diff options
| -rw-r--r-- | environment.yaml | 18 | ||||
| -rw-r--r-- | infer.py | 7 | ||||
| -rw-r--r-- | schedulers/scheduling_unipc_multistep.py | 615 | ||||
| -rw-r--r-- | train_dreambooth.py | 4 | ||||
| -rw-r--r-- | train_lora.py | 4 | ||||
| -rw-r--r-- | train_ti.py | 6 | ||||
| -rw-r--r-- | training/functional.py | 3 | ||||
| -rw-r--r-- | training/strategy/dreambooth.py | 8 | ||||
| -rw-r--r-- | training/strategy/lora.py | 2 | ||||
| -rw-r--r-- | training/strategy/ti.py | 4 |
10 files changed, 27 insertions, 644 deletions
diff --git a/environment.yaml b/environment.yaml index f5632bf..8010c09 100644 --- a/environment.yaml +++ b/environment.yaml | |||
| @@ -1,28 +1,24 @@ | |||
| 1 | name: ldd | 1 | name: ldd |
| 2 | channels: | 2 | channels: |
| 3 | - pytorch | 3 | - pytorch-nightly |
| 4 | - nvidia | 4 | - nvidia |
| 5 | - xformers/label/dev | 5 | - xformers/label/dev |
| 6 | - defaults | 6 | - defaults |
| 7 | dependencies: | 7 | dependencies: |
| 8 | - cudatoolkit=11.3 | 8 | - cudatoolkit=11.7 |
| 9 | - libcufile=1.4.0.31 | ||
| 10 | - matplotlib=3.6.2 | 9 | - matplotlib=3.6.2 |
| 11 | - numpy=1.23.4 | 10 | - numpy=1.23.4 |
| 12 | - pip=22.3.1 | 11 | - pip=22.3.1 |
| 13 | - python=3.10.8 | 12 | - python=3.10.8 |
| 14 | - pytorch=1.13.1=*cuda* | 13 | - pytorch=2.0.0.dev20230216=*cuda* |
| 15 | - torchvision=0.14.1 | 14 | - torchvision=0.15.0.dev20230216 |
| 16 | - pip: | 15 | - pip: |
| 17 | - -e . | 16 | - -e . |
| 18 | - -e git+https://github.com/huggingface/diffusers#egg=diffusers | 17 | - -e git+https://github.com/huggingface/diffusers#egg=diffusers |
| 19 | - -e git+https://github.com/cloneofsimo/lora#egg=lora-diffusion | 18 | - accelerate==0.16.0 |
| 20 | - accelerate==0.15.0 | ||
| 21 | - bitsandbytes==0.37.0 | 19 | - bitsandbytes==0.37.0 |
| 22 | - python-slugify>=6.1.2 | 20 | - python-slugify>=6.1.2 |
| 23 | - safetensors==0.2.7 | 21 | - safetensors==0.2.8 |
| 24 | - setuptools==65.6.3 | 22 | - setuptools==65.6.3 |
| 25 | - test-tube>=0.7.5 | 23 | - test-tube>=0.7.5 |
| 26 | - transformers==4.25.1 | 24 | - transformers==4.26.1 |
| 27 | - triton==2.0.0.dev20221202 | ||
| 28 | - xformers==0.0.17.dev443 | ||
| @@ -21,7 +21,8 @@ from diffusers import ( | |||
| 21 | LMSDiscreteScheduler, | 21 | LMSDiscreteScheduler, |
| 22 | EulerAncestralDiscreteScheduler, | 22 | EulerAncestralDiscreteScheduler, |
| 23 | KDPM2DiscreteScheduler, | 23 | KDPM2DiscreteScheduler, |
| 24 | KDPM2AncestralDiscreteScheduler | 24 | KDPM2AncestralDiscreteScheduler, |
| 25 | UniPCMultistepScheduler | ||
| 25 | ) | 26 | ) |
| 26 | from transformers import CLIPTextModel | 27 | from transformers import CLIPTextModel |
| 27 | 28 | ||
| @@ -29,7 +30,6 @@ from data.keywords import prompt_to_keywords, keywords_to_prompt | |||
| 29 | from models.clip.embeddings import patch_managed_embeddings | 30 | from models.clip.embeddings import patch_managed_embeddings |
| 30 | from models.clip.tokenizer import MultiCLIPTokenizer | 31 | from models.clip.tokenizer import MultiCLIPTokenizer |
| 31 | from pipelines.stable_diffusion.vlpn_stable_diffusion import VlpnStableDiffusion | 32 | from pipelines.stable_diffusion.vlpn_stable_diffusion import VlpnStableDiffusion |
| 32 | from schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler | ||
| 33 | from util import load_config, load_embeddings_from_dir | 33 | from util import load_config, load_embeddings_from_dir |
| 34 | 34 | ||
| 35 | 35 | ||
| @@ -245,7 +245,8 @@ def create_pipeline(model, dtype): | |||
| 245 | tokenizer=tokenizer, | 245 | tokenizer=tokenizer, |
| 246 | scheduler=scheduler, | 246 | scheduler=scheduler, |
| 247 | ) | 247 | ) |
| 248 | pipeline.enable_xformers_memory_efficient_attention() | 248 | # pipeline.enable_xformers_memory_efficient_attention() |
| 249 | pipeline.unet = torch.compile(pipeline.unet) | ||
| 249 | pipeline.enable_vae_slicing() | 250 | pipeline.enable_vae_slicing() |
| 250 | pipeline.to("cuda") | 251 | pipeline.to("cuda") |
| 251 | 252 | ||
diff --git a/schedulers/scheduling_unipc_multistep.py b/schedulers/scheduling_unipc_multistep.py deleted file mode 100644 index ff5db24..0000000 --- a/schedulers/scheduling_unipc_multistep.py +++ /dev/null | |||
| @@ -1,615 +0,0 @@ | |||
| 1 | # Copyright 2022 TSAIL 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/LuChengTHU/dpm-solver | ||
| 16 | |||
| 17 | import math | ||
| 18 | from typing import List, Optional, Union | ||
| 19 | |||
| 20 | import numpy as np | ||
| 21 | import torch | ||
| 22 | |||
| 23 | from diffusers.configuration_utils import ConfigMixin, register_to_config | ||
| 24 | from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput | ||
| 25 | |||
| 26 | |||
| 27 | def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999): | ||
| 28 | """ | ||
