Spaces:
Running
on
Zero
Running
on
Zero
add func of get_torch_dtype
Browse files- sampler_invsr.py +10 -2
- trainer.py +0 -1643
sampler_invsr.py
CHANGED
|
@@ -10,8 +10,6 @@ from pathlib import Path
|
|
| 10 |
from loguru import logger
|
| 11 |
from omegaconf import OmegaConf
|
| 12 |
|
| 13 |
-
from trainer import get_torch_dtype
|
| 14 |
-
|
| 15 |
from utils import util_net
|
| 16 |
from utils import util_image
|
| 17 |
from utils import util_common
|
|
@@ -30,6 +28,16 @@ _positive= 'Cinematic, high-contrast, photo-realistic, 8k, ultra HD, ' +\
|
|
| 30 |
_negative= 'Low quality, blurring, jpeg artifacts, deformed, over-smooth, cartoon, noisy,' +\
|
| 31 |
'painting, drawing, sketch, oil painting'
|
| 32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
class BaseSampler:
|
| 34 |
def __init__(self, configs):
|
| 35 |
'''
|
|
|
|
| 10 |
from loguru import logger
|
| 11 |
from omegaconf import OmegaConf
|
| 12 |
|
|
|
|
|
|
|
| 13 |
from utils import util_net
|
| 14 |
from utils import util_image
|
| 15 |
from utils import util_common
|
|
|
|
| 28 |
_negative= 'Low quality, blurring, jpeg artifacts, deformed, over-smooth, cartoon, noisy,' +\
|
| 29 |
'painting, drawing, sketch, oil painting'
|
| 30 |
|
| 31 |
+
def get_torch_dtype(torch_dtype: str):
|
| 32 |
+
if torch_dtype == 'torch.float16':
|
| 33 |
+
return torch.float16
|
| 34 |
+
elif torch_dtype == 'torch.bfloat16':
|
| 35 |
+
return torch.bfloat16
|
| 36 |
+
elif torch_dtype == 'torch.float32':
|
| 37 |
+
return torch.float32
|
| 38 |
+
else:
|
| 39 |
+
raise ValueError(f'Unexpected torch dtype:{torch_dtype}')
|
| 40 |
+
|
| 41 |
class BaseSampler:
|
| 42 |
def __init__(self, configs):
|
| 43 |
'''
|
trainer.py
DELETED
|
@@ -1,1643 +0,0 @@
|
|
| 1 |
-
#!/usr/bin/env python
|
| 2 |
-
# -*- coding:utf-8 -*-
|
| 3 |
-
# Power by Zongsheng Yue 2022-05-18 13:04:06
|
| 4 |
-
|
| 5 |
-
import os, sys, math, time, random, datetime
|
| 6 |
-
import numpy as np
|
| 7 |
-
from box import Box
|
| 8 |
-
from pathlib import Path
|
| 9 |
-
from loguru import logger
|
| 10 |
-
from copy import deepcopy
|
| 11 |
-
from omegaconf import OmegaConf
|
| 12 |
-
from einops import rearrange
|
| 13 |
-
from typing import Any, Dict, List, Optional, Tuple, Union
|
| 14 |
-
|
| 15 |
-
from datapipe.datasets import create_dataset
|
| 16 |
-
|
| 17 |
-
import torch
|
| 18 |
-
import torch.nn as nn
|
| 19 |
-
import torch.nn.functional as F
|
| 20 |
-
import torch.utils.data as udata
|
| 21 |
-
import torch.distributed as dist
|
| 22 |
-
import torch.multiprocessing as mp
|
| 23 |
-
import torchvision.utils as vutils
|
| 24 |
-
from torch.nn.parallel import DistributedDataParallel as DDP
|
| 25 |
-
|
| 26 |
-
from utils import util_net
|
| 27 |
-
from utils import util_common
|
| 28 |
-
from utils import util_image
|
| 29 |
-
from utils.util_ops import append_dims
|
| 30 |
-
|
| 31 |
-
import pyiqa
|
| 32 |
-
from basicsr.utils import DiffJPEG, USMSharp
|
| 33 |
-
from basicsr.utils.img_process_util import filter2D
|
| 34 |
-
from basicsr.data.transforms import paired_random_crop
|
| 35 |
-
from basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt
|
| 36 |
-
|
| 37 |
-
from diffusers import EulerDiscreteScheduler
|
| 38 |
-
from diffusers.models.autoencoders.vae import DiagonalGaussianDistribution
|
| 39 |
-
from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img import retrieve_timesteps
|
| 40 |
-
|
| 41 |
-
_base_seed = 10**6
|
| 42 |
-
_INTERPOLATION_MODE = 'bicubic'
|
| 43 |
-
_Latent_bound = {'min':-10.0, 'max':10.0}
|
| 44 |
-
_positive= 'Cinematic, high-contrast, photo-realistic, 8k, ultra HD, ' +\
|
| 45 |
-
'meticulous detailing, hyper sharpness, perfect without deformations'
|
| 46 |
-
_negative= 'Low quality, blurring, jpeg artifacts, deformed, over-smooth, cartoon, noisy,' +\
|
| 47 |
-
'painting, drawing, sketch, oil painting'
|
| 48 |
-
|
| 49 |
-
class TrainerBase:
|
| 50 |
-
def __init__(self, configs):
|
| 51 |
-
self.configs = configs
|
| 52 |
-
|
| 53 |
-
# setup distributed training: self.num_gpus, self.rank
|
| 54 |
-
self.setup_dist()
|
| 55 |
-
|
| 56 |
-
# setup seed
|
| 57 |
-
self.setup_seed()
|
| 58 |
-
|
| 59 |
-
def setup_dist(self):
|
| 60 |
-
num_gpus = torch.cuda.device_count()
|
| 61 |
-
|
| 62 |
-
if num_gpus > 1:
|
| 63 |
-
if mp.get_start_method(allow_none=True) is None:
|
| 64 |
-
mp.set_start_method('spawn')
|
| 65 |
-
rank = int(os.environ['LOCAL_RANK'])
|
| 66 |
-
torch.cuda.set_device(rank % num_gpus)
|
| 67 |
-
dist.init_process_group(
|
| 68 |
-
timeout=datetime.timedelta(seconds=3600),
|
| 69 |
-
backend='nccl',
|
| 70 |
-
init_method='env://',
|
| 71 |
-
)
|
| 72 |
-
|
| 73 |
-
self.num_gpus = num_gpus
|
| 74 |
-
self.rank = int(os.environ['LOCAL_RANK']) if num_gpus > 1 else 0
|
| 75 |
-
|
| 76 |
-
def setup_seed(self, seed=None, global_seeding=None):
|
| 77 |
-
if seed is None:
|
| 78 |
-
seed = self.configs.train.get('seed', 12345)
|
| 79 |
-
if global_seeding is None:
|
| 80 |
-
global_seeding = self.configs.train.get('global_seeding', False)
