patrickvonplaten
commited on
Commit
•
871f1b5
1
Parent(s):
a53c3fa
Create my_scheduler.py
Browse files- scheduler/my_scheduler.py +514 -0
scheduler/my_scheduler.py
ADDED
@@ -0,0 +1,514 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2023 UC Berkeley Team and The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
|
15 |
+
# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim
|
16 |
+
|
17 |
+
import math
|
18 |
+
from dataclasses import dataclass
|
19 |
+
from typing import List, Optional, Tuple, Union
|
20 |
+
|
21 |
+
import numpy as np
|
22 |
+
import torch
|
23 |
+
|
24 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
25 |
+
from diffusers.utils import BaseOutput
|
26 |
+
from diffusers.utils.torch_utils import randn_tensor
|
27 |
+
from diffusers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
|
28 |
+
|
29 |
+
|
30 |
+
@dataclass
|
31 |
+
class MySchedulerOutput(BaseOutput):
|
32 |
+
"""
|
33 |
+
Output class for the scheduler's `step` function output.
|
34 |
+
|
35 |
+
Args:
|
36 |
+
prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
|
37 |
+
Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
|
38 |
+
denoising loop.
|
39 |
+
pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
|
40 |
+
The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
|
41 |
+
`pred_original_sample` can be used to preview progress or for guidance.
|
42 |
+
"""
|
43 |
+
|
44 |
+
prev_sample: torch.FloatTensor
|
45 |
+
pred_original_sample: Optional[torch.FloatTensor] = None
|
46 |
+
|
47 |
+
|
48 |
+
def betas_for_alpha_bar(
|
49 |
+
num_diffusion_timesteps,
|
50 |
+
max_beta=0.999,
|
51 |
+
alpha_transform_type="cosine",
|
52 |
+
):
|
53 |
+
"""
|
54 |
+
Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
|
55 |
+
(1-beta) over time from t = [0,1].
|
56 |
+
|
57 |
+
Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
|
58 |
+
to that part of the diffusion process.
|
59 |
+
|
60 |
+
|
61 |
+
Args:
|
62 |
+
num_diffusion_timesteps (`int`): the number of betas to produce.
|
63 |
+
max_beta (`float`): the maximum beta to use; use values lower than 1 to
|
64 |
+
prevent singularities.
|
65 |
+
alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
|
66 |
+
Choose from `cosine` or `exp`
|
67 |
+
|
68 |
+
Returns:
|
69 |
+
betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
|
70 |
+
"""
|
71 |
+
if alpha_transform_type == "cosine":
|
72 |
+
|
73 |
+
def alpha_bar_fn(t):
|
74 |
+
return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
|
75 |
+
|
76 |
+
elif alpha_transform_type == "exp":
|
77 |
+
|
78 |
+
def alpha_bar_fn(t):
|
79 |
+
return math.exp(t * -12.0)
|
80 |
+
|
81 |
+
else:
|
82 |
+
raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
|
83 |
+
|
84 |
+
betas = []
|
85 |
+
for i in range(num_diffusion_timesteps):
|
86 |
+
t1 = i / num_diffusion_timesteps
|
87 |
+
t2 = (i + 1) / num_diffusion_timesteps
|
88 |
+
betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
|
89 |
+
return torch.tensor(betas, dtype=torch.float32)
|
90 |
+
|
91 |
+
|
92 |
+
class MyScheduler(SchedulerMixin, ConfigMixin):
|
93 |
+
"""
|
94 |
+
`MyScheduler` explores the connections between denoising score matching and Langevin dynamics sampling.
|
95 |
+
|
96 |
+
This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
|
97 |
+
methods the library implements for all schedulers such as loading and saving.
|
98 |
+
|
99 |
+
Args:
|
100 |
+
num_train_timesteps (`int`, defaults to 1000):
|
101 |
+
The number of diffusion steps to train the model.
|
102 |
+
beta_start (`float`, defaults to 0.0001):
|
103 |
+
The starting `beta` value of inference.