| 29 | Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of | ||
| 30 | (1-beta) over time from t = [0,1]. | ||
| 31 | |||
| 32 | Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up | ||
| 33 | to that part of the diffusion process. | ||
| 34 | |||
| 35 | |||
| 36 | Args: | ||
| 37 | num_diffusion_timesteps (`int`): the number of betas to produce. | ||
| 38 | max_beta (`float`): the maximum beta to use; use values lower than 1 to | ||
| 39 | prevent singularities. | ||
| 40 | |||
| 41 | Returns: | ||
| 42 | betas (`np.ndarray`): the betas used by the scheduler to step the model outputs | ||
| 43 | """ | ||
| 44 | |||
| 45 | def alpha_bar(time_step): | ||
| 46 | return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2 | ||
| 47 | |||
| 48 | betas = [] | ||
| 49 | for i in range(num_diffusion_timesteps): | ||
| 50 | t1 = i / num_diffusion_timesteps | ||
| 51 | t2 = (i + 1) / num_diffusion_timesteps | ||
| 52 | betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) | ||
| 53 | return torch.tensor(betas, dtype=torch.float32) | ||
| 54 | |||
| 55 | |||
| 56 | class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin): | ||
| 57 | """ | ||
| 58 | UniPC is a training-free framework designed for the fast sampling of diffusion models, which consists of | ||
| 59 | a corrector (UniC) and a predictor (UniP) that share a unified analytical form and support arbitrary orders. | ||
| 60 | UniPC is by desinged model-agnostic, supporting pixel-space/latent-space DPMs on unconditional/conditional | ||
| 61 | sampling. It can also be applied to both noise prediction model and data prediction model. The corrector | ||
| 62 | UniC can be also applied after any off-the-shelf solvers to increase the order of accuracy. | ||
| 63 | |||
| 64 | For more details, see the original paper: https://arxiv.org/abs/2302.04867 | ||
| 65 | |||
| 66 | Currently, we support the multistep UniPC for both noise prediction models and data prediction models. We | ||
| 67 | recommend to use `solver_order=2` for guided sampling, and `solver_order=3` for unconditional sampling. | ||
| 68 | |||
| 69 | We also support the "dynamic thresholding" method in Imagen (https://arxiv.org/abs/2205.11487). For pixel-space | ||
| 70 | diffusion models, you can set both `algorithm_type="dpmsolver++"` and `thresholding=True` to use the dynamic | ||
| 71 | thresholding. Note that the thresholding method is unsuitable for latent-space diffusion models (such as | ||
| 72 | stable-diffusion). | ||
| 73 | |||
| 74 | [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` | ||
| 75 | function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. | ||
| 76 | [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and | ||
| 77 | [`~SchedulerMixin.from_pretrained`] functions. | ||
| 78 | |||
| 79 | Args: | ||
| 80 | num_train_timesteps (`int`): number of diffusion steps used to train the model. | ||
| 81 | beta_start (`float`): the starting `beta` value of inference. | ||
| 82 | beta_end (`float`): the final `beta` value. | ||
| 83 | beta_schedule (`str`): | ||
| 84 | the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from | ||
| 85 | `linear`, `scaled_linear`, or `squaredcos_cap_v2`. | ||
| 86 | trained_betas (`np.ndarray`, optional): | ||
| 87 | option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. | ||
| 88 | solver_order (`int`, default `2`): | ||
| 89 | the order of UniPC, also the p in UniPC-p; can be any positive integer. Note that the effective order of | ||
| 90 | accuracy is `solver_order + 1` due to the UniC. We recommend to use `solver_order=2` for guided | ||
| 91 | sampling, and `solver_order=3` for unconditional sampling. | ||
| 92 | prediction_type (`str`, default `epsilon`, optional): | ||
| 93 | prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion | ||
| 94 | process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 | ||
| 95 | https://imagen.research.google/video/paper.pdf) | ||
| 96 | thresholding (`bool`, default `False`): | ||
| 97 | whether to use the "dynamic thresholding" method (introduced by Imagen, https://arxiv.org/abs/2205.11487). | ||
| 98 | For pixel-space diffusion models, you can set both `algorithm_type=dpmsolver++` and `thresholding=True` to | ||
| 99 | use the dynamic thresholding. Note that the thresholding method is unsuitable for latent-space diffusion | ||
| 100 | models (such as stable-diffusion). | ||
| 101 | dynamic_thresholding_ratio (`float`, default `0.995`): | ||
| 102 | the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen | ||
| 103 | (https://arxiv.org/abs/2205.11487). | ||
| 104 | sample_max_value (`float`, default `1.0`): | ||
| 105 | the threshold value for dynamic thresholding. Valid only when `thresholding=True` and | ||
| 106 | `predict_x0=True`. | ||