|
| 81 |
-
if not global_seeding:
|
| 82 |
-
seed += self.rank
|
| 83 |
-
torch.cuda.manual_seed(seed)
|
| 84 |
-
else:
|
| 85 |
-
torch.cuda.manual_seed_all(seed)
|
| 86 |
-
random.seed(seed)
|
| 87 |
-
np.random.seed(seed)
|
| 88 |
-
torch.manual_seed(seed)
|
| 89 |
-
|
| 90 |
-
def init_logger(self):
|
| 91 |
-
if self.configs.resume:
|
| 92 |
-
assert self.configs.resume.endswith(".pth")
|
| 93 |
-
save_dir = Path(self.configs.resume).parents[1]
|
| 94 |
-
project_id = save_dir.name
|
| 95 |
-
else:
|
| 96 |
-
project_id = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M")
|
| 97 |
-
save_dir = Path(self.configs.save_dir) / project_id
|
| 98 |
-
if not save_dir.exists() and self.rank == 0:
|
| 99 |
-
save_dir.mkdir(parents=True)
|
| 100 |
-
|
| 101 |
-
# setting log counter
|
| 102 |
-
if self.rank == 0:
|
| 103 |
-
self.log_step = {phase: 1 for phase in ['train', 'val']}
|
| 104 |
-
self.log_step_img = {phase: 1 for phase in ['train', 'val']}
|
| 105 |
-
|
| 106 |
-
# text logging
|
| 107 |
-
logtxet_path = save_dir / 'training.log'
|
| 108 |
-
if self.rank == 0:
|
| 109 |
-
if logtxet_path.exists():
|
| 110 |
-
assert self.configs.resume
|
| 111 |
-
self.logger = logger
|
| 112 |
-
self.logger.remove()
|
| 113 |
-
self.logger.add(logtxet_path, format="{message}", mode='a', level='INFO')
|
| 114 |
-
self.logger.add(sys.stdout, format="{message}")
|
| 115 |
-
|
| 116 |
-
# tensorboard logging
|
| 117 |
-
log_dir = save_dir / 'tf_logs'
|
| 118 |
-
self.tf_logging = self.configs.train.tf_logging
|
| 119 |
-
if self.rank == 0 and self.tf_logging:
|
| 120 |
-
if not log_dir.exists():
|
| 121 |
-
log_dir.mkdir()
|
| 122 |
-
self.writer = SummaryWriter(str(log_dir))
|
| 123 |
-
|
| 124 |
-
# checkpoint saving
|
| 125 |
-
ckpt_dir = save_dir / 'ckpts'
|
| 126 |
-
self.ckpt_dir = ckpt_dir
|
| 127 |
-
if self.rank == 0 and (not ckpt_dir.exists()):
|
| 128 |
-
ckpt_dir.mkdir()
|
| 129 |
-
if 'ema_rate' in self.configs.train:
|
| 130 |
-
self.ema_rate = self.configs.train.ema_rate
|
| 131 |
-
assert isinstance(self.ema_rate, float), "Ema rate must be a float number"
|
| 132 |
-
ema_ckpt_dir = save_dir / 'ema_ckpts'
|
| 133 |
-
self.ema_ckpt_dir = ema_ckpt_dir
|
| 134 |
-
if self.rank == 0 and (not ema_ckpt_dir.exists()):
|
| 135 |
-
ema_ckpt_dir.mkdir()
|
| 136 |
-
|
| 137 |
-
# save images into local disk
|
| 138 |
-
self.local_logging = self.configs.train.local_logging
|
| 139 |
-
if self.rank == 0 and self.local_logging:
|
| 140 |
-
image_dir = save_dir / 'images'
|
| 141 |
-
if not image_dir.exists():
|
| 142 |
-
(image_dir / 'train').mkdir(parents=True)
|
| 143 |
-
(image_dir / 'val').mkdir(parents=True)
|
| 144 |
-
self.image_dir = image_dir
|
| 145 |
-
|
| 146 |
-
# logging the configurations
|
| 147 |
-
if self.rank == 0:
|
| 148 |
-
self.logger.info(OmegaConf.to_yaml(self.configs))
|
| 149 |
-
|
| 150 |
-
def close_logger(self):
|
| 151 |
-
if self.rank == 0 and self.tf_logging:
|
| 152 |
-
self.writer.close()
|
| 153 |
-
|
| 154 |
-
def resume_from_ckpt(self):
|
| 155 |
-
if self.configs.resume:
|
| 156 |
-
assert self.configs.resume.endswith(".pth") and os.path.isfile(self.configs.resume)
|
| 157 |
-
|
| 158 |
-
if self.rank == 0:
|
| 159 |
-
self.logger.info(f"=> Loading checkpoint from {self.configs.resume}")
|
| 160 |
-
ckpt = torch.load(self.configs.resume, map_location=f"cuda:{self.rank}")
|
| 161 |
-
util_net.reload_model(self.model, ckpt['state_dict'])
|
| 162 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 163 |
-
util_net.reload_model(self.discriminator, ckpt['state_dict_dis'])
|
| 164 |
-
torch.cuda.empty_cache()
|
| 165 |
-
|
| 166 |
-
# learning rate scheduler
|
| 167 |
-
self.iters_start = ckpt['iters_start']
|
| 168 |
-
for ii in range(1, self.iters_start+1):
|
| 169 |
-
self.adjust_lr(ii)
|
| 170 |
-
|
| 171 |
-
# logging
|
| 172 |
-
if self.rank == 0:
|
| 173 |
-
self.log_step = ckpt['log_step']
|
| 174 |
-
self.log_step_img = ckpt['log_step_img']
|
| 175 |
-
|
| 176 |
-
# EMA model
|
| 177 |
-
if self.rank == 0 and hasattr(self.configs.train, 'ema_rate'):
|
| 178 |
-
ema_ckpt_path = self.ema_ckpt_dir / ("ema_"+Path(self.configs.resume).name)
|
| 179 |
-
self.logger.info(f"=> Loading EMA checkpoint from {str(ema_ckpt_path)}")
|
| 180 |
-
ema_ckpt = torch.load(ema_ckpt_path, map_location=f"cuda:{self.rank}")
|
| 181 |
-
util_net.reload_model(self.ema_model, ema_ckpt)
|
| 182 |
-
torch.cuda.empty_cache()
|
| 183 |
-
|
| 184 |
-
# AMP scaler
|
| 185 |
-
if self.amp_scaler is not None:
|
| 186 |
-
if "amp_scaler" in ckpt:
|
| 187 |
-
self.amp_scaler.load_state_dict(ckpt["amp_scaler"])
|
| 188 |
-
if self.rank == 0:
|
| 189 |
-
self.logger.info("Loading scaler from resumed state...")
|
| 190 |
-
if self.configs.get('discriminator', None) is not None:
|
| 191 |
-
if "amp_scaler_dis" in ckpt:
|
| 192 |
-
self.amp_scaler_dis.load_state_dict(ckpt["amp_scaler_dis"])
|
| 193 |
-
if self.rank == 0:
|
| 194 |
-
self.logger.info("Loading scaler (discriminator) from resumed state...")
