File size: 13,757 Bytes
599a397 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 |
# Copyright (c) MONAI Consortium
# 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.
#
# =========================================================================
# Adapted from https://github.com/hpcaitech/Open-Sora/blob/main/opensora/schedulers/rf/rectified_flow.py
# which has the following license:
# https://github.com/hpcaitech/Open-Sora/blob/main/LICENSE
# 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 __future__ import annotations
from typing import Union
import numpy as np
import torch
from torch.distributions import LogisticNormal
from monai.utils import StrEnum
from .ddpm import DDPMPredictionType
from .scheduler import Scheduler
class RFlowPredictionType(StrEnum):
"""
Set of valid prediction type names for the RFlow scheduler's `prediction_type` argument.
v_prediction: velocity prediction, see section 2.4 https://imagen.research.google/video/paper.pdf
"""
V_PREDICTION = DDPMPredictionType.V_PREDICTION
def timestep_transform(
t, input_img_size_numel, base_img_size_numel=32 * 32 * 32, scale=1.0, num_train_timesteps=1000, spatial_dim=3
):
"""
Applies a transformation to the timestep based on image resolution scaling.
Args:
t (torch.Tensor): The original timestep(s).
input_img_size_numel (torch.Tensor): The input image's size (H * W * D).
base_img_size_numel (int): reference H*W*D size, usually smaller than input_img_size_numel.
scale (float): Scaling factor for the transformation.
num_train_timesteps (int): Total number of training timesteps.
spatial_dim (int): Number of spatial dimensions in the image.
Returns:
torch.Tensor: Transformed timestep(s).
"""
t = t / num_train_timesteps
ratio_space = (input_img_size_numel / base_img_size_numel) ** (1.0 / spatial_dim)
ratio = ratio_space * scale
new_t = ratio * t / (1 + (ratio - 1) * t)
new_t = new_t * num_train_timesteps
return new_t
class RFlowScheduler(Scheduler):
"""
A rectified flow scheduler for guiding the diffusion process in a generative model.
Supports uniform and logit-normal sampling methods, timestep transformation for
different resolutions, and noise addition during diffusion.
Args:
num_train_timesteps (int): Total number of training timesteps.
use_discrete_timesteps (bool): Whether to use discrete timesteps.
sample_method (str): Training time step sampling method ('uniform' or 'logit-normal').
loc (float): Location parameter for logit-normal distribution, used only if sample_method='logit-normal'.
scale (float): Scale parameter for logit-normal distribution, used only if sample_method='logit-normal'.
use_timestep_transform (bool): Whether to apply timestep transformation.
If true, there will be more inference timesteps at early(noisy) stages for larger image volumes.
transform_scale (float): Scaling factor for timestep transformation, used only if use_timestep_transform=True.
steps_offset (int): Offset added to computed timesteps, used only if use_timestep_transform=True.
base_img_size_numel (int): Reference image volume size for scaling, used only if use_timestep_transform=True.
spatial_dim (int): 2 or 3, incidcating 2D or 3D images, used only if use_timestep_transform=True.
Example:
.. code-block:: python
# define a scheduler
noise_scheduler = RFlowScheduler(
num_train_timesteps = 1000,
use_discrete_timesteps = True,
sample_method = 'logit-normal',
use_timestep_transform = True,
base_img_size_numel = 32 * 32 * 32,
spatial_dim = 3
)
# during training
inputs = torch.ones(2,4,64,64,32)
noise = torch.randn_like(inputs)
timesteps = noise_scheduler.sample_timesteps(inputs)
noisy_inputs = noise_scheduler.add_noise(original_samples=inputs, noise=noise, timesteps=timesteps)
predicted_velocity = diffusion_unet(
x=noisy_inputs,
timesteps=timesteps
)
loss = loss_l1(predicted_velocity, (inputs - noise))
# during inference
noisy_inputs = torch.randn(2,4,64,64,32)
input_img_size_numel = torch.prod(torch.tensor(noisy_inputs.shape[-3:])
noise_scheduler.set_timesteps(
num_inference_steps=30, input_img_size_numel=input_img_size_numel)
)
all_next_timesteps = torch.cat(
(noise_scheduler.timesteps[1:], torch.tensor([0], dtype=noise_scheduler.timesteps.dtype))
)
for t, next_t in tqdm(
zip(noise_scheduler.timesteps, all_next_timesteps),
total=min(len(noise_scheduler.timesteps), len(all_next_timesteps)),
):
predicted_velocity = diffusion_unet(
x=noisy_inputs,
timesteps=timesteps
)
noisy_inputs, _ = noise_scheduler.step(predicted_velocity, t, noisy_inputs, next_t)
final_output = noisy_inputs
"""
def __init__(
self,
num_train_timesteps: int = 1000,
use_discrete_timesteps: bool = True,
sample_method: str = "uniform",
loc: float = 0.0,
scale: float = 1.0,
use_timestep_transform: bool = False,
transform_scale: float = 1.0,
steps_offset: int = 0,
base_img_size_numel: int = 32 * 32 * 32,
spatial_dim: int = 3,
):
# rectified flow only accepts velocity prediction
self.prediction_type = RFlowPredictionType.V_PREDICTION
self.num_train_timesteps = num_train_timesteps
self.use_discrete_timesteps = use_discrete_timesteps
self.base_img_size_numel = base_img_size_numel
self.spatial_dim = spatial_dim
# sample method
if sample_method not in ["uniform", "logit-normal"]:
raise ValueError(
f"sample_method = {sample_method}, which has to be chosen from ['uniform', 'logit-normal']."