| 107 | predict_x0 (`bool`, default `True`): | ||
| 108 | whether to use the updating algrithm on the predicted x0. See https://arxiv.org/abs/2211.01095 for details | ||
| 109 | solver_type (`str`, default `bh1`): | ||
| 110 | the solver type of UniPC. We recommend use `bh1` for unconditional sampling when steps < 10, and use | ||
| 111 | `bh2` otherwise. | ||
| 112 | lower_order_final (`bool`, default `True`): | ||
| 113 | whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. We empirically | ||
| 114 | find this trick can stabilize the sampling of DPM-Solver for steps < 15, especially for steps <= 10. | ||
| 115 | disable_corrector (`list`, default `[]`): | ||
| 116 | decide which step to disable the corrector. For large guidance scale, the misalignment between the | ||
| 117 | `epsilon_theta(x_t, c)`and `epsilon_theta(x_t^c, c)` might influence the convergence. This can be | ||
| 118 | mitigated by disable the corrector at the first few steps (e.g., disable_corrector=[0]) | ||
| 119 | solver_p (`SchedulerMixin`): | ||
| 120 | can be any other scheduler. If specified, the algorithm will become solver_p + UniC. | ||
| 121 | """ | ||
| 122 | |||
| 123 | _compatibles = [e.name for e in KarrasDiffusionSchedulers] | ||
| 124 | order = 1 | ||
| 125 | |||
| 126 | @register_to_config | ||
| 127 | def __init__( | ||
| 128 | self, | ||
| 129 | num_train_timesteps: int = 1000, | ||
| 130 | beta_start: float = 0.0001, | ||
| 131 | beta_end: float = 0.02, | ||
| 132 | beta_schedule: str = "linear", | ||
| 133 | trained_betas: Optional[Union[np.ndarray, List[float]]] = None, | ||
| 134 | solver_order: int = 2, | ||
| 135 | prediction_type: str = "epsilon", | ||
| 136 | thresholding: bool = False, | ||
| 137 | dynamic_thresholding_ratio: float = 0.995, | ||
| 138 | sample_max_value: float = 1.0, | ||
| 139 | predict_x0: bool = True, | ||
| 140 | solver_type: str = "bh1", | ||
| 141 | lower_order_final: bool = True, | ||
| 142 | disable_corrector: List[int] = [], | ||
| 143 | solver_p: SchedulerMixin = None, | ||
| 144 | ): | ||
| 145 | if trained_betas is not None: | ||
| 146 | self.betas = torch.tensor(trained_betas, dtype=torch.float32) | ||
| 147 | elif beta_schedule == "linear": | ||
| 148 | self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) | ||
| 149 | elif beta_schedule == "scaled_linear": | ||
| 150 | # this schedule is very specific to the latent diffusion model. | ||
| 151 | self.betas = ( | ||
| 152 | torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 | ||
| 153 | ) | ||
| 154 | elif beta_schedule == "squaredcos_cap_v2": | ||
| 155 | # Glide cosine schedule | ||
| 156 | self.betas = betas_for_alpha_bar(num_train_timesteps) | ||
| 157 | else: | ||
| 158 | raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}") | ||
| 159 | |||
| 160 | self.alphas = 1.0 - self.betas | ||
| 161 | self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) | ||
| 162 | # Currently we only support VP-type noise schedule | ||
| 163 | self.alpha_t = torch.sqrt(self.alphas_cumprod) | ||
| 164 | self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) | ||
| 165 | self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) | ||
| 166 | |||
| 167 | # standard deviation of the initial noise distribution | ||
| 168 | self.init_noise_sigma = 1.0 | ||
| 169 | |||
| 170 | if solver_type not in ["bh1", "bh2"]: | ||
| 171 | raise NotImplementedError(f"{solver_type} does is not implemented for {self.__class__}") | ||
| 172 | |||
| 173 | self.predict_x0 = predict_x0 | ||
| 174 | # setable values | ||
| 175 | self.num_inference_steps = None | ||
| 176 | timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() | ||
| 177 | self.timesteps = torch.from_numpy(timesteps) | ||
| 178 | self.model_outputs = [None] * solver_order | ||
| 179 | self.timestep_list = [None] * solver_order | ||
| 180 | self.lower_order_nums = 0 | ||
| 181 | self.disable_corrector = disable_corrector | ||
| 182 | self.solver_p = solver_p | ||
| 183 | |||
| 184 | def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): | ||
| 185 | """ | ||
| 186 | Sets the timesteps used for the diffusion chain. Supporting function to be run before inference. | ||
| 187 | |||
| 188 | Args: | ||
| 189 | num_inference_steps (`int`): | ||
| 190 | the number of diffusion steps used when generating samples with a pre-trained model. | ||
| 191 | device (`str` or `torch.device`, optional): | ||
| 192 | the device to which the timesteps should be moved to. If `None`, the timesteps are not moved. | ||
| 193 | """ | ||
| 194 | self.num_inference_steps = num_inference_steps | ||
| 195 | timesteps = ( | ||
| 196 | np.linspace(0, self.num_train_timesteps - 1, num_inference_steps + 1) | ||
| 197 | .round()[::-1][:-1] | ||
| 198 | .copy() | ||