|
| 195 |
-
|
| 196 |
-
# reset the seed
|
| 197 |
-
self.setup_seed(seed=self.iters_start)
|
| 198 |
-
else:
|
| 199 |
-
self.iters_start = 0
|
| 200 |
-
|
| 201 |
-
def setup_optimizaton(self):
|
| 202 |
-
self.optimizer = torch.optim.AdamW(self.model.parameters(),
|
| 203 |
-
lr=self.configs.train.lr,
|
| 204 |
-
weight_decay=self.configs.train.weight_decay)
|
| 205 |
-
|
| 206 |
-
# amp settings
|
| 207 |
-
self.amp_scaler = torch.amp.GradScaler('cuda') if self.configs.train.use_amp else None
|
| 208 |
-
|
| 209 |
-
if self.configs.train.lr_schedule == 'cosin':
|
| 210 |
-
self.lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
|
| 211 |
-
optimizer=self.optimizer,
|
| 212 |
-
T_max=self.configs.train.iterations - self.configs.train.warmup_iterations,
|
| 213 |
-
eta_min=self.configs.train.lr_min,
|
| 214 |
-
)
|
| 215 |
-
|
| 216 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 217 |
-
self.optimizer_dis = torch.optim.Adam(
|
| 218 |
-
self.discriminator.parameters(),
|
| 219 |
-
lr=self.configs.train.lr_dis,
|
| 220 |
-
weight_decay=self.configs.train.weight_decay_dis,
|
| 221 |
-
)
|
| 222 |
-
self.amp_scaler_dis = torch.amp.GradScaler('cuda') if self.configs.train.use_amp else None
|
| 223 |
-
|
| 224 |
-
def prepare_compiling(self):
|
| 225 |
-
# https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/stable_diffusion_3#stable-diffusion-3
|
| 226 |
-
if not hasattr(self, "prepare_compiling_well") or (not self.prepare_compiling_well):
|
| 227 |
-
torch.set_float32_matmul_precision("high")
|
| 228 |
-
torch._inductor.config.conv_1x1_as_mm = True
|
| 229 |
-
torch._inductor.config.coordinate_descent_tuning = True
|
| 230 |
-
torch._inductor.config.epilogue_fusion = False
|
| 231 |
-
torch._inductor.config.coordinate_descent_check_all_directions = True
|
| 232 |
-
self.prepare_compiling_well = True
|
| 233 |
-
|
| 234 |
-
def build_model(self):
|
| 235 |
-
if self.configs.train.get("compile", True):
|
| 236 |
-
self.prepare_compiling()
|
| 237 |
-
|
| 238 |
-
params = self.configs.model.get('params', dict)
|
| 239 |
-
model = util_common.get_obj_from_str(self.configs.model.target)(**params)
|
| 240 |
-
model.cuda()
|
| 241 |
-
if not self.configs.train.start_mode: # Loading the starting model for evaluation
|
| 242 |
-
self.start_model = deepcopy(model)
|
| 243 |
-
assert self.configs.model.ckpt_start_path is not None
|
| 244 |
-
ckpt_start_path = self.configs.model.ckpt_start_path
|
| 245 |
-
if self.rank == 0:
|
| 246 |
-
self.logger.info(f"Loading the starting model from {ckpt_start_path}")
|
| 247 |
-
ckpt = torch.load(ckpt_start_path, map_location=f"cuda:{self.rank}")
|
| 248 |
-
if 'state_dict' in ckpt:
|
| 249 |
-
ckpt = ckpt['state_dict']
|
| 250 |
-
util_net.reload_model(self.start_model, ckpt)
|
| 251 |
-
self.freeze_model(self.start_model)
|
| 252 |
-
self.start_model.eval()
|
| 253 |
-
# delete the started timestep
|
| 254 |
-
start_timestep = max(self.configs.train.timesteps)
|
| 255 |
-
self.configs.train.timesteps.remove(start_timestep)
|
| 256 |
-
# end_timestep = min(self.configs.train.timesteps)
|
| 257 |
-
# self.configs.train.timesteps.remove(end_timestep)
|
| 258 |
-
|
| 259 |
-
# setting the training model
|
| 260 |
-
if self.configs.model.get('ckpt_path', None): # initialize if necessary
|
| 261 |
-
ckpt_path = self.configs.model.ckpt_path
|
| 262 |
-
if self.rank == 0:
|
| 263 |
-
self.logger.info(f"Initializing model from {ckpt_path}")
|
| 264 |
-
ckpt = torch.load(ckpt_path, map_location=f"cuda:{self.rank}")
|
| 265 |
-
if 'state_dict' in ckpt:
|
| 266 |
-
ckpt = ckpt['state_dict']
|
| 267 |
-
util_net.reload_model(model, ckpt)
|
| 268 |
-
if self.configs.model.get("compile", False):
|
| 269 |
-
if self.rank == 0:
|
| 270 |
-
self.logger.info("Compile the model...")
|
| 271 |
-
model.to(memory_format=torch.channels_last)
|
| 272 |
-
model = torch.compile(model, mode="max-autotune", fullgraph=False)
|
| 273 |
-
if self.num_gpus > 1:
|
| 274 |
-
model = DDP(model, device_ids=[self.rank,]) # wrap the network
|
| 275 |
-
if self.rank == 0 and hasattr(self.configs.train, 'ema_rate'):
|
| 276 |
-
self.ema_model = deepcopy(model)
|
| 277 |
-
self.freeze_model(self.ema_model)
|
| 278 |
-
self.model = model
|
| 279 |
-
|
| 280 |
-
# discriminator if necessary
|
| 281 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 282 |
-
assert hasattr(self.configs, 'discriminator')
|
| 283 |
-
params = self.configs.discriminator.get('params', dict)
|
| 284 |
-
discriminator = util_common.get_obj_from_str(self.configs.discriminator.target)(**params)
|
| 285 |
-
discriminator.cuda()
|
| 286 |
-
if self.configs.discriminator.get("compile", False):
|
| 287 |
-
if self.rank == 0:
|
| 288 |
-
self.logger.info("Compile the discriminator...")
|
| 289 |
-
discriminator.to(memory_format=torch.channels_last)
|
| 290 |
-
discriminator = torch.compile(discriminator, mode="max-autotune", fullgraph=False)
|
| 291 |
-
if self.num_gpus > 1:
|
| 292 |
-
discriminator = DDP(discriminator, device_ids=[self.rank,]) # wrap the network
|
| 293 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 294 |
-
if self.configs.discriminator.enable_grad_checkpoint:
|
| 295 |
-
if self.rank == 0:
|
| 296 |
-
self.logger.info("Activating gradient checkpointing for discriminator...")
|
| 297 |
-
self.set_grad_checkpointing(discriminator)
|
| 298 |
-
self.discriminator = discriminator
|
| 299 |
-
|
| 300 |
-
# build the stable diffusion
|
| 301 |
-
params = dict(self.configs.sd_pipe.params)
|
| 302 |
-
torch_dtype = params.pop('torch_dtype')
|
| 303 |
-
params['torch_dtype'] = get_torch_dtype(torch_dtype)
|
| 304 |
-
# loading the fp16 robust vae for sdxl: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
|
| 305 |
-
if self.configs.get('vae_fp16', None) is not None:
|
| 306 |
-
params_vae = dict(self.configs.vae_fp16.params)
|
| 307 |
-
params_vae['torch_dtype'] = torch.float16
|
| 308 |
-
pipe_id = self.configs.vae_fp16.params.pretrained_model_name_or_path
|
| 309 |
-
if self.rank == 0:
|
| 310 |
-
self.logger.info(f'Loading improved vae from {pipe_id}...')
|
| 311 |
-
vae_pipe = util_common.get_obj_from_str(self.configs.vae_fp16.target).from_pretrained(**params_vae)
|
| 312 |
-
if self.rank == 0:
|
| 313 |
-
self.logger.info('Loaded Done')
|
| 314 |
-
params['vae'] = vae_pipe
|
| 315 |
-
if ("StableDiffusion3" in self.configs.sd_pipe.target.split('.')[-1]
|
| 316 |
-
and self.configs.sd_pipe.get("model_quantization", False)):
|
| 317 |
-
if self.rank == 0:
|
| 318 |
-
self.logger.info(f'Loading the quantized transformer for SD3...')
|
| 319 |
-
nf4_config = BitsAndBytesConfig(
|
| 320 |
-
load_in_4bit=True,
|
| 321 |
-
bnb_4bit_quant_type="nf4",
|
| 322 |
-
bnb_4bit_compute_dtype=torch.bfloat16
|
| 323 |
-
)
|
| 324 |
-
params_model = dict(self.configs.model_nf4.params)
|
| 325 |
-
torch_dtype = params_model.pop('torch_dtype')
|
| 326 |
-
params_model['torch_dtype'] = get_torch_dtype(torch_dtype)
|
| 327 |
-
params_model['quantization_config'] = nf4_config
|
| 328 |
-
model_nf4 = util_common.get_obj_from_str(self.configs.model_nf4.target).from_pretrained(
|
| 329 |
-
**params_model
|
| 330 |
-
)
|
| 331 |
-
params['transformer'] = model_nf4
|
| 332 |
-
sd_pipe = util_common.get_obj_from_str(self.configs.sd_pipe.target).from_pretrained(**params)
|
| 333 |
-
if self.configs.get('scheduler', None) is not None:
|
| 334 |
-
pipe_id = self.configs.scheduler.target.split('.')[-1]
|
| 335 |
-
if self.rank == 0:
|
| 336 |
-
self.logger.info(f'Loading scheduler of {pipe_id}...')
