157 lines
5.9 KiB
Plaintext
157 lines
5.9 KiB
Plaintext
# Copyright 2022 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.
|
|
|
|
|
|
import torch
|
|
|
|
import tqdm
|
|
|
|
from ..pipeline_utils import DiffusionPipeline
|
|
|
|
|
|
class PNDM(DiffusionPipeline):
|
|
def __init__(self, unet, noise_scheduler):
|
|
super().__init__()
|
|
noise_scheduler = noise_scheduler.set_format("pt")
|
|
self.register_modules(unet=unet, noise_scheduler=noise_scheduler)
|
|
|
|
def __call__(self, batch_size=1, generator=None, torch_device=None, num_inference_steps=50):
|
|
# eta corresponds to η in paper and should be between [0, 1]
|
|
if torch_device is None:
|
|
torch_device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
|
num_trained_timesteps = self.noise_scheduler.timesteps
|
|
inference_step_times = range(0, num_trained_timesteps, num_trained_timesteps // num_inference_steps)
|
|
|
|
self.unet.to(torch_device)
|
|
|
|
# Sample gaussian noise to begin loop
|
|
image = torch.randn(
|
|
(batch_size, self.unet.in_channels, self.unet.resolution, self.unet.resolution),
|
|
generator=generator,
|
|
)
|
|
image = image.to(torch_device)
|
|
|
|
seq = list(inference_step_times)
|
|
seq_next = [-1] + list(seq[:-1])
|
|
model = self.unet
|
|
|
|
warmup_steps = [len(seq) - (i // 4 + 1) for i in range(3 * 4)]
|
|
|
|
ets = []
|
|
prev_image = image
|
|
for i, step_idx in enumerate(warmup_steps):
|
|
i = seq[step_idx]
|
|
j = seq_next[step_idx]
|
|
|
|
t = (torch.ones(image.shape[0]) * i)
|
|
t_next = (torch.ones(image.shape[0]) * j)
|
|
|
|
residual = model(image.to("cuda"), t.to("cuda"))
|
|
residual = residual.to("cpu")
|
|
|
|
image = image.to("cpu")
|
|
image = self.noise_scheduler.transfer(prev_image.to("cpu"), t_list[0], t_list[1], residual)
|
|
|
|
if i % 4 == 0:
|
|
ets.append(residual)
|
|
prev_image = image
|
|
|
|
for
|
|
|
|
ets = []
|
|
step_idx = len(seq) - 1
|
|
while step_idx >= 0:
|
|
i = seq[step_idx]
|
|
j = seq_next[step_idx]
|
|
|
|
t = (torch.ones(image.shape[0]) * i)
|
|
t_next = (torch.ones(image.shape[0]) * j)
|
|
|
|
residual = model(image.to("cuda"), t.to("cuda"))
|
|
residual = residual.to("cpu")
|
|
|
|
t_list = [t, (t+t_next)/2, t_next]
|
|
|
|
ets.append(residual)
|
|
if len(ets) <= 3:
|
|
image = image.to("cpu")
|
|
x_2 = self.noise_scheduler.transfer(image.to("cpu"), t_list[0], t_list[1], residual)
|
|
|
|
e_2 = model(x_2.to("cuda"), t_list[1].to("cuda")).to("cpu")
|
|
x_3 = self.noise_scheduler.transfer(image, t_list[0], t_list[1], e_2)
|
|
e_3 = model(x_3.to("cuda"), t_list[1].to("cuda")).to("cpu")
|
|
x_4 = self.noise_scheduler.transfer(image, t_list[0], t_list[2], e_3)
|
|
e_4 = model(x_4.to("cuda"), t_list[2].to("cuda")).to("cpu")
|
|
residual = (1 / 6) * (residual + 2 * e_2 + 2 * e_3 + e_4)
|
|
else:
|
|
residual = (1 / 24) * (55 * ets[-1] - 59 * ets[-2] + 37 * ets[-3] - 9 * ets[-4])
|
|
|
|
img_next = self.noise_scheduler.transfer(image.to("cpu"), t, t_next, residual)
|
|
image = img_next
|
|
|
|
step_idx = step_idx - 1
|
|
|
|
# if len(prev_noises) in [1, 2]:
|
|
# t = (t + t_next) / 2
|
|
# elif len(prev_noises) == 3:
|
|
# t = t_next / 2
|
|
|
|
# if len(prev_noises) == 0:
|
|
# ets.append(residual)
|
|
#
|
|
# if len(ets) > 3:
|
|
# residual = (1 / 24) * (55 * ets[-1] - 59 * ets[-2] + 37 * ets[-3] - 9 * ets[-4])
|
|
# step_idx = step_idx - 1
|
|
# elif len(ets) <= 3 and len(prev_noises) == 3:
|
|
# residual = (1 / 6) * (prev_noises[-3] + 2 * prev_noises[-2] + 2 * prev_noises[-1] + residual)
|
|
# prev_noises = []
|
|
# step_idx = step_idx - 1
|
|
# elif len(ets) <= 3 and len(prev_noises) < 3:
|
|
# prev_noises.append(residual)
|
|
# if len(prev_noises) < 2:
|
|
# t_next = (t + t_next) / 2
|
|
#
|
|
# image = self.noise_scheduler.transfer(image.to("cpu"), t, t_next, residual)
|
|
|
|
return image
|
|
|
|
# See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
|
|
# Ideally, read DDIM paper in-detail understanding
|
|
|
|
# Notation (<variable name> -> <name in paper>
|
|
# - pred_noise_t -> e_theta(x_t, t)
|
|
# - pred_original_image -> f_theta(x_t, t) or x_0
|
|
# - std_dev_t -> sigma_t
|
|
# - eta -> η
|
|
# - pred_image_direction -> "direction pointingc to x_t"
|
|
# - pred_prev_image -> "x_t-1"
|
|
# for t in tqdm.tqdm(reversed(range(num_inference_steps)), total=num_inference_steps):
|
|
# 1. predict noise residual
|
|
# with torch.no_grad():
|
|
# residual = self.unet(image, inference_step_times[t])
|
|
#
|
|
# 2. predict previous mean of image x_t-1
|
|
# pred_prev_image = self.noise_scheduler.step(residual, image, t, num_inference_steps, eta)
|
|
#
|
|
# 3. optionally sample variance
|
|
# variance = 0
|
|
# if eta > 0:
|
|
# noise = torch.randn(image.shape, generator=generator).to(image.device)
|
|
# variance = self.noise_scheduler.get_variance(t, num_inference_steps).sqrt() * eta * noise
|
|
#
|
|
# 4. set current image to prev_image: x_t -> x_t-1
|
|
# image = pred_prev_image + variance
|