| 199 | .astype(np.int64) | ||
| 200 | ) | ||
| 201 | self.timesteps = torch.from_numpy(timesteps).to(device) | ||
| 202 | self.model_outputs = [ | ||
| 203 | None, | ||
| 204 | ] * self.config.solver_order | ||
| 205 | self.lower_order_nums = 0 | ||
| 206 | if self.solver_p: | ||
| 207 | self.solver_p.set_timesteps(num_inference_steps, device=device) | ||
| 208 | |||
| 209 | def convert_model_output( | ||
| 210 | self, model_output: torch.FloatTensor, timestep: int, sample: torch.FloatTensor | ||
| 211 | ): | ||
| 212 | r""" | ||
| 213 | Convert the model output to the corresponding type that the algorithm PC needs. | ||
| 214 | |||
| 215 | Args: | ||
| 216 | model_output (`torch.FloatTensor`): direct output from learned diffusion model. | ||
| 217 | timestep (`int`): current discrete timestep in the diffusion chain. | ||
| 218 | sample (`torch.FloatTensor`): | ||
| 219 | current instance of sample being created by diffusion process. | ||
| 220 | |||
| 221 | Returns: | ||
| 222 | `torch.FloatTensor`: the converted model output. | ||
| 223 | """ | ||
| 224 | if self.predict_x0: | ||
| 225 | if self.config.prediction_type == "epsilon": | ||
| 226 | alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep] | ||
| 227 | x0_pred = (sample - sigma_t * model_output) / alpha_t | ||
| 228 | elif self.config.prediction_type == "sample": | ||
| 229 | x0_pred = model_output | ||
| 230 | elif self.config.prediction_type == "v_prediction": | ||
| 231 | alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep] | ||
| 232 | x0_pred = alpha_t * sample - sigma_t * model_output | ||
| 233 | else: | ||
| 234 | raise ValueError( | ||
| 235 | f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" | ||
| 236 | " `v_prediction` for the DPMSolverMultistepScheduler." | ||
| 237 | ) | ||
| 238 | |||
| 239 | if self.config.thresholding: | ||
| 240 | # Dynamic thresholding in https://arxiv.org/abs/2205.11487 | ||
| 241 | orig_dtype = x0_pred.dtype | ||
| 242 | if orig_dtype not in [torch.float, torch.double]: | ||
| 243 | x0_pred = x0_pred.float() | ||
| 244 | dynamic_max_val = torch.quantile( | ||
| 245 | torch.abs(x0_pred).reshape((x0_pred.shape[0], -1)), self.config.dynamic_thresholding_ratio, dim=1 | ||
| 246 | ) | ||
| 247 | dynamic_max_val = torch.maximum( | ||
| 248 | dynamic_max_val, | ||
| 249 | self.config.sample_max_value * torch.ones_like(dynamic_max_val).to(dynamic_max_val.device), | ||
| 250 | )[(...,) + (None,) * (x0_pred.ndim - 1)] | ||
| 251 | x0_pred = torch.clamp(x0_pred, -dynamic_max_val, dynamic_max_val) / dynamic_max_val | ||
| 252 | x0_pred = x0_pred.type(orig_dtype) | ||
| 253 | return x0_pred | ||
| 254 | else: | ||
| 255 | if self.config.prediction_type == "epsilon": | ||
| 256 | return model_output | ||
| 257 | elif self.config.prediction_type == "sample": | ||
| 258 | alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep] | ||
| 259 | epsilon = (sample - alpha_t * model_output) / sigma_t | ||
| 260 | return epsilon | ||
| 261 | elif self.config.prediction_type == "v_prediction": | ||
| 262 | alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep] | ||
| 263 | epsilon = alpha_t * model_output + sigma_t * sample | ||
| 264 | return epsilon | ||
| 265 | else: | ||
| 266 | raise ValueError( | ||
| 267 | f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" | ||
| 268 | " `v_prediction` for the DPMSolverMultistepScheduler." | ||
| 269 | ) | ||
| 270 | |||
| 271 | def multistep_uni_p_bh_update( | ||
| 272 | self, | ||
| 273 | model_output: torch.FloatTensor, | ||
| 274 | prev_timestep: int, | ||
| 275 | sample: torch.FloatTensor, | ||
| 276 | order: int, | ||
| 277 | ): | ||
| 278 | """ | ||
| 279 | One step for the UniP (B(h) version). Alternatively, `self.solver_p` is used if is specified. | ||
| 280 | |||
| 281 | Args: | ||
| 282 | model_output (`torch.FloatTensor`): | ||
| 283 | direct outputs from learned diffusion model at the current timestep. | ||
| 284 | prev_timestep (`int`): previous discrete timestep in the diffusion chain. | ||
| 285 | sample (`torch.FloatTensor`): | ||
| 286 | current instance of sample being created by diffusion process. | ||
| 287 | order (`int`): the order of UniP at this step, also the p in UniPC-p. | ||
| 288 | |||
| 289 | Returns: | ||
| 290 | `torch.FloatTensor`: the sample tensor at the previous timestep. | ||
| 291 | """ | ||
| 292 | timestep_list = self.timestep_list | ||
| 293 | model_output_list = self.model_outputs | ||
| 294 | |||
| 295 | s0, t = self.timestep_list[-1], prev_timestep | ||
| 296 | m0 = model_output_list[-1] | ||
| 297 | x = sample | ||
| 298 | |||
| 299 | if self.solver_p: | ||
| 300 | x_t = self.solver_p.step(model_output, s0, x).prev_sample | ||
| 301 | return x_t | ||
| 302 | |||