|
| 337 |
-
sd_pipe.scheduler = util_common.get_obj_from_str(self.configs.scheduler.target).from_config(
|
| 338 |
-
sd_pipe.scheduler.config
|
| 339 |
-
)
|
| 340 |
-
if self.rank == 0:
|
| 341 |
-
self.logger.info('Loaded Done')
|
| 342 |
-
if ("StableDiffusion3" in self.configs.sd_pipe.target.split('.')[-1]
|
| 343 |
-
and self.configs.sd_pipe.get("model_quantization", False)):
|
| 344 |
-
sd_pipe.enable_model_cpu_offload(gpu_id=self.rank,device='cuda')
|
| 345 |
-
else:
|
| 346 |
-
sd_pipe.to(f"cuda:{self.rank}")
|
| 347 |
-
# freezing model parameters
|
| 348 |
-
if hasattr(sd_pipe, 'unet'):
|
| 349 |
-
self.freeze_model(sd_pipe.unet)
|
| 350 |
-
if hasattr(sd_pipe, 'transformer'):
|
| 351 |
-
self.freeze_model(sd_pipe.transformer)
|
| 352 |
-
self.freeze_model(sd_pipe.vae)
|
| 353 |
-
# compiling
|
| 354 |
-
if self.configs.sd_pipe.get('compile', True):
|
| 355 |
-
if self.rank == 0:
|
| 356 |
-
self.logger.info('Compile the SD model...')
|
| 357 |
-
sd_pipe.set_progress_bar_config(disable=True)
|
| 358 |
-
if hasattr(sd_pipe, 'unet'):
|
| 359 |
-
sd_pipe.unet.to(memory_format=torch.channels_last)
|
| 360 |
-
sd_pipe.unet = torch.compile(sd_pipe.unet, mode="max-autotune", fullgraph=False)
|
| 361 |
-
if hasattr(sd_pipe, 'transformer'):
|
| 362 |
-
sd_pipe.transformer.to(memory_format=torch.channels_last)
|
| 363 |
-
sd_pipe.transformer = torch.compile(sd_pipe.transformer, mode="max-autotune", fullgraph=False)
|
| 364 |
-
sd_pipe.vae.to(memory_format=torch.channels_last)
|
| 365 |
-
sd_pipe.vae = torch.compile(sd_pipe.vae, mode="max-autotune", fullgraph=True)
|
| 366 |
-
# setting gradient checkpoint for vae
|
| 367 |
-
if self.configs.sd_pipe.get("enable_grad_checkpoint_vae", True):
|
| 368 |
-
if self.rank == 0:
|
| 369 |
-
self.logger.info("Activating gradient checkpointing for VAE...")
|
| 370 |
-
sd_pipe.vae._set_gradient_checkpointing(sd_pipe.vae.encoder)
|
| 371 |
-
sd_pipe.vae._set_gradient_checkpointing(sd_pipe.vae.decoder)
|
| 372 |
-
# setting gradient checkpoint for diffusion model
|
| 373 |
-
if self.configs.sd_pipe.enable_grad_checkpoint:
|
| 374 |
-
if self.rank == 0:
|
| 375 |
-
self.logger.info("Activating gradient checkpointing for SD...")
|
| 376 |
-
if hasattr(sd_pipe, 'unet'):
|
| 377 |
-
self.set_grad_checkpointing(sd_pipe.unet)
|
| 378 |
-
if hasattr(sd_pipe, 'transformer'):
|
| 379 |
-
self.set_grad_checkpointing(sd_pipe.transformer)
|
| 380 |
-
self.sd_pipe = sd_pipe
|
| 381 |
-
|
| 382 |
-
# latent LPIPS loss
|
| 383 |
-
if self.configs.train.loss_coef.get('llpips', 0) > 0:
|
| 384 |
-
params = self.configs.llpips.get('params', dict)
|
| 385 |
-
llpips_loss = util_common.get_obj_from_str(self.configs.llpips.target)(**params)
|
| 386 |
-
llpips_loss.cuda()
|
| 387 |
-
self.freeze_model(llpips_loss)
|
| 388 |
-
|
| 389 |
-
# loading the pre-trained model
|
| 390 |
-
ckpt_path = self.configs.llpips.ckpt_path
|
| 391 |
-
self.load_model(llpips_loss, ckpt_path, tag='latent lpips')
|
| 392 |
-
|
| 393 |
-
if self.configs.llpips.get("compile", True):
|
| 394 |
-
if self.rank == 0:
|
| 395 |
-
self.logger.info('Compile the llpips loss...')
|
| 396 |
-
llpips_loss.to(memory_format=torch.channels_last)
|
| 397 |
-
llpips_loss = torch.compile(llpips_loss, mode="max-autotune", fullgraph=True)
|
| 398 |
-
|
| 399 |
-
self.llpips_loss = llpips_loss
|
| 400 |
-
|
| 401 |
-
# model information
|
| 402 |
-
self.print_model_info()
|
| 403 |
-
|
| 404 |
-
torch.cuda.empty_cache()
|
| 405 |
-
|
| 406 |
-
def set_grad_checkpointing(self, model):
|
| 407 |
-
if hasattr(model, 'down_blocks'):
|
| 408 |
-
for module in model.down_blocks:
|
| 409 |
-
module.gradient_checkpointing = True
|
| 410 |
-
module.training = True
|
| 411 |
-
|
| 412 |
-
if hasattr(model, 'up_blocks'):
|
| 413 |
-
for module in model.up_blocks:
|
| 414 |
-
module.gradient_checkpointing = True
|
| 415 |
-
module.training = True
|
| 416 |
-
|
| 417 |
-
if hasattr(model, 'mid_blocks'):
|
| 418 |
-
model.mid_block.gradient_checkpointing = True
|
| 419 |
-
model.mid_block.training = True
|
| 420 |
-
|
| 421 |
-
def build_dataloader(self):
|
| 422 |
-
def _wrap_loader(loader):
|
| 423 |
-
while True: yield from loader
|
| 424 |
-
|
| 425 |
-
# make datasets
|
| 426 |
-
datasets = {'train': create_dataset(self.configs.data.get('train', dict)), }
|
| 427 |
-
if hasattr(self.configs.data, 'val') and self.rank == 0:
|
| 428 |
-
datasets['val'] = create_dataset(self.configs.data.get('val', dict))
|
| 429 |
-
if self.rank == 0:
|
| 430 |
-
for phase in datasets.keys():
|
| 431 |
-
length = len(datasets[phase])
|
| 432 |
-
self.logger.info('Number of images in {:s} data set: {:d}'.format(phase, length))
|
| 433 |
-
|
| 434 |
-
# make dataloaders
|
| 435 |
-
if self.num_gpus > 1:
|
| 436 |
-
sampler = udata.distributed.DistributedSampler(
|
| 437 |
-
datasets['train'],
|
| 438 |
-
num_replicas=self.num_gpus,
|
| 439 |
-
rank=self.rank,
|
| 440 |
-
)
|
| 441 |
-
else:
|
| 442 |
-
sampler = None
|
| 443 |
-
dataloaders = {'train': _wrap_loader(udata.DataLoader(
|
| 444 |
-
datasets['train'],
|
| 445 |
-
batch_size=self.configs.train.batch // self.num_gpus,
|
| 446 |
-
shuffle=False if self.num_gpus > 1 else True,
|
| 447 |
-
drop_last=True,
|
| 448 |
-
num_workers=min(self.configs.train.num_workers, 4),
|
| 449 |
-
pin_memory=True,
|
| 450 |
-
prefetch_factor=self.configs.train.get('prefetch_factor', 2),
|
| 451 |
-
worker_init_fn=my_worker_init_fn,
|
| 452 |
-
sampler=sampler,
|
| 453 |
-
))}
|
| 454 |
-
if hasattr(self.configs.data, 'val') and self.rank == 0:
|
| 455 |
-
dataloaders['val'] = udata.DataLoader(datasets['val'],
|
| 456 |
-
batch_size=self.configs.validate.batch,
|
| 457 |
-
shuffle=False,
|
| 458 |
-
drop_last=False,
|
| 459 |
-
num_workers=0,
|
| 460 |
-
pin_memory=True,
|
| 461 |
-
)
|
| 462 |
-
|
| 463 |
-
self.datasets = datasets
|
| 464 |
-
self.dataloaders = dataloaders
|
| 465 |
-
self.sampler = sampler
|
| 466 |
-
|
| 467 |
-
def print_model_info(self):
|
| 468 |
-
if self.rank == 0:
|
| 469 |
-
num_params = util_net.calculate_parameters(self.model) / 1000**2
|
| 470 |
-
# self.logger.info("Detailed network architecture:")
|
| 471 |
-
# self.logger.info(self.model.__repr__())
|
| 472 |
-