)
self.sample_method = sample_method
if sample_method == "logit-normal":
self.distribution = LogisticNormal(torch.tensor([loc]), torch.tensor([scale]))
self.sample_t = lambda x: self.distribution.sample((x.shape[0],))[:, 0].to(x.device)
# timestep transform
self.use_timestep_transform = use_timestep_transform
self.transform_scale = transform_scale
self.steps_offset = steps_offset
def add_noise(self, original_samples: torch.Tensor, noise: torch.Tensor, timesteps: torch.Tensor) -> torch.Tensor:
"""
Add noise to the original samples.
Args:
original_samples: original samples
noise: noise to add to samples
timesteps: timesteps tensor with shape of (N,), indicating the timestep to be computed for each sample.
Returns:
noisy_samples: sample with added noise
"""
timepoints: torch.Tensor = timesteps.float() / self.num_train_timesteps
timepoints = 1 - timepoints # [1,1/1000]
# expand timepoint to noise shape
if noise.ndim == 5:
timepoints = timepoints[..., None, None, None, None].expand(-1, *noise.shape[1:])
elif noise.ndim == 4:
timepoints = timepoints[..., None, None, None].expand(-1, *noise.shape[1:])
else:
raise ValueError(f"noise tensor has to be 4D or 5D tensor, yet got shape of {noise.shape}")
noisy_samples: torch.Tensor = timepoints * original_samples + (1 - timepoints) * noise
return noisy_samples
def set_timesteps(
self,
num_inference_steps: int,
device: str | torch.device | None = None,
input_img_size_numel: int | None = None,
) -> None:
"""
Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference.
Args:
num_inference_steps: number of diffusion steps used when generating samples with a pre-trained model.
device: target device to put the data.
input_img_size_numel: int, H*W*D of the image, used with self.use_timestep_transform is True.
"""
if num_inference_steps > self.num_train_timesteps or num_inference_steps < 1:
raise ValueError(
f"`num_inference_steps`: {num_inference_steps} should be at least 1, "
"and cannot be larger than `self.num_train_timesteps`:"
f" {self.num_train_timesteps} as the unet model trained with this scheduler can only handle"
f" maximal {self.num_train_timesteps} timesteps."
)
self.num_inference_steps = num_inference_steps
# prepare timesteps
timesteps = [
(1.0 - i / self.num_inference_steps) * self.num_train_timesteps for i in range(self.num_inference_steps)
]
if self.use_discrete_timesteps:
timesteps = [int(round(t)) for t in timesteps]
if self.use_timestep_transform:
timesteps = [
timestep_transform(
t,
input_img_size_numel=input_img_size_numel,
base_img_size_numel=self.base_img_size_numel,
num_train_timesteps=self.num_train_timesteps,
spatial_dim=self.spatial_dim,
)
for t in timesteps
]
timesteps_np = np.array(timesteps).astype(np.float16)
if self.use_discrete_timesteps:
timesteps_np = timesteps_np.astype(np.int64)
self.timesteps = torch.from_numpy(timesteps_np).to(device)
self.timesteps += self.steps_offset
def sample_timesteps(self, x_start):
"""
Randomly samples training timesteps using the chosen sampling method.
Args:
x_start (torch.Tensor): The input tensor for sampling.
Returns:
torch.Tensor: Sampled timesteps.
"""
if self.sample_method == "uniform":
t = torch.rand((x_start.shape[0],), device=x_start.device) * self.num_train_timesteps
elif self.sample_method == "logit-normal":
t = self.sample_t(x_start) * self.num_train_timesteps
if self.use_discrete_timesteps:
t = t.long()
if self.use_timestep_transform:
input_img_size_numel = torch.prod(torch.tensor(x_start.shape[2:]))
t = timestep_transform(
t,
input_img_size_numel=input_img_size_numel,
base_img_size_numel=self.base_img_size_numel,
num_train_timesteps=self.num_train_timesteps,
spatial_dim=len(x_start.shape) - 2,
)
return t
def step(
self, model_output: torch.Tensor, timestep: int, sample: torch.Tensor, next_timestep: Union[int, None] = None
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Predicts the next sample in the diffusion process.
Args:
model_output (torch.Tensor): Output from the trained diffusion model.
timestep (int): Current timestep in the diffusion chain.
sample (torch.Tensor): Current sample in the process.
next_timestep (Union[int, None]): Optional next timestep.
Returns:
tuple[torch.Tensor, torch.Tensor]: Predicted sample at the next step and additional info.
"""
# Ensure num_inference_steps exists and is a valid integer
if not hasattr(self, "num_inference_steps") or not isinstance(self.num_inference_steps, int):
raise AttributeError(
"num_inference_steps is missing or not an integer in the class."
"Please run self.set_timesteps(num_inference_steps,device,input_img_size_numel) to set it."
)
v_pred = model_output
if next_timestep is not None:
next_timestep = int(next_timestep)
dt: float = (
float(timestep - next_timestep) / self.num_train_timesteps
) # Now next_timestep is guaranteed to be int
else:
dt = (
1.0 / float(self.num_inference_steps) if self.num_inference_steps > 0 else 0.0
) # Avoid division by zero
pred_post_sample = sample + v_pred * dt
pred_original_sample = sample + v_pred * timestep / self.num_train_timesteps
return pred_post_sample, pred_original_sample
|