| 303 | lambda_t, lambda_s0 = self.lambda_t[t], self.lambda_t[s0] | ||
| 304 | alpha_t, alpha_s0 = self.alpha_t[t], self.alpha_t[s0] | ||
| 305 | sigma_t, sigma_s0 = self.sigma_t[t], self.sigma_t[s0] | ||
| 306 | |||
| 307 | h = lambda_t - lambda_s0 | ||
| 308 | device = sample.device | ||
| 309 | |||
| 310 | rks = [] | ||
| 311 | D1s = [] | ||
| 312 | for i in range(1, order): | ||
| 313 | si = timestep_list[-(i + 1)] | ||
| 314 | mi = model_output_list[-(i + 1)] | ||
| 315 | lambda_si = self.lambda_t[si] | ||
| 316 | rk = ((lambda_si - lambda_s0) / h) | ||
| 317 | rks.append(rk) | ||
| 318 | D1s.append((mi - m0) / rk) | ||
| 319 | |||
| 320 | rks.append(1.) | ||
| 321 | rks = torch.tensor(rks, device=device) | ||
| 322 | |||
| 323 | R = [] | ||
| 324 | b = [] | ||
| 325 | |||
| 326 | hh = -h if self.predict_x0 else h | ||
| 327 | h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1 | ||
| 328 | h_phi_k = h_phi_1 / hh - 1 | ||
| 329 | |||
| 330 | factorial_i = 1 | ||
| 331 | |||
| 332 | if self.config.solver_type == 'bh1': | ||
| 333 | B_h = hh | ||
| 334 | elif self.config.solver_type == 'bh2': | ||
| 335 | B_h = torch.expm1(hh) | ||
| 336 | else: | ||
| 337 | raise NotImplementedError() | ||
| 338 | |||
| 339 | for i in range(1, order + 1): | ||
| 340 | R.append(torch.pow(rks, i - 1)) | ||
| 341 | b.append(h_phi_k * factorial_i / B_h) | ||
| 342 | factorial_i *= (i + 1) | ||
| 343 | h_phi_k = h_phi_k / hh - 1 / factorial_i | ||
| 344 | |||
| 345 | R = torch.stack(R) | ||
| 346 | b = torch.tensor(b, device=device) | ||
| 347 | |||
| 348 | if len(D1s) > 0: | ||
| 349 | D1s = torch.stack(D1s, dim=1) # (B, K) | ||
| 350 | # for order 2, we use a simplified version | ||
| 351 | if order == 2: | ||
| 352 | rhos_p = torch.tensor([0.5], dtype=x.dtype, device=device) | ||
| 353 | else: | ||
| 354 | rhos_p = torch.linalg.solve(R[:-1, :-1], b[:-1]) | ||
| 355 | else: | ||
| 356 | D1s = None | ||
| 357 | |||
| 358 | if self.predict_x0: | ||
| 359 | x_t_ = ( | ||
| 360 | sigma_t / sigma_s0 * x | ||
| 361 | - alpha_t * h_phi_1 * m0 | ||
| 362 | ) | ||
| 363 | if D1s is not None: | ||
| 364 | pred_res = torch.einsum('k,bkchw->bchw', rhos_p, D1s) | ||
| 365 | else: | ||
| 366 | pred_res = 0 | ||
| 367 | x_t = x_t_ - alpha_t * B_h * pred_res | ||
| 368 | else: | ||
| 369 | x_t_ = ( | ||
| 370 | alpha_t / alpha_s0 * x | ||
| 371 | - sigma_t * h_phi_1 * m0 | ||
| 372 | ) | ||
| 373 | if D1s is not None: | ||
| 374 | pred_res = torch.einsum('k,bkchw->bchw', rhos_p, D1s) | ||
| 375 | else: | ||
| 376 | pred_res = 0 | ||
| 377 | x_t = x_t_ - sigma_t * B_h * pred_res | ||
| 378 | |||
| 379 | x_t = x_t.to(x.dtype) | ||
| 380 | return x_t | ||
| 381 | |||
| 382 | def multistep_uni_c_bh_update( | ||
| 383 | self, | ||
| 384 | this_model_output: torch.FloatTensor, | ||
| 385 | this_timestep: int, | ||
| 386 | last_sample: torch.FloatTensor, | ||
| 387 | this_sample: torch.FloatTensor, | ||
| 388 | order: int, | ||
| 389 | ): | ||
| 390 | """ | ||
| 391 | One step for the UniC (B(h) version). | ||
| 392 | |||
| 393 | Args: | ||
| 394 | this_model_output (`torch.FloatTensor`): the model outputs at `x_t` | ||
| 395 | this_timestep (`int`): the current timestep `t` | ||
| 396 | last_sample (`torch.FloatTensor`): the generated sample before the last predictor: `x_{t-1}` | ||
| 397 | this_sample (`torch.FloatTensor`): the generated sample after the last predictor: `x_{t}` | ||
| 398 | order (`int`): the `p` of UniC-p at this step. Note that the effective order of accuracy | ||
| 399 | should be order + 1 | ||
| 400 | |||
| 401 | Returns: | ||
| 402 | `torch.FloatTensor`: the corrected sample tensor at the current timestep. | ||
| 403 | """ | ||
| 404 | timestep_list = self.timestep_list | ||
| 405 | model_output_list = self.model_outputs | ||
| 406 | |||
| 407 | s0, t = timestep_list[-1], this_timestep | ||
| 408 | m0 = model_output_list[-1] | ||
| 409 | x = last_sample | ||
| 410 | x_t = this_sample | ||
| 411 | model_t = this_model_output | ||
| 412 | |||
| 413 | lambda_t, lambda_s0 = self.lambda_t[t], self.lambda_t[s0] | ||
| 414 | alpha_t, alpha_s0 = self.alpha_t[t], self.alpha_t[s0] | ||
| 415 | sigma_t, sigma_s0 = self.sigma_t[t], self.sigma_t[s0] | ||
| 416 | |||
| 417 | h = lambda_t - lambda_s0 | ||
| 418 | device = this_sample.device | ||
| 419 | |||
| 420 | rks = [] | ||
| 421 | D1s = [] | ||
| 422 | for i in range(1, order): | ||
| 423 | si = timestep_list[-(i + 1)] | ||
| 424 | mi = model_output_list[-(i + 1)] | ||
| 425 | lambda_si = self.lambda_t[si] | ||
| 426 | rk = ((lambda_si - lambda_s0) / h) | ||
| 427 | rks.append(rk) | ||
| 428 | D1s.append((mi - m0) / rk) | ||
| 429 | |||
| 430 | rks.append(1.) | ||
| 431 | rks = torch.tensor(rks, device=device) | ||
| 432 | |||
| 433 | R = [] | ||
| 434 | b = [] | ||