if self.configs.train.get('use_fsdp', False):
|
| 473 |
-
num_params *= self.num_gpus
|
| 474 |
-
self.logger.info(f"Number of parameters: {num_params:.2f}M")
|
| 475 |
-
|
| 476 |
-
if hasattr(self, 'discriminator'):
|
| 477 |
-
num_params = util_net.calculate_parameters(self.discriminator) / 1000**2
|
| 478 |
-
self.logger.info(f"Number of parameters in discriminator: {num_params:.2f}M")
|
| 479 |
-
|
| 480 |
-
def prepare_data(self, data, dtype=torch.float32, phase='train'):
|
| 481 |
-
data = {key:value.cuda().to(dtype=dtype) for key, value in data.items()}
|
| 482 |
-
return data
|
| 483 |
-
|
| 484 |
-
def validation(self):
|
| 485 |
-
pass
|
| 486 |
-
|
| 487 |
-
def train(self):
|
| 488 |
-
self.init_logger() # setup logger: self.logger
|
| 489 |
-
|
| 490 |
-
self.build_dataloader() # prepare data: self.dataloaders, self.datasets, self.sampler
|
| 491 |
-
|
| 492 |
-
self.build_model() # build model: self.model, self.loss
|
| 493 |
-
|
| 494 |
-
self.setup_optimizaton() # setup optimization: self.optimzer, self.sheduler
|
| 495 |
-
|
| 496 |
-
self.resume_from_ckpt() # resume if necessary
|
| 497 |
-
|
| 498 |
-
self.model.train()
|
| 499 |
-
num_iters_epoch = math.ceil(len(self.datasets['train']) / self.configs.train.batch)
|
| 500 |
-
for ii in range(self.iters_start, self.configs.train.iterations):
|
| 501 |
-
self.current_iters = ii + 1
|
| 502 |
-
|
| 503 |
-
# prepare data
|
| 504 |
-
data = self.prepare_data(next(self.dataloaders['train']), phase='train')
|
| 505 |
-
|
| 506 |
-
# training phase
|
| 507 |
-
self.training_step(data)
|
| 508 |
-
|
| 509 |
-
# update ema model
|
| 510 |
-
if hasattr(self.configs.train, 'ema_rate') and self.rank == 0:
|
| 511 |
-
self.update_ema_model()
|
| 512 |
-
|
| 513 |
-
# validation phase
|
| 514 |
-
if ((ii+1) % self.configs.train.save_freq == 0 and
|
| 515 |
-
'val' in self.dataloaders and
|
| 516 |
-
self.rank == 0
|
| 517 |
-
):
|
| 518 |
-
self.validation()
|
| 519 |
-
|
| 520 |
-
#update learning rate
|
| 521 |
-
self.adjust_lr()
|
| 522 |
-
|
| 523 |
-
# save checkpoint
|
| 524 |
-
if (ii+1) % self.configs.train.save_freq == 0 and self.rank == 0:
|
| 525 |
-
self.save_ckpt()
|
| 526 |
-
|
| 527 |
-
if (ii+1) % num_iters_epoch == 0 and self.sampler is not None:
|
| 528 |
-
self.sampler.set_epoch(ii+1)
|
| 529 |
-
|
| 530 |
-
# close the tensorboard
|
| 531 |
-
self.close_logger()
|
| 532 |
-
|
| 533 |
-
def adjust_lr(self, current_iters=None):
|
| 534 |
-
base_lr = self.configs.train.lr
|
| 535 |
-
warmup_steps = self.configs.train.get("warmup_iterations", 0)
|
| 536 |
-
current_iters = self.current_iters if current_iters is None else current_iters
|
| 537 |
-
if current_iters <= warmup_steps:
|
| 538 |
-
for params_group in self.optimizer.param_groups:
|
| 539 |
-
params_group['lr'] = (current_iters / warmup_steps) * base_lr
|
| 540 |
-
else:
|
| 541 |
-
if hasattr(self, 'lr_scheduler'):
|
| 542 |
-
self.lr_scheduler.step()
|
| 543 |
-
|
| 544 |
-
def save_ckpt(self):
|
| 545 |
-
ckpt_path = self.ckpt_dir / 'model_{:d}.pth'.format(self.current_iters)
|
| 546 |
-
ckpt = {
|
| 547 |
-
'iters_start': self.current_iters,
|
| 548 |
-
'log_step': {phase:self.log_step[phase] for phase in ['train', 'val']},
|
| 549 |
-
'log_step_img': {phase:self.log_step_img[phase] for phase in ['train', 'val']},
|
| 550 |
-
'state_dict': self.model.state_dict(),
|
| 551 |
-
}
|
| 552 |
-
if self.amp_scaler is not None:
|
| 553 |
-
ckpt['amp_scaler'] = self.amp_scaler.state_dict()
|
| 554 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 555 |
-
ckpt['state_dict_dis'] = self.discriminator.state_dict()
|
| 556 |
-
if self.amp_scaler_dis is not None:
|
| 557 |
-
ckpt['amp_scaler_dis'] = self.amp_scaler_dis.state_dict()
|
| 558 |
-
torch.save(ckpt, ckpt_path)
|
| 559 |
-
if hasattr(self.configs.train, 'ema_rate'):
|
| 560 |
-
ema_ckpt_path = self.ema_ckpt_dir / 'ema_model_{:d}.pth'.format(self.current_iters)
|
| 561 |
-
torch.save(self.ema_model.state_dict(), ema_ckpt_path)
|
| 562 |
-
|
| 563 |
-
def logging_image(self, im_tensor, tag, phase, add_global_step=False, nrow=8):
|
| 564 |
-
"""
|
| 565 |
-
Args:
|
| 566 |
-
im_tensor: b x c x h x w tensor
|
| 567 |
-
im_tag: str
|
| 568 |
-
phase: 'train' or 'val'
|
| 569 |
-
nrow: number of displays in each row
|
| 570 |
-
"""
|
| 571 |
-
assert self.tf_logging or self.local_logging
|
| 572 |
-
im_tensor = vutils.make_grid(im_tensor, nrow=nrow, normalize=True, scale_each=True) # c x H x W
|
| 573 |
-
if self.local_logging:
|
| 574 |
-
im_path = str(self.image_dir / phase / f"{tag}-{self.log_step_img[phase]}.png")
|
| 575 |
-
im_np = im_tensor.cpu().permute(1,2,0).numpy()
|
| 576 |
-
util_image.imwrite(im_np, im_path)
|
| 577 |
-
if self.tf_logging:
|
| 578 |
-
self.writer.add_image(
|
| 579 |
-
f"{phase}-{tag}-{self.log_step_img[phase]}",
|
| 580 |
-
im_tensor,
|
| 581 |
-
self.log_step_img[phase],
|
| 582 |
-
)
|
| 583 |
-
if add_global_step:
|
| 584 |
-
self.log_step_img[phase] += 1
|
| 585 |
-
|
| 586 |
-
def logging_text(self, text_list, phase):
|
| 587 |
-
"""
|
| 588 |
-
Args:
|
| 589 |
-
text_list: (b,) list
|
| 590 |
-
phase: 'train' or 'val'
|
| 591 |
-
"""
|
| 592 |
-
assert self.local_logging
|
| 593 |
-
if self.local_logging:
|
| 594 |
-
text_path = str(self.image_dir / phase / f"text-{self.log_step_img[phase]}.txt")
|
| 595 |
-
with open(text_path, 'w') as ff:
|
| 596 |
-
for text in text_list:
|
| 597 |
-
ff.write(text + '\n')
|
| 598 |
-
|
| 599 |
-
def logging_metric(self, metrics, tag, phase, add_global_step=False):
|
| 600 |
-
"""
|
| 601 |
-
Args:
|
| 602 |
-
metrics: dict
|
| 603 |
-
tag: str
|
| 604 |
-
phase: 'train' or 'val'
|
| 605 |
-
"""
|
| 606 |
-
if self.tf_logging:
|
| 607 |
-
tag = f"{phase}-{tag}"
|
| 608 |
-
if isinstance(metrics, dict):
|
| 609 |
-
self.writer.add_scalars(tag, metrics, self.log_step[phase])
|
| 610 |
-
else:
|
| 611 |
-
self.writer.add_scalar(tag, metrics, self.log_step[phase])
|
| 612 |
-
if add_global_step:
|
| 613 |
-
self.log_step[phase] += 1
|
| 614 |
-
else:
|
| 615 |
-
pass
|
| 616 |
-
|
| 617 |
-
def load_model(self, model, ckpt_path=None, tag='model'):
|
| 618 |
-
if self.rank == 0:
|
| 619 |
-
self.logger.info(f'Loading {tag} from {ckpt_path}...')
