current_order = min(i + 1, self.config.solver_order)
if current_order == 1:
coeffs = [sigma_next - sigma_t]
else:
ts = [self._sigmas_cpu[i - j] for j in range(current_order)]
t_next = sigma_next
if current_order == 2:
t_cur, t_prev1 = ts[0], ts[1]
coeff_cur = ((t_next - t_prev1) ** 2 - (t_cur - t_prev1) ** 2) / (2 * (t_cur - t_prev1))
coeff_prev1 = (t_next - t_cur) ** 2 / (2 * (t_prev1 - t_cur))
coeffs = [coeff_cur, coeff_prev1]
elif current_order == 3:
t_cur, t_prev1, t_prev2 = ts[0], ts[1], ts[2]
coeffs = [
get_def_integral_2(t_prev1, t_prev2, t_cur, t_next, t_cur),
get_def_integral_2(t_cur, t_prev2, t_cur, t_next, t_prev1),
get_def_integral_2(t_cur, t_prev1, t_cur, t_next, t_prev2),
]
elif current_order == 4:
t_cur, t_prev1, t_prev2, t_prev3 = ts[0], ts[1], ts[2], ts[3]
coeffs = [
get_def_integral_3(t_prev1, t_prev2, t_prev3, t_cur, t_next, t_cur),
get_def_integral_3(t_cur, t_prev2, t_prev3, t_cur, t_next, t_prev1),
get_def_integral_3(t_cur, t_prev1, t_prev3, t_cur, t_next, t_prev2),
get_def_integral_3(t_cur, t_prev1, t_prev2, t_cur, t_next, t_prev3),
]
else:
coeffs = [(sigma_next - sigma_t) / sigma_t] # Fallback to Euler
self.all_coeffs.append(coeffs)
# Reset history
self.model_outputs = []
self.hist_samples = []
self._step_index = 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:
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:
if self._step_index is None:
self._init_step_index(timestep)
step_index = self._step_index
sigma_t = self.sigmas[step_index]
# RECONSTRUCT X0 (Matching PEC pattern)
if self.config.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
elif self.config.prediction_type == "flow_prediction":
denoised = sample - sigma_t * model_output
elif self.config.prediction_type == "sample":
denoised = model_output
else:
raise ValueError(f"prediction_type error: {self.config.prediction_type}")
if self.config.clip_sample:
denoised = denoised.clamp(-self.config.sample_max_value, self.config.sample_max_value)
if self.config.prediction_type == "flow_prediction":
# Variable Step Adams-Bashforth for Flow Matching
self.model_outputs.append(model_output)
self.prev_sigmas.append(sigma_t)
# Note: deis uses hist_samples for x0? I'll use model_outputs for v.
if len(self.model_outputs) > 4:
self.model_outputs.pop(0)
self.prev_sigmas.pop(0)
dt = self.sigmas[step_index + 1] - sigma_t
v_n = model_output
curr_order = min(len(self.prev_sigmas), 3)
if curr_order == 1:
x_next = sample + dt * v_n
elif curr_order == 2:
sigma_prev = self.prev_sigmas[-2]
dt_prev = sigma_t - sigma_prev
r = dt / dt_prev if abs(dt_prev) > 1e-8 else 0.0
if dt_prev == 0 or r < -0.9 or r > 2.0:
x_next = sample + dt * v_n
else:
c0 = 1 + 0.5 * r
c1 = -0.5 * r
x_next = sample + dt * (c0 * v_n + c1 * self.model_outputs[-2])
else:
# AB2 fallback
sigma_prev = self.prev_sigmas[-2]
dt_prev = sigma_t - sigma_prev
r = dt / dt_prev if abs(dt_prev) > 1e-8 else 0.0
c0 = 1 + 0.5 * r
c1 = -0.5 * r
x_next = sample + dt * (c0 * v_n + c1 * self.model_outputs[-2])
self._step_index += 1
if not return_dict:
return (x_next,)
return SchedulerOutput(prev_sample=x_next)
sigma_next = self.sigmas[step_index + 1]
if self.config.solver_order == 1:
# 1st order step (Euler) in x-space
x_next = (sigma_next / sigma_t) * sample + (1 - sigma_next / sigma_t) * denoised
prev_sample = x_next
else:
# Multistep weights based on phi functions (consistent with RESMultistep)
h = -torch.log(sigma_next / sigma_t) if sigma_t > 0 and sigma_next > 0 else torch.zeros_like(sigma_t)
phi = Phi(h, [0], getattr(self.config, "use_analytic_solution", True))
phi_1 = phi(1)
# History of denoised samples
x0s = [denoised] + self.model_outputs[::-1]
orders = min(len(x0s), self.config.solver_order)
# Force Order 1 at the end of schedule
if self.num_inference_steps is not None and step_index >= self.num_inference_steps - 3:
res = phi_1 * denoised
elif orders == 1:
res = phi_1 * denoised
elif orders == 2:
# Use phi(2) for 2nd order interpolation
h_prev = -np.log(self._sigmas_cpu[step_index] / (self._sigmas_cpu[step_index - 1] + 1e-9))
h_prev_t = torch.tensor(h_prev, device=sample.device, dtype=sample.dtype)
r = h_prev_t / (h + 1e-9)
h_prev = -np.log(self._sigmas_cpu[step_index] / (self._sigmas_cpu[step_index - 1] + 1e-9))
h_prev_t = torch.tensor(h_prev, device=sample.device, dtype=sample.dtype)
r = h_prev_t / (h + 1e-9)
# Hard Restart
if r < 0.5 or r > 2.0:
res = phi_1 * denoised
else:
phi_2 = phi(2)
# Correct Adams-Bashforth-like coefficients: b2 = -phi_2 / r
b2 = -phi_2 / (r + 1e-9)
b1 = phi_1 - b2
res = b1 * x0s[0] + b2 * x0s[1]
elif orders == 3:
# 3rd order with varying step sizes
# 3rd order with varying step sizes
h_p1 = -np.log(self._sigmas_cpu[step_index] / (self._sigmas_cpu[step_index - 1] + 1e-9))
h_p2 = -np.log(self._sigmas_cpu[step_index] / (self._sigmas_cpu[step_index - 2] + 1e-9))
r1 = torch.tensor(h_p1, device=sample.device, dtype=sample.dtype) / (h + 1e-9)
r2 = torch.tensor(h_p2, device=sample.device, dtype=sample.dtype) / (h + 1e-9)
h_p1 = -np.log(self._sigmas_cpu[step_index] / (self._sigmas_cpu[step_index - 1] + 1e-9))
h_p2 = -np.log(self._sigmas_cpu[step_index] / (self._sigmas_cpu[step_index - 2] + 1e-9))
r1 = torch.tensor(h_p1, device=sample.device, dtype=sample.dtype) / (h + 1e-9)
r2 = torch.tensor(h_p2, device=sample.device, dtype=sample.dtype) / (h + 1e-9)
# Hard Restart
if r1 < 0.5 or r1 > 2.0 or r2 < 0.5 or r2 > 2.0:
res = phi_1 * denoised
else:
phi_2, phi_3 = phi(2), phi(3)
denom = r2 - r1 + 1e-9
b3 = (phi_3 + r1 * phi_2) / (r2 * denom)
b2 = -(phi_3 + r2 * phi_2) / (r1 * denom)
b1 = phi_1 - b2 - b3
res = b1 * x0s[0] + b2 * x0s[1] + b3 * x0s[2]
else:
# Fallback to Euler or lower order
res = phi_1 * denoised
# Stable update in x-space
if sigma_next == 0:
x_next = denoised
else:
x_next = torch.exp(-h) * sample + h * res
prev_sample = x_next
# Store state (always store x0)
self.model_outputs.append(denoised)
self.hist_samples.append(sample)
if len(self.model_outputs) > 4:
self.model_outputs.pop(0)
self.hist_samples.pop(0)
if self._step_index is not 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