| 435 | |||
| 436 | hh = -h if self.predict_x0 else h | ||
| 437 | h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1 | ||
| 438 | h_phi_k = h_phi_1 / hh - 1 | ||
| 439 | |||
| 440 | factorial_i = 1 | ||
| 441 | |||
| 442 | if self.config.solver_type == 'bh1': | ||
| 443 | B_h = hh | ||
| 444 | elif self.config.solver_type == 'bh2': | ||
| 445 | B_h = torch.expm1(hh) | ||
| 446 | else: | ||
| 447 | raise NotImplementedError() | ||
| 448 | |||
| 449 | for i in range(1, order + 1): | ||
| 450 | R.append(torch.pow(rks, i - 1)) | ||
| 451 | b.append(h_phi_k * factorial_i / B_h) | ||
| 452 | factorial_i *= (i + 1) | ||
| 453 | h_phi_k = h_phi_k / hh - 1 / factorial_i | ||
| 454 | |||
| 455 | R = torch.stack(R) | ||
| 456 | b = torch.tensor(b, device=device) | ||
| 457 | |||
| 458 | if len(D1s) > 0: | ||
| 459 | D1s = torch.stack(D1s, dim=1) | ||
| 460 | else: | ||
| 461 | D1s = None | ||
| 462 | |||
| 463 | # for order 1, we use a simplified version | ||
| 464 | if order == 1: | ||
| 465 | rhos_c = torch.tensor([0.5], dtype=x.dtype, device=device) | ||
| 466 | else: | ||
| 467 | rhos_c = torch.linalg.solve(R, b) | ||
| 468 | |||
| 469 | if self.predict_x0: | ||
| 470 | x_t_ = ( | ||
| 471 | sigma_t / sigma_s0 * x | ||
| 472 | - alpha_t * h_phi_1 * m0 | ||
| 473 | ) | ||
| 474 | if D1s is not None: | ||
| 475 | corr_res = torch.einsum('k,bkchw->bchw', rhos_c[:-1], D1s) | ||
| 476 | else: | ||
| 477 | corr_res = 0 | ||
| 478 | D1_t = (model_t - m0) | ||
| 479 | x_t = x_t_ - alpha_t * B_h * (corr_res + rhos_c[-1] * D1_t) | ||
| 480 | else: | ||
| 481 | x_t_ = ( | ||
| 482 | alpha_t / alpha_s0 * x | ||
| 483 | - sigma_t * h_phi_1 * m0 | ||
| 484 | ) | ||
| 485 | if D1s is not None: | ||
| 486 | corr_res = torch.einsum('k,bkchw->bchw', rhos_c[:-1], D1s) | ||
| 487 | else: | ||
| 488 | corr_res = 0 | ||
| 489 | D1_t = (model_t - m0) | ||
| 490 | x_t = x_t_ - sigma_t * B_h * (corr_res + rhos_c[-1] * D1_t) | ||
| 491 | x_t = x_t.to(x.dtype) | ||
| 492 | return x_t | ||
| 493 | |||
| 494 | def step( | ||
| 495 | self, | ||
| 496 | model_output: torch.FloatTensor, | ||
| 497 | timestep: int, | ||
| 498 | sample: torch.FloatTensor, | ||
| 499 | return_dict: bool = True, | ||
| 500 | ): | ||
| 501 | # -> Union[SchedulerOutput, Tuple]: | ||
| 502 | """ | ||
| 503 | Step function propagating the sample with the multistep UniPC. | ||
| 504 | |||
| 505 | Args: | ||
| 506 | model_output (`torch.FloatTensor`): direct output from learned diffusion model. | ||
| 507 | timestep (`int`): current discrete timestep in the diffusion chain. | ||
| 508 | sample (`torch.FloatTensor`): | ||
| 509 | current instance of sample being created by diffusion process. | ||
| 510 | return_dict (`bool`): option for returning tuple rather than SchedulerOutput class | ||
| 511 | |||
| 512 | Returns: | ||
| 513 | [`~scheduling_utils.SchedulerOutput`] or `tuple`: [`~scheduling_utils.SchedulerOutput`] if `return_dict` is | ||
| 514 | True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. | ||
| 515 | |||
| 516 | """ | ||
| 517 | |||
| 518 | if self.num_inference_steps is None: | ||
| 519 | raise ValueError( | ||
| 520 | "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" | ||
| 521 | ) | ||
| 522 | |||
| 523 | if isinstance(timestep, torch.Tensor): | ||
| 524 | timestep = timestep.to(self.timesteps.device) | ||
| 525 | step_index = (self.timesteps == timestep).nonzero() | ||
| 526 | if len(step_index) == 0: | ||
| 527 | step_index = len(self.timesteps) - 1 | ||
| 528 | else: | ||
| 529 | step_index = step_index.item() | ||
| 530 | |||
| 531 | use_corrector = step_index > 0 and step_index - 1 not in self.disable_corrector # step_index not in self.disable_corrector | ||
| 532 | |||
| 533 | model_output_convert = self.convert_model_output(model_output, timestep, sample) | ||
| 534 | if use_corrector: | ||
| 535 | sample = self.multistep_uni_c_bh_update( | ||
| 536 | this_model_output=model_output_convert, | ||
| 537 | this_timestep=timestep, | ||
| 538 | last_sample=self.last_sample, | ||
| 539 | this_sample=sample, | ||
| 540 | order=self.this_order, | ||
| 541 | ) | ||
| 542 | |||
| 543 | # now prepare to run the predictor | ||
| 544 | prev_timestep = 0 if step_index == len(self.timesteps) - 1 else self.timesteps[step_index + 1] | ||
| 545 | |||
| 546 | for i in range(self.config.solver_order - 1): | ||
| 547 | self.model_outputs[i] = self.model_outputs[i + 1] | ||
| 548 | self.timestep_list[i] = self.timestep_list[i + 1] | ||
| 549 | |||
| 550 | self.model_outputs[-1] = model_output_convert | ||
| 551 | self.timestep_list[-1] = timestep | ||
| 552 | |||
| 553 | if self.config.lower_order_final: | ||
| 554 | this_order = min(self.config.solver_order, len(self.timesteps) - step_index) | ||
| 555 | else: | ||