|
104 |
+
beta_end (`float`, defaults to 0.02):
|
105 |
+
The final `beta` value.
|
106 |
+
beta_schedule (`str`, defaults to `"linear"`):
|
107 |
+
The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
|
108 |
+
`linear`, `scaled_linear`, or `squaredcos_cap_v2`.
|
109 |
+
variance_type (`str`, defaults to `"fixed_small"`):
|
110 |
+
Clip the variance when adding noise to the denoised sample. Choose from `fixed_small`, `fixed_small_log`,
|
111 |
+
`fixed_large`, `fixed_large_log`, `learned` or `learned_range`.
|
112 |
+
clip_sample (`bool`, defaults to `True`):
|
113 |
+
Clip the predicted sample for numerical stability.
|
114 |
+
clip_sample_range (`float`, defaults to 1.0):
|
115 |
+
The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
|
116 |
+
prediction_type (`str`, defaults to `epsilon`, *optional*):
|
117 |
+
Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
|
118 |
+
`sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
|
119 |
+
Video](https://imagen.research.google/video/paper.pdf) paper).
|
120 |
+
thresholding (`bool`, defaults to `False`):
|
121 |
+
Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
|
122 |
+
as Stable Diffusion.
|
123 |
+
dynamic_thresholding_ratio (`float`, defaults to 0.995):
|
124 |
+
The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
|
125 |
+
sample_max_value (`float`, defaults to 1.0):
|
126 |
+
The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
|
127 |
+
timestep_spacing (`str`, defaults to `"leading"`):
|
128 |
+
The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
|
129 |
+
Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
|
130 |
+
steps_offset (`int`, defaults to 0):
|
131 |
+
An offset added to the inference steps. You can use a combination of `offset=1` and
|
132 |
+
`set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
|
133 |
+
Diffusion.
|
134 |
+
"""
|
135 |
+
|
136 |
+
_compatibles = [e.name for e in KarrasDiffusionSchedulers]
|
137 |
+
order = 1
|
138 |
+
|
139 |
+
@register_to_config
|
140 |
+
def __init__(
|
141 |
+
self,
|
142 |
+
num_train_timesteps: int = 1000,
|
143 |
+
beta_start: float = 0.0001,
|
144 |
+
beta_end: float = 0.02,
|
145 |
+
beta_schedule: str = "linear",
|
146 |
+
trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
|
147 |
+
variance_type: str = "fixed_small",
|
148 |
+
clip_sample: bool = True,
|
149 |
+
prediction_type: str = "epsilon",
|
150 |
+
thresholding: bool = False,
|
151 |
+
dynamic_thresholding_ratio: float = 0.995,
|
152 |
+
clip_sample_range: float = 1.0,
|
153 |
+
sample_max_value: float = 1.0,
|
154 |
+
timestep_spacing: str = "leading",
|
155 |
+
steps_offset: int = 0,
|
156 |
+
):
|
157 |
+
if trained_betas is not None:
|
158 |
+
self.betas = torch.tensor(trained_betas, dtype=torch.float32)
|
159 |
+
elif beta_schedule == "linear":
|
160 |
+
self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
|
161 |
+
elif beta_schedule == "scaled_linear":
