134 lines
4.9 KiB
Python
134 lines
4.9 KiB
Python
# Copyright 2022 Katherine Crowson and The HuggingFace Team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
from typing import List, Union
|
|
|
|
import numpy as np
|
|
import torch
|
|
|
|
from scipy import integrate
|
|
|
|
from ..configuration_utils import ConfigMixin, register_to_config
|
|
from .scheduling_utils import SchedulerMixin
|
|
|
|
|
|
class LMSDiscreteScheduler(SchedulerMixin, ConfigMixin):
|
|
@register_to_config
|
|
def __init__(
|
|
self,
|
|
num_train_timesteps=1000,
|
|
beta_start=0.0001,
|
|
beta_end=0.02,
|
|
beta_schedule="linear",
|
|
trained_betas=None,
|
|
timestep_values=None,
|
|
tensor_format="pt",
|
|
):
|
|
"""
|
|
Linear Multistep Scheduler for discrete beta schedules.
|
|
Based on the original k-diffusion implementation by Katherine Crowson:
|
|
https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L181
|
|
"""
|
|
|
|
if beta_schedule == "linear":
|
|
self.betas = np.linspace(beta_start, beta_end, num_train_timesteps, dtype=np.float32)
|
|
elif beta_schedule == "scaled_linear":
|
|
# this schedule is very specific to the latent diffusion model.
|
|
self.betas = np.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=np.float32) ** 2
|
|
else:
|
|
raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
|
|
|
|
self.alphas = 1.0 - self.betas
|
|
self.alphas_cumprod = np.cumprod(self.alphas, axis=0)
|
|
|
|
self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
|
|
|
|
# setable values
|
|
self.num_inference_steps = None
|
|
self.timesteps = np.arange(0, num_train_timesteps)[::-1].copy()
|
|
self.derivatives = []
|
|
|
|
self.tensor_format = tensor_format
|
|
self.set_format(tensor_format=tensor_format)
|
|
|
|
def get_lms_coefficient(self, order, t, current_order):
|
|
"""
|
|
Compute a linear multistep coefficient
|
|
"""
|
|
|
|
def lms_derivative(tau):
|
|
prod = 1.0
|
|
for k in range(order):
|
|
if current_order == k:
|
|
continue
|
|
prod *= (tau - self.sigmas[t - k]) / (self.sigmas[t - current_order] - self.sigmas[t - k])
|
|
return prod
|
|
|
|
integrated_coeff = integrate.quad(lms_derivative, self.sigmas[t], self.sigmas[t + 1], epsrel=1e-4)[0]
|
|
|
|
return integrated_coeff
|
|
|
|
def set_timesteps(self, num_inference_steps):
|
|
self.num_inference_steps = num_inference_steps
|
|
self.timesteps = np.linspace(self.num_train_timesteps - 1, 0, num_inference_steps, dtype=float)
|
|
|
|
low_idx = np.floor(self.timesteps).astype(int)
|
|
high_idx = np.ceil(self.timesteps).astype(int)
|
|
frac = np.mod(self.timesteps, 1.0)
|
|
sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
|
|
sigmas = (1 - frac) * sigmas[low_idx] + frac * sigmas[high_idx]
|
|
self.sigmas = np.concatenate([sigmas, [0.0]])
|
|
|
|
self.derivatives = []
|
|
|
|
self.set_format(tensor_format=self.tensor_format)
|
|
|
|
def step(
|
|
self,
|
|
model_output: Union[torch.FloatTensor, np.ndarray],
|
|
timestep: int,
|
|
sample: Union[torch.FloatTensor, np.ndarray],
|
|
order: int = 4,
|
|
):
|
|
sigma = self.sigmas[timestep]
|
|
|
|
# 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
|
|
pred_original_sample = sample - sigma * model_output
|
|
|
|
# 2. Convert to an ODE derivative
|
|
derivative = (sample - pred_original_sample) / sigma
|
|
self.derivatives.append(derivative)
|
|
if len(self.derivatives) > order:
|
|
self.derivatives.pop(0)
|
|
|
|
# 3. Compute linear multistep coefficients
|
|
order = min(timestep + 1, order)
|
|
lms_coeffs = [self.get_lms_coefficient(order, timestep, curr_order) for curr_order in range(order)]
|
|
|
|
# 4. Compute previous sample based on the derivatives path
|
|
prev_sample = sample + sum(
|
|
coeff * derivative for coeff, derivative in zip(lms_coeffs, reversed(self.derivatives))
|
|
)
|
|
|
|
return {"prev_sample": prev_sample}
|
|
|
|
def add_noise(self, original_samples, noise, timesteps):
|
|
alpha_prod = self.alphas_cumprod[timesteps]
|
|
alpha_prod = self.match_shape(alpha_prod, original_samples)
|
|
|
|
noisy_samples = (alpha_prod**0.5) * original_samples + ((1 - alpha_prod) ** 0.5) * noise
|
|
return noisy_samples
|
|
|
|
def __len__(self):
|
|
return self.config.num_train_timesteps
|