| 556 | this_order = self.config.solver_order | ||
| 557 | |||
| 558 | self.this_order = min(this_order, self.lower_order_nums + 1) # warmup for multistep | ||
| 559 | assert self.this_order > 0 | ||
| 560 | |||
| 561 | self.last_sample = sample | ||
| 562 | prev_sample = self.multistep_uni_p_bh_update( | ||
| 563 | model_output=model_output, # pass the original non-converted model output, in case solver-p is used | ||
| 564 | prev_timestep=prev_timestep, | ||
| 565 | sample=sample, | ||
| 566 | order=self.this_order, | ||
| 567 | ) | ||
| 568 | |||
| 569 | if self.lower_order_nums < self.config.solver_order: | ||
| 570 | self.lower_order_nums += 1 | ||
| 571 | |||
| 572 | if not return_dict: | ||
| 573 | return (prev_sample,) | ||
| 574 | |||
| 575 | return SchedulerOutput(prev_sample=prev_sample) | ||
| 576 | |||
| 577 | def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs): # -> torch.FloatTensor: | ||
| 578 | """ | ||
| 579 | Ensures interchangeability with schedulers that need to scale the denoising model input depending on the | ||
| 580 | current timestep. | ||
| 581 | |||
| 582 | Args: | ||
| 583 | sample (`torch.FloatTensor`): input sample | ||
| 584 | |||
| 585 | Returns: | ||
| 586 | `torch.FloatTensor`: scaled input sample | ||
| 587 | """ | ||
| 588 | return sample | ||
| 589 | |||
| 590 | def add_noise( | ||
| 591 | self, | ||
| 592 | original_samples: torch.FloatTensor, | ||
| 593 | noise: torch.FloatTensor, | ||
| 594 | timesteps: torch.IntTensor, | ||
| 595 | ): | ||
| 596 | # -> torch.FloatTensor: | ||
| 597 | # Make sure alphas_cumprod and timestep have same device and dtype as original_samples | ||
| 598 | self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype) | ||
| 599 | timesteps = timesteps.to(original_samples.device) | ||
| 600 | |||
| 601 | sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5 | ||
| 602 | sqrt_alpha_prod = sqrt_alpha_prod.flatten() | ||
| 603 | while len(sqrt_alpha_prod.shape) < len(original_samples.shape): | ||
| 604 | sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) | ||
| 605 | |||
| 606 | sqrt_one_minus_alpha_prod = (1 - self.alphas_cumprod[timesteps]) ** 0.5 | ||
| 607 | sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() | ||
| 608 | while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): | ||
| 609 | sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) | ||
| 610 | |||
| 611 | noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise | ||
| 612 | return noisy_samples | ||
| 613 | |||
| 614 | def __len__(self): | ||
| 615 | return self.config.num_train_timesteps | ||
diff --git a/train_dreambooth.py b/train_dreambooth.py index 5a7911c..85b756c 100644 --- a/train_dreambooth.py +++ b/train_dreambooth.py | |||
| @@ -464,8 +464,8 @@ def main(): | |||
| 464 | tokenizer.set_dropout(args.vector_dropout) | 464 | tokenizer.set_dropout(args.vector_dropout) |
| 465 | 465 | ||
| 466 | vae.enable_slicing() | 466 | vae.enable_slicing() |
| 467 | vae.set_use_memory_efficient_attention_xformers(True) | 467 | # vae.set_use_memory_efficient_attention_xformers(True) |
| 468 | unet.enable_xformers_memory_efficient_attention() | 468 | # unet.enable_xformers_memory_efficient_attention() |
| 469 | 469 | ||
| 470 | if args.gradient_checkpointing: | 470 | if args.gradient_checkpointing: |
| 471 | unet.enable_gradient_checkpointing() | 471 | unet.enable_gradient_checkpointing() |
diff --git a/train_lora.py b/train_lora.py index 330bcd6..8a06ae8 100644 --- a/train_lora.py +++ b/train_lora.py | |||
| @@ -421,8 +421,8 @@ def main(): | |||
| 421 | args.pretrained_model_name_or_path) | 421 | args.pretrained_model_name_or_path) |
| 422 | 422 | ||
| 423 | vae.enable_slicing() | 423 | vae.enable_slicing() |
| 424 | vae.set_use_memory_efficient_attention_xformers(True) | 424 | # vae.set_use_memory_efficient_attention_xformers(True) |
| 425 | unet.enable_xformers_memory_efficient_attention() | 425 | # unet.enable_xformers_memory_efficient_attention() |
| 426 | 426 | ||
| 427 | unet.to(accelerator.device, dtype=weight_dtype) | 427 | unet.to(accelerator.device, dtype=weight_dtype) |
| 428 | text_encoder.to(accelerator.device, dtype=weight_dtype) | 428 | text_encoder.to(accelerator.device, dtype=weight_dtype) |
diff --git a/train_ti.py b/train_ti.py index d1defb3..7d10317 100644 --- a/train_ti.py +++ b/train_ti.py | |||
| @@ -538,8 +538,10 @@ def main(): | |||
| 538 | tokenizer.set_dropout(args.vector_dropout) | 538 | tokenizer.set_dropout(args.vector_dropout) |
| 539 | 539 | ||
| 540 | vae.enable_slicing() | 540 | vae.enable_slicing() |
| 541 | vae.set_use_memory_efficient_attention_xformers(True) | 541 | # vae.set_use_memory_efficient_attention_xformers(True) |
| 542 | unet.enable_xformers_memory_efficient_attention() | 542 | # unet.enable_xformers_memory_efficient_attention() |