|
| 620 |
-
ckpt = torch.load(ckpt_path, map_location=f"cuda:{self.rank}")
|
| 621 |
-
if 'state_dict' in ckpt:
|
| 622 |
-
ckpt = ckpt['state_dict']
|
| 623 |
-
util_net.reload_model(model, ckpt)
|
| 624 |
-
if self.rank == 0:
|
| 625 |
-
self.logger.info('Loaded Done')
|
| 626 |
-
|
| 627 |
-
def freeze_model(self, net):
|
| 628 |
-
for params in net.parameters():
|
| 629 |
-
params.requires_grad = False
|
| 630 |
-
|
| 631 |
-
def unfreeze_model(self, net):
|
| 632 |
-
for params in net.parameters():
|
| 633 |
-
params.requires_grad = True
|
| 634 |
-
|
| 635 |
-
@torch.no_grad()
|
| 636 |
-
def update_ema_model(self):
|
| 637 |
-
decay = min(self.configs.train.ema_rate, (1 + self.current_iters) / (10 + self.current_iters))
|
| 638 |
-
target_params = dict(self.model.named_parameters())
|
| 639 |
-
# if hasattr(self.configs.train, 'ema_rate'):
|
| 640 |
-
# with FSDP.summon_full_params(self.model, writeback=True):
|
| 641 |
-
# target_params = dict(self.model.named_parameters())
|
| 642 |
-
# else:
|
| 643 |
-
# target_params = dict(self.model.named_parameters())
|
| 644 |
-
|
| 645 |
-
one_minus_decay = 1.0 - decay
|
| 646 |
-
|
| 647 |
-
for key, source_value in self.ema_model.named_parameters():
|
| 648 |
-
target_value = target_params[key]
|
| 649 |
-
if target_value.requires_grad:
|
| 650 |
-
source_value.sub_(one_minus_decay * (source_value - target_value.data))
|
| 651 |
-
|
| 652 |
-
class TrainerBaseSR(TrainerBase):
|
| 653 |
-
@torch.no_grad()
|
| 654 |
-
def _dequeue_and_enqueue(self):
|
| 655 |
-
"""It is the training pair pool for increasing the diversity in a batch.
|
| 656 |
-
|
| 657 |
-
Batch processing limits the diversity of synthetic degradations in a batch. For example, samples in a
|
| 658 |
-
batch could not have different resize scaling factors. Therefore, we employ this training pair pool
|
| 659 |
-
to increase the degradation diversity in a batch.
|
| 660 |
-
"""
|
| 661 |
-
# initialize
|
| 662 |
-
b, c, h, w = self.lq.size()
|
| 663 |
-
if not hasattr(self, 'queue_size'):
|
| 664 |
-
self.queue_size = self.configs.degradation.get('queue_size', b*10)
|
| 665 |
-
if not hasattr(self, 'queue_lr'):
|
| 666 |
-
assert self.queue_size % b == 0, f'queue size {self.queue_size} should be divisible by batch size {b}'
|
| 667 |
-
self.queue_lr = torch.zeros(self.queue_size, c, h, w).cuda()
|
| 668 |
-
_, c, h, w = self.gt.size()
|
| 669 |
-
self.queue_gt = torch.zeros(self.queue_size, c, h, w).cuda()
|
| 670 |
-
_, c, h, w = self.gt_latent.size()
|
| 671 |
-
self.queue_gt_latent = torch.zeros(self.queue_size, c, h, w).cuda()
|
| 672 |
-
self.queue_txt = ["", ] * self.queue_size
|
| 673 |
-
self.queue_ptr = 0
|
| 674 |
-
if self.queue_ptr == self.queue_size: # the pool is full
|
| 675 |
-
# do dequeue and enqueue
|
| 676 |
-
# shuffle
|
| 677 |
-
idx = torch.randperm(self.queue_size)
|
| 678 |
-
self.queue_lr = self.queue_lr[idx]
|
| 679 |
-
self.queue_gt = self.queue_gt[idx]
|
| 680 |
-
self.queue_gt_latent = self.queue_gt_latent[idx]
|
| 681 |
-
self.queue_txt = [self.queue_txt[ii] for ii in idx]
|
| 682 |
-
# get first b samples
|
| 683 |
-
lq_dequeue = self.queue_lr[0:b, :, :, :].clone()
|
| 684 |
-
gt_dequeue = self.queue_gt[0:b, :, :, :].clone()
|
| 685 |
-
gt_latent_dequeue = self.queue_gt_latent[0:b, :, :, :].clone()
|
| 686 |
-
txt_dequeue = deepcopy(self.queue_txt[0:b])
|
| 687 |
-
# update the queue
|
| 688 |
-
self.queue_lr[0:b, :, :, :] = self.lq.clone()
|
| 689 |
-
self.queue_gt[0:b, :, :, :] = self.gt.clone()
|
| 690 |
-
self.queue_gt_latent[0:b, :, :, :] = self.gt_latent.clone()
|
| 691 |
-
self.queue_txt[0:b] = deepcopy(self.txt)
|
| 692 |
-
|
| 693 |
-
self.lq = lq_dequeue
|
| 694 |
-
self.gt = gt_dequeue
|
| 695 |
-
self.gt_latent = gt_latent_dequeue
|
| 696 |
-
self.txt = txt_dequeue
|
| 697 |
-
else:
|
| 698 |
-
# only do enqueue
|
| 699 |
-
self.queue_lr[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.lq.clone()
|
| 700 |
-
self.queue_gt[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.gt.clone()
|
| 701 |
-
self.queue_gt_latent[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.gt_latent.clone()
|
| 702 |
-
self.queue_txt[self.queue_ptr:self.queue_ptr + b] = deepcopy(self.txt)
|
| 703 |
-
self.queue_ptr = self.queue_ptr + b
|
| 704 |
-
|
| 705 |
-
@torch.no_grad()
|
| 706 |
-
def prepare_data(self, data, phase='train'):
|
| 707 |
-
if phase == 'train' and self.configs.data.get(phase).get('type') == 'realesrgan':
|
| 708 |
-
if not hasattr(self, 'jpeger'):
|
| 709 |
-
self.jpeger = DiffJPEG(differentiable=False).cuda() # simulate JPEG compression artifacts
|
| 710 |
-
if (not hasattr(self, 'sharpener')) and self.configs.degradation.get('use_sharp', False):
|
| 711 |
-
self.sharpener = USMSharp().cuda()
|
| 712 |
-
|
| 713 |
-
im_gt = data['gt'].cuda()
|
| 714 |
-
kernel1 = data['kernel1'].cuda()
|
| 715 |
-
kernel2 = data['kernel2'].cuda()
|
| 716 |
-
sinc_kernel = data['sinc_kernel'].cuda()
|
| 717 |
-
|
| 718 |
-
ori_h, ori_w = im_gt.size()[2:4]
|
| 719 |
-
if isinstance(self.configs.degradation.sf, int):
|
| 720 |
-
sf = self.configs.degradation.sf
|
| 721 |
-
else:
|
| 722 |
-
assert len(self.configs.degradation.sf) == 2
|
| 723 |
-
sf = random.uniform(*self.configs.degradation.sf)
|
| 724 |
-
|
| 725 |
-
if self.configs.degradation.use_sharp:
|
| 726 |
-
im_gt = self.sharpener(im_gt)
|
| 727 |
-
|
| 728 |
-
# ----------------------- The first degradation process ----------------------- #
|
| 729 |
-
# blur
|
| 730 |
-
out = filter2D(im_gt, kernel1)
|
| 731 |
-
# random resize
|
| 732 |
-
updown_type = random.choices(
|
| 733 |
-
['up', 'down', 'keep'],
|
| 734 |
-
self.configs.degradation['resize_prob'],
|
| 735 |
-
)[0]
|
| 736 |
-
if updown_type == 'up':
|
| 737 |
-
scale = random.uniform(1, self.configs.degradation['resize_range'][1])
|
| 738 |
-
elif updown_type == 'down':
|
| 739 |
-
scale = random.uniform(self.configs.degradation['resize_range'][0], 1)
|
| 740 |
-
else:
|
| 741 |
-
scale = 1
|
| 742 |
-
mode = random.choice(['area', 'bilinear', 'bicubic'])
|
| 743 |
-
out = F.interpolate(out, scale_factor=scale, mode=mode)
|
| 744 |
-
# add noise
|
| 745 |
-
gray_noise_prob = self.configs.degradation['gray_noise_prob']
|
| 746 |
-
if random.random() < self.configs.degradation['gaussian_noise_prob']:
|
| 747 |
-
out = random_add_gaussian_noise_pt(
|
| 748 |
-
out,
|
| 749 |
-
sigma_range=self.configs.degradation['noise_range'],
|
| 750 |
-
clip=True,
|
| 751 |
-
rounds=False,
|
| 752 |
-
gray_prob=gray_noise_prob,
|
| 753 |
-
)
|
| 754 |
-
else:
|
| 755 |
-
out = random_add_poisson_noise_pt(
|
| 756 |
-
out,
|
| 757 |
-
scale_range=self.configs.degradation['poisson_scale_range'],
|
| 758 |
-
gray_prob=gray_noise_prob,
|
| 759 |
-
clip=True,
|
| 760 |
-
rounds=False)
|
| 761 |
-
# JPEG compression
|
| 762 |
-
jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.configs.degradation['jpeg_range'])
|
| 763 |
-
out = torch.clamp(out, 0, 1) # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts
|
| 764 |
-
out = self.jpeger(out, quality=jpeg_p)
|
| 765 |
-
|
| 766 |
-
# ----------------------- The second degradation process ----------------------- #
|
| 767 |
-
if random.random() < self.configs.degradation['second_order_prob']:
|
| 768 |
-
# blur
|
| 769 |
-
if random.random() < self.configs.degradation['second_blur_prob']:
|
| 770 |
-
out = filter2D(out, kernel2)
|
| 771 |
-
# random resize
|
| 772 |
-
updown_type = random.choices(
|
| 773 |
-
['up', 'down', 'keep'],
|
| 774 |
-
self.configs.degradation['resize_prob2'],
|
| 775 |
-
)[0]
|
| 776 |
-
if updown_type == 'up':
|
| 777 |
-
scale = random.uniform(1, self.configs.degradation['resize_range2'][1])
|
| 778 |
-
elif updown_type == 'down':
|
| 779 |
-
scale = random.uniform(self.configs.degradation['resize_range2'][0], 1)
|
| 780 |
-
else:
|
| 781 |
-
scale = 1
|
| 782 |
-
mode = random.choice(['area', 'bilinear', 'bicubic'])
|
| 783 |
-
out = F.interpolate(
|
| 784 |
-
out,
|
| 785 |
-
size=(int(ori_h / sf * scale), int(ori_w / sf * scale)),
|
| 786 |
-
mode=mode,
|
| 787 |
-
)
|
| 788 |
-
# add noise
|
| 789 |
-
gray_noise_prob = self.configs.degradation['gray_noise_prob2']
|
| 790 |
-
if random.random() < self.configs.degradation['gaussian_noise_prob2']:
|
| 791 |
-
out = random_add_gaussian_noise_pt(
|
| 792 |
-
out,
|
| 793 |
-
sigma_range=self.configs.degradation['noise_range2'],
|
| 794 |
-
clip=True,
|
| 795 |
-
rounds=False,
|
| 796 |
-
gray_prob=gray_noise_prob,
|
| 797 |
-
)
|
| 798 |
-
else:
|
| 799 |
-
out = random_add_poisson_noise_pt(
|
| 800 |
-
out,
|
| 801 |
-
scale_range=self.configs.degradation['poisson_scale_range2'],
|
| 802 |
-
gray_prob=gray_noise_prob,
|
| 803 |
-
clip=True,
|
| 804 |
-
rounds=False,
|
| 805 |
-
)
|
| 806 |
-
|
| 807 |
-
# JPEG compression + the final sinc filter
|
| 808 |
-
# We also need to resize images to desired sizes. We group [resize back + sinc filter] together
|
| 809 |
-
# as one operation.
|
| 810 |
-
# We consider two orders:
|
| 811 |
-
# 1. [resize back + sinc filter] + JPEG compression
|
| 812 |
-
# 2. JPEG compression + [resize back + sinc filter]
|
| 813 |
-
# Empirically, we find other combinations (sinc + JPEG + Resize) will introduce twisted lines.
|
| 814 |
-
if random.random() < 0.5:
|
| 815 |
-
# resize back + the final sinc filter
|
| 816 |
-
mode = random.choice(['area', 'bilinear', 'bicubic'])
|
| 817 |
-
out = F.interpolate(
|
| 818 |
-
out,
|
| 819 |
-
size=(ori_h // sf, ori_w // sf),
|
| 820 |
-
mode=mode,
|
| 821 |
-
)
|
| 822 |
-
out = filter2D(out, sinc_kernel)
|
| 823 |
-
# JPEG compression
|
| 824 |
-
jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.configs.degradation['jpeg_range2'])
|
| 825 |
-
out = torch.clamp(out, 0, 1)
|
| 826 |
-
out = self.jpeger(out, quality=jpeg_p)
|
| 827 |
-
else:
|
| 828 |
-
# JPEG compression
|
| 829 |
-
jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.configs.degradation['jpeg_range2'])
|
| 830 |
-
out = torch.clamp(out, 0, 1)
|
| 831 |
-
out = self.jpeger(out, quality=jpeg_p)
|
| 832 |
-
# resize back + the final sinc filter
|
| 833 |
-
mode = random.choice(['area', 'bilinear', 'bicubic'])
|
| 834 |
-
out = F.interpolate(
|
| 835 |
-
out,
|
| 836 |
-
size=(ori_h // sf, ori_w // sf),
|
| 837 |
-
mode=mode,
|
| 838 |
-
)
|
| 839 |
-
out = filter2D(out, sinc_kernel)
|
| 840 |
-
|
| 841 |
-
# resize back
|
| 842 |
-
if self.configs.degradation.resize_back:
|
| 843 |
-
out = F.interpolate(out, size=(ori_h, ori_w), mode=_INTERPOLATION_MODE)
|
| 844 |
-
|
| 845 |
-
# clamp and round
|
| 846 |
-
im_lq = torch.clamp((out * 255.0).round(), 0, 255) / 255.