|
162 |
+
# this schedule is very specific to the latent diffusion model.
|
163 |
+
self.betas = (
|
164 |
+
torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
|
165 |
+
)
|
166 |
+
elif beta_schedule == "squaredcos_cap_v2":
|
167 |
+
# Glide cosine schedule
|
168 |
+
self.betas = betas_for_alpha_bar(num_train_timesteps)
|
169 |
+
elif beta_schedule == "sigmoid":
|
170 |
+
# GeoDiff sigmoid schedule
|
171 |
+
betas = torch.linspace(-6, 6, num_train_timesteps)
|
172 |
+
self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start
|
173 |
+
else:
|
174 |
+
raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
|
175 |
+
|
176 |
+
self.alphas = 1.0 - self.betas
|
177 |
+
self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
|
178 |
+
self.one = torch.tensor(1.0)
|
179 |
+
|
180 |
+
# standard deviation of the initial noise distribution
|
181 |
+
self.init_noise_sigma = 1.0
|
182 |
+
|
183 |
+
# setable values
|
184 |
+
self.custom_timesteps = False
|
185 |
+
self.num_inference_steps = None
|
186 |
+
self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy())
|
187 |
+
|
188 |
+
self.variance_type = variance_type
|
189 |
+
|
190 |
+
def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
|
191 |
+
"""
|
192 |
+
Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
|
193 |
+
current timestep.
|
194 |
+
|
195 |
+
Args:
|
196 |
+
sample (`torch.FloatTensor`):
|
197 |
+
The input sample.
|
198 |
+
timestep (`int`, *optional*):
|
199 |
+
The current timestep in the diffusion chain.
|
200 |
+
|
201 |
+
Returns:
|
202 |
+
`torch.FloatTensor`:
|
203 |
+
A scaled input sample.
|
204 |
+
"""
|
205 |
+
return sample
|
206 |
+
|
207 |
+
def set_timesteps(
|
208 |
+
self,
|
209 |
+
num_inference_steps: Optional[int] = None,
|
210 |
+
device: Union[str, torch.device] = None,
|
211 |
+
timesteps: Optional[List[int]] = None,
|
212 |
+
):
|
213 |
+
"""
|
214 |
+
Sets the discrete timesteps used for the diffusion chain (to be run before inference).
|
215 |
+
|
216 |
+
Args:
|
217 |
+
num_inference_steps (`int`):
|
218 |
+
The number of diffusion steps used when generating samples with a pre-trained model. If used,
|
219 |
+
`timesteps` must be `None`.
|
220 |
+
device (`str` or `torch.device`, *optional*):
|
221 |
+
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
222 |
+
timesteps (`List[int]`, *optional*):
|
223 |
+
Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
|
224 |
+
timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed,
|
225 |
+
`num_inference_steps` must be `None`.
|
226 |
+
|
227 |
+
"""
|
228 |
+
if num_inference_steps is not None and timesteps is not None:
|
229 |
+
raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.")
|
230 |
+
|
231 |
+
if timesteps is not None:
|
232 |
+
for i in range(1, len(timesteps)):
|
233 |
+
if timesteps[i] >= timesteps[i - 1]:
|
234 |
+
raise ValueError("`custom_timesteps` must be in descending order.")
|
235 |
+
|
236 |
+
if timesteps[0] >= self.config.num_train_timesteps:
|
237 |
+
raise ValueError(
|
238 |
+
f"`timesteps` must start before `self.config.train_timesteps`:"
|
239 |
+
f" {self.config.num_train_timesteps}."
|
240 |
+
)
|
241 |
+
|
242 |
+
timesteps = np.array(timesteps, dtype=np.int64)
|
243 |
+
self.custom_timesteps = True
|
244 |
+
else:
|
245 |
+
if num_inference_steps > self.config.num_train_timesteps:
|
246 |
+
raise ValueError(
|
247 |
+
f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
|
248 |
+
f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
|
249 |
+
f" maximal {self.config.num_train_timesteps} timesteps."