| 543 | |||
| 544 | # unet = torch.compile(unet) | ||
| 543 | 545 | ||
| 544 | if args.gradient_checkpointing: | 546 | if args.gradient_checkpointing: |
| 545 | unet.enable_gradient_checkpointing() | 547 | unet.enable_gradient_checkpointing() |
diff --git a/training/functional.py b/training/functional.py index 78a2b10..41794ea 100644 --- a/training/functional.py +++ b/training/functional.py | |||
| @@ -12,7 +12,7 @@ from torch.utils.data import DataLoader | |||
| 12 | 12 | ||
| 13 | from accelerate import Accelerator | 13 | from accelerate import Accelerator |
| 14 | from transformers import CLIPTextModel | 14 | from transformers import CLIPTextModel |
| 15 | from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel | 15 | from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel, UniPCMultistepScheduler |
| 16 | 16 | ||
| 17 | from tqdm.auto import tqdm | 17 | from tqdm.auto import tqdm |
| 18 | from PIL import Image | 18 | from PIL import Image |
| @@ -22,7 +22,6 @@ from pipelines.stable_diffusion.vlpn_stable_diffusion import VlpnStableDiffusion | |||
| 22 | from models.clip.embeddings import ManagedCLIPTextEmbeddings, patch_managed_embeddings | 22 | from models.clip.embeddings import ManagedCLIPTextEmbeddings, patch_managed_embeddings |
| 23 | from models.clip.util import get_extended_embeddings | 23 | from models.clip.util import get_extended_embeddings |
| 24 | from models.clip.tokenizer import MultiCLIPTokenizer | 24 | from models.clip.tokenizer import MultiCLIPTokenizer |
| 25 | from schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler | ||
| 26 | from training.util import AverageMeter | 25 | from training.util import AverageMeter |
| 27 | 26 | ||
| 28 | 27 | ||
diff --git a/training/strategy/dreambooth.py b/training/strategy/dreambooth.py index 8aaed3a..d697554 100644 --- a/training/strategy/dreambooth.py +++ b/training/strategy/dreambooth.py | |||
| @@ -144,8 +144,8 @@ def dreambooth_strategy_callbacks( | |||
| 144 | 144 | ||
| 145 | print("Saving model...") | 145 | print("Saving model...") |
| 146 | 146 | ||
| 147 | unet_ = accelerator.unwrap_model(unet) | 147 | unet_ = accelerator.unwrap_model(unet, False) |
| 148 | text_encoder_ = accelerator.unwrap_model(text_encoder) | 148 | text_encoder_ = accelerator.unwrap_model(text_encoder, False) |
| 149 | 149 | ||
| 150 | with ema_context(): | 150 | with ema_context(): |
| 151 | pipeline = VlpnStableDiffusion( | 151 | pipeline = VlpnStableDiffusion( |
| @@ -167,8 +167,8 @@ def dreambooth_strategy_callbacks( | |||
| 167 | @torch.no_grad() | 167 | @torch.no_grad() |
| 168 | def on_sample(step): | 168 | def on_sample(step): |
| 169 | with ema_context(): | 169 | with ema_context(): |
| 170 | unet_ = accelerator.unwrap_model(unet) | 170 | unet_ = accelerator.unwrap_model(unet, False) |
| 171 | text_encoder_ = accelerator.unwrap_model(text_encoder) | 171 | text_encoder_ = accelerator.unwrap_model(text_encoder, False) |
| 172 | 172 | ||
| 173 | orig_unet_dtype = unet_.dtype | 173 | orig_unet_dtype = unet_.dtype |
| 174 | orig_text_encoder_dtype = text_encoder_.dtype | 174 | orig_text_encoder_dtype = text_encoder_.dtype |
diff --git a/training/strategy/lora.py b/training/strategy/lora.py index 4dd1100..ccec215 100644 --- a/training/strategy/lora.py +++ b/training/strategy/lora.py | |||
| @@ -90,7 +90,7 @@ def lora_strategy_callbacks( | |||
| 90 | def on_checkpoint(step, postfix): | 90 | def on_checkpoint(step, postfix): |
| 91 | print(f"Saving checkpoint for step {step}...") | 91 | print(f"Saving checkpoint for step {step}...") |
| 92 | 92 | ||
| 93 | unet_ = accelerator.unwrap_model(unet) | 93 | unet_ = accelerator.unwrap_model(unet, False) |
| 94 | unet_.save_attn_procs(checkpoint_output_dir / f"{step}_{postfix}") | 94 | unet_.save_attn_procs(checkpoint_output_dir / f"{step}_{postfix}") |
| 95 | del unet_ | 95 | del unet_ |
| 96 | 96 | ||
diff --git a/training/strategy/ti.py b/training/strategy/ti.py index 0de3cb0..66d3129 100644 --- a/training/strategy/ti.py +++ b/training/strategy/ti.py | |||
| @@ -144,8 +144,8 @@ def textual_inversion_strategy_callbacks( | |||
| 144 | @torch.no_grad() | 144 | @torch.no_grad() |
| 145 | def on_sample(step): | 145 | def on_sample(step): |
| 146 | with ema_context(): | 146 | with ema_context(): |
| 147 | unet_ = accelerator.unwrap_model(unet) | 147 | unet_ = accelerator.unwrap_model(unet, False) |
| 148 | text_encoder_ = accelerator.unwrap_model(text_encoder) | 148 | text_encoder_ = accelerator.unwrap_model(text_encoder, False) |
| 149 | 149 | ||
| 150 | orig_unet_dtype = unet_.dtype | 150 | orig_unet_dtype = unet_.dtype |
| 151 | orig_text_encoder_dtype = text_encoder_.dtype | 151 | orig_text_encoder_dtype = text_encoder_.dtype |