|
250 |
+
)
|
251 |
+
|
252 |
+
self.num_inference_steps = num_inference_steps
|
253 |
+
self.custom_timesteps = False
|
254 |
+
|
255 |
+
# "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
|
256 |
+
if self.config.timestep_spacing == "linspace":
|
257 |
+
timesteps = (
|
258 |
+
np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
|
259 |
+
.round()[::-1]
|
260 |
+
.copy()
|
261 |
+
.astype(np.int64)
|
262 |
+
)
|
263 |
+
elif self.config.timestep_spacing == "leading":
|
264 |
+
step_ratio = self.config.num_train_timesteps // self.num_inference_steps
|
265 |
+
# creates integer timesteps by multiplying by ratio
|
266 |
+
# casting to int to avoid issues when num_inference_step is power of 3
|
267 |
+
timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
|
268 |
+
timesteps += self.config.steps_offset
|
269 |
+
elif self.config.timestep_spacing == "trailing":
|
270 |
+
step_ratio = self.config.num_train_timesteps / self.num_inference_steps
|
271 |
+
# creates integer timesteps by multiplying by ratio
|
272 |
+
# casting to int to avoid issues when num_inference_step is power of 3
|
273 |
+
timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64)
|
274 |
+
timesteps -= 1
|
275 |
+
else:
|
276 |
+
raise ValueError(
|
277 |
+
f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
|
278 |
+
)
|
279 |
+
|
280 |
+
self.timesteps = torch.from_numpy(timesteps).to(device)
|
281 |
+
|
282 |
+
def _get_variance(self, t, predicted_variance=None, variance_type=None):
|
283 |
+
prev_t = self.previous_timestep(t)
|
284 |
+
|
285 |
+
alpha_prod_t = self.alphas_cumprod[t]
|
286 |
+
alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one
|
287 |
+
current_beta_t = 1 - alpha_prod_t / alpha_prod_t_prev
|
288 |
+
|
289 |
+
# For t > 0, compute predicted variance βt (see formula (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
|
290 |
+
# and sample from it to get previous sample
|
291 |
+
# x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
|
292 |
+
variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t
|
293 |
+
|
294 |
+
# we always take the log of variance, so clamp it to ensure it's not 0
|
295 |
+
variance = torch.clamp(variance, min=1e-20)
|
296 |
+
|
297 |
+
if variance_type is None:
|
298 |
+
variance_type = self.config.variance_type
|
299 |
+
|
300 |
+
# hacks - were probably added for training stability
|
301 |
+
if variance_type == "fixed_small":
|
302 |
+
variance = variance
|
303 |
+
# for rl-diffuser https://arxiv.org/abs/2205.09991
|
304 |
+
elif variance_type == "fixed_small_log":
|
305 |
+
variance = torch.log(variance)
|
306 |
+
variance = torch.exp(0.5 * variance)
|
307 |
+
elif variance_type == "fixed_large":
|
308 |
+
variance = current_beta_t
|
309 |
+
elif variance_type == "fixed_large_log":
|
310 |
+
# Glide max_log
|
311 |
+
variance = torch.log(current_beta_t)
|
312 |
+
elif variance_type == "learned":
|
313 |
+
return predicted_variance
|
314 |
+
elif variance_type == "learned_range":
|
315 |
+
min_log = torch.log(variance)
|
316 |
+
max_log = torch.log(current_beta_t)
|
317 |
+
frac = (predicted_variance + 1) / 2
|
318 |
+
variance = frac * max_log + (1 - frac) * min_log
|
319 |
+
|
320 |
+
return variance
|
321 |
+
|
322 |
+
def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
|
323 |
+
"""
|
324 |
+
"Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
|
325 |
+
prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
|
326 |
+
s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
|
327 |
+
pixels from saturation at each step. We find that dynamic thresholding results in significantly better
|
328 |
+
photorealism as well as better image-text alignment, especially when using very large guidance weights."
|
329 |
+
|
330 |
+
https://arxiv.org/abs/2205.11487
|
331 |
+
"""
|
332 |
+
dtype = sample.dtype
|
333 |
+
batch_size, channels, *remaining_dims = sample.shape
|
334 |
+
|
335 |
+
if dtype not in (torch.float32, torch.float64):
|
336 |
+
sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
|
337 |
+
|
338 |
+
# Flatten sample for doing quantile calculation along each image
|
339 |
+
sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
|
340 |
+
|
341 |
+
abs_sample = sample.abs() # "a certain percentile absolute pixel value"
|
342 |
+
|
343 |
+
s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
|
344 |
+
s = torch.clamp(
|
345 |
+
s, min=1, max=self.config.sample_max_value
|
346 |
+
) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
|
347 |
+
s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
|
348 |
+
sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
|
349 |
+
|
350 |
+
sample = sample.reshape(batch_size, channels, *remaining_dims)
|
351 |
+
sample = sample.to(dtype)
|
352 |
+
|
353 |
+
return sample
|
354 |
+
|
355 |
+
def step(
|
356 |
+
self,
|
357 |
+
model_output: torch.FloatTensor,
|
358 |
+
timestep: int,
|
359 |
+
sample: torch.FloatTensor,
|
360 |
+
generator=None,
|
361 |
+
return_dict: bool = True,
|
362 |
+
) -> Union[MySchedulerOutput, Tuple]:
|
363 |
+
"""
|
364 |
+
Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
|
365 |
+
process from the learned model outputs (most often the predicted noise).
|
366 |
+
|
367 |
+
Args:
|
368 |
+
model_output (`torch.FloatTensor`):
|
369 |
+
The direct output from learned diffusion model.
|
370 |
+
timestep (`float`):
|
371 |
+
The current discrete timestep in the diffusion chain.
|
372 |
+
sample (`torch.FloatTensor`):
|
373 |
+
A current instance of a sample created by the diffusion process.
|
374 |
+
generator (`torch.Generator`, *optional*):
|
375 |
+
A random number generator.
|
376 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
377 |
+
Whether or not to return a [`~schedulers.scheduling_ddpm.MySchedulerOutput`] or `tuple`.
|
378 |
+
|
379 |
+
Returns:
|
380 |
+
[`~schedulers.scheduling_ddpm.MySchedulerOutput`] or `tuple`:
|
381 |
+
If return_dict is `True`, [`~schedulers.scheduling_ddpm.MySchedulerOutput`] is returned, otherwise a
|
382 |
+
tuple is returned where the first element is the sample tensor.
|
383 |
+
|
384 |
+
"""
|
385 |
+
t = timestep
|
386 |
+
|
387 |
+
prev_t = self.previous_timestep(t)
|
388 |
+
|
389 |
+
if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]:
|
390 |
+
model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1)
|
391 |
+
else:
|
392 |
+
predicted_variance = None
|
393 |
+
|
394 |
+
# 1. compute alphas, betas
|
395 |
+
alpha_prod_t = self.alphas_cumprod[t]
|
396 |
+
alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one
|
397 |
+
beta_prod_t = 1 - alpha_prod_t
|
398 |
+
beta_prod_t_prev = 1 - alpha_prod_t_prev
|
399 |
+
current_alpha_t = alpha_prod_t / alpha_prod_t_prev
|
400 |
+
current_beta_t = 1 - current_alpha_t
|
401 |
+
|
402 |
+
# 2. compute predicted original sample from predicted noise also called
|
403 |
+
# "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
|
404 |
+
if self.config.prediction_type == "epsilon":
|
405 |
+
pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
|
406 |
+
elif self.config.prediction_type == "sample":
|
407 |
+
pred_original_sample = model_output
|
408 |
+
elif self.config.prediction_type == "v_prediction":
|
409 |
+
pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
|
410 |
+
else:
|
411 |
+
raise ValueError(
|
412 |
+
f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or"
|
413 |
+
" `v_prediction` for the MyScheduler."
|
414 |
+
)
|
415 |
+
|
416 |
+
# 3. Clip or threshold "predicted x_0"
|
417 |
+
if self.config.thresholding:
|
418 |
+
pred_original_sample = self._threshold_sample(pred_original_sample)
|
419 |
+
elif self.config.clip_sample:
|
420 |
+
pred_original_sample = pred_original_sample.clamp(
|
421 |
+
-self.config.clip_sample_range, self.config.clip_sample_range
|
422 |
+
)
|
423 |
+
|
424 |
+
# 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
|
425 |
+
# See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
|
426 |
+
pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t
|
427 |
+
current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t
|
428 |
+
|
429 |
+
# 5. Compute predicted previous sample µ_t
|
430 |
+
# See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
|
431 |
+
pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample
|
432 |
+
|
433 |
+
# 6. Add noise
|
434 |
+
variance = 0
|
435 |
+
if t > 0:
|
436 |
+
device = model_output.device
|
437 |
+
variance_noise = randn_tensor(
|
438 |
+
model_output.shape, generator=generator, device=device, dtype=model_output.dtype
|
439 |
+
)
|
440 |
+
if self.variance_type == "fixed_small_log":
|
441 |
+
variance = self._get_variance(t, predicted_variance=predicted_variance) * variance_noise
|
442 |
+
elif self.variance_type == "learned_range":
|
443 |
+
variance = self._get_variance(t, predicted_variance=predicted_variance)
|
444 |
+
variance = torch.exp(0.5 * variance) * variance_noise
|
445 |
+
else:
|
446 |
+
variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * variance_noise
|
447 |
+
|
448 |
+
pred_prev_sample = pred_prev_sample + variance
|
449 |
+
|
450 |
+
if not return_dict:
|
451 |
+
return (pred_prev_sample,)
|
452 |
+
|
453 |
+
return MySchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample)
|
454 |
+
|
455 |
+
def add_noise(
|
456 |
+
self,
|
457 |
+
original_samples: torch.FloatTensor,
|
458 |
+
noise: torch.FloatTensor,
|
459 |
+
timesteps: torch.IntTensor,
|
460 |
+
) -> torch.FloatTensor:
|
461 |
+
# Make sure alphas_cumprod and timestep have same device and dtype as original_samples
|
462 |
+
alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype)
|
463 |
+
timesteps = timesteps.to(original_samples.device)
|
464 |
+
|
465 |
+
sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
|
466 |
+
sqrt_alpha_prod = sqrt_alpha_prod.flatten()
|
467 |
+
while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
|
468 |
+
sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
|
469 |
+
|
470 |
+
sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
|
471 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
|
472 |
+
while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
|
473 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
|
474 |
+
|
475 |
+
noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
|
476 |
+
return noisy_samples
|
477 |
+
|
478 |
+
def get_velocity(
|
479 |
+
self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
|
480 |
+
) -> torch.FloatTensor:
|
481 |
+
# Make sure alphas_cumprod and timestep have same device and dtype as sample
|
482 |
+
alphas_cumprod = self.alphas_cumprod.to(device=sample.device, dtype=sample.dtype)
|
483 |
+
timesteps = timesteps.to(sample.device)
|
484 |
+
|
485 |
+
sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
|
486 |
+
sqrt_alpha_prod = sqrt_alpha_prod.flatten()
|
487 |
+
while len(sqrt_alpha_prod.shape) < len(sample.shape):
|
488 |
+
sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
|
489 |
+
|
490 |
+
sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
|
491 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
|
492 |
+
while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
|
493 |
+
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
|
494 |
+
|
495 |
+
velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
|
496 |
+
return velocity
|
497 |
+
|
498 |
+
def __len__(self):
|
499 |
+
return self.config.num_train_timesteps
|
500 |
+
|
501 |
+
def previous_timestep(self, timestep):
|
502 |
+
if self.custom_timesteps:
|
503 |
+
index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0]
|
504 |
+
if index == self.timesteps.shape[0] - 1:
|
505 |
+
prev_t = torch.tensor(-1)
|
506 |
+
else:
|
507 |
+
prev_t = self.timesteps[index + 1]
|
508 |
+
else:
|
509 |
+
num_inference_steps = (
|
510 |
+
self.num_inference_steps if self.num_inference_steps else self.config.num_train_timesteps
|
511 |
+
)
|
512 |
+
prev_t = timestep - self.config.num_train_timesteps // num_inference_steps
|
513 |
+
|
514 |
+
return prev_t
|