|
| 847 |
-
|
| 848 |
-
self.lq, self.gt, self.txt = im_lq, im_gt, data['txt']
|
| 849 |
-
if "gt_moment" not in data:
|
| 850 |
-
self.gt_latent = self.encode_first_stage(
|
| 851 |
-
im_gt.cuda(),
|
| 852 |
-
center_input_sample=True,
|
| 853 |
-
deterministic=self.configs.train.loss_coef.get('rkl', 0) > 0,
|
| 854 |
-
)
|
| 855 |
-
else:
|
| 856 |
-
self.gt_latent = self.encode_from_moment(
|
| 857 |
-
data['gt_moment'].cuda(),
|
| 858 |
-
deterministic=self.configs.train.loss_coef.get('rkl', 0) > 0,
|
| 859 |
-
)
|
| 860 |
-
|
| 861 |
-
if (not self.configs.train.use_text) or self.configs.data.train.params.random_crop:
|
| 862 |
-
self.txt = [_positive,] * im_lq.shape[0]
|
| 863 |
-
|
| 864 |
-
# training pair pool
|
| 865 |
-
self._dequeue_and_enqueue()
|
| 866 |
-
self.lq = self.lq.contiguous() # for the warning: grad and param do not obey the gradient layout contract
|
| 867 |
-
|
| 868 |
-
batch = {'lq':self.lq, 'gt':self.gt, 'gt_latent':self.gt_latent, 'txt':self.txt}
|
| 869 |
-
elif phase == 'val':
|
| 870 |
-
resolution = self.configs.data.train.params.gt_size // self.configs.degradation.sf
|
| 871 |
-
batch = {}
|
| 872 |
-
batch['lq'] = data['lq'].cuda()
|
| 873 |
-
if 'gt' in data:
|
| 874 |
-
batch['gt'] = data['gt'].cuda()
|
| 875 |
-
batch['txt'] = [_positive, ] * data['lq'].shape[0]
|
| 876 |
-
else:
|
| 877 |
-
batch = {key:value.cuda().to(dtype=torch.float32) for key, value in data.items()}
|
| 878 |
-
|
| 879 |
-
return batch
|
| 880 |
-
|
| 881 |
-
@torch.no_grad()
|
| 882 |
-
def encode_from_moment(self, z, deterministic=True):
|
| 883 |
-
dist = DiagonalGaussianDistribution(z)
|
| 884 |
-
init_latents = dist.mode() if deterministic else dist.sample()
|
| 885 |
-
|
| 886 |
-
latents_mean = latents_std = None
|
| 887 |
-
if hasattr(self.sd_pipe.vae.config, "latents_mean") and self.sd_pipe.vae.config.latents_mean is not None:
|
| 888 |
-
latents_mean = torch.tensor(self.sd_pipe.vae.config.latents_mean).view(1, 4, 1, 1)
|
| 889 |
-
if hasattr(self.sd_pipe.vae.config, "latents_std") and self.sd_pipe.vae.config.latents_std is not None:
|
| 890 |
-
latents_std = torch.tensor(self.sd_pipe.vae.config.latents_std).view(1, 4, 1, 1)
|
| 891 |
-
|
| 892 |
-
scaling_factor = self.sd_pipe.vae.config.scaling_factor
|
| 893 |
-
if latents_mean is not None and latents_std is not None:
|
| 894 |
-
latents_mean = latents_mean.to(device=z.device, dtype=z.dtype)
|
| 895 |
-
latents_std = latents_std.to(device=z.device, dtype=z.dtype)
|
| 896 |
-
init_latents = (init_latents - latents_mean) * scaling_factor / latents_std
|
| 897 |
-
else:
|
| 898 |
-
init_latents = init_latents * scaling_factor
|
| 899 |
-
|
| 900 |
-
return init_latents
|
| 901 |
-
|
| 902 |
-
@torch.no_grad()
|
| 903 |
-
@torch.amp.autocast('cuda')
|
| 904 |
-
def encode_first_stage(self, x, deterministic=False, center_input_sample=True):
|
| 905 |
-
if center_input_sample:
|
| 906 |
-
x = x * 2.0 - 1.0
|
| 907 |
-
latents_mean = latents_std = None
|
| 908 |
-
if hasattr(self.sd_pipe.vae.config, "latents_mean") and self.sd_pipe.vae.config.latents_mean is not None:
|
| 909 |
-
latents_mean = torch.tensor(self.sd_pipe.vae.config.latents_mean).view(1, -1, 1, 1)
|
| 910 |
-
if hasattr(self.sd_pipe.vae.config, "latents_std") and self.sd_pipe.vae.config.latents_std is not None:
|
| 911 |
-
latents_std = torch.tensor(self.sd_pipe.vae.config.latents_std).view(1, -1, 1, 1)
|
| 912 |
-
|
| 913 |
-
if deterministic:
|
| 914 |
-
partial_encode = lambda xx: self.sd_pipe.vae.encode(xx).latent_dist.mode()
|
| 915 |
-
else:
|
| 916 |
-
partial_encode = lambda xx: self.sd_pipe.vae.encode(xx).latent_dist.sample()
|
| 917 |
-
|
| 918 |
-
trunk_size = self.configs.sd_pipe.vae_split
|
| 919 |
-
if trunk_size < x.shape[0]:
|
| 920 |
-
init_latents = torch.cat([partial_encode(xx) for xx in x.split(trunk_size, 0)], dim=0)
|
| 921 |
-
else:
|
| 922 |
-
init_latents = partial_encode(x)
|
| 923 |
-
|
| 924 |
-
scaling_factor = self.sd_pipe.vae.config.scaling_factor
|
| 925 |
-
if latents_mean is not None and latents_std is not None:
|
| 926 |
-
latents_mean = latents_mean.to(device=x.device, dtype=x.dtype)
|
| 927 |
-
latents_std = latents_std.to(device=x.device, dtype=x.dtype)
|
| 928 |
-
init_latents = (init_latents - latents_mean) * scaling_factor / latents_std
|
| 929 |
-
else:
|
| 930 |
-
init_latents = init_latents * scaling_factor
|
| 931 |
-
|
| 932 |
-
return init_latents
|
| 933 |
-
|
| 934 |
-
@torch.no_grad()
|
| 935 |
-
@torch.amp.autocast('cuda')
|
| 936 |
-
def decode_first_stage(self, z, clamp=True):
|
| 937 |
-
z = z / self.sd_pipe.vae.config.scaling_factor
|
| 938 |
-
|
| 939 |
-
trunk_size = 1
|
| 940 |
-
if trunk_size < z.shape[0]:
|
| 941 |
-
out = torch.cat(
|
| 942 |
-
[self.sd_pipe.vae.decode(xx).sample for xx in z.split(trunk_size, 0)], dim=0,
|
| 943 |
-
)
|
| 944 |
-
else:
|
| 945 |
-
out = self.sd_pipe.vae.decode(z).sample
|
| 946 |
-
if clamp:
|
| 947 |
-
out = out.clamp(-1.0, 1.0)
|
| 948 |
-
return out
|
| 949 |
-
|
| 950 |
-
def get_loss_from_discrimnator(self, logits_fake):
|
| 951 |
-
if not (isinstance(logits_fake, list) or isinstance(logits_fake, tuple)):
|
| 952 |
-
g_loss = -torch.mean(logits_fake, dim=list(range(1, logits_fake.ndim)))
|
| 953 |
-
else:
|
| 954 |
-
g_loss = -torch.mean(logits_fake[0], dim=list(range(1, logits_fake[0].ndim)))
|
| 955 |
-
for current_logits in logits_fake[1:]:
|
| 956 |
-
g_loss += -torch.mean(current_logits, dim=list(range(1, current_logits.ndim)))
|
| 957 |
-
g_loss /= len(logits_fake)
|
| 958 |
-
|
| 959 |
-
return g_loss
|
| 960 |
-
|
| 961 |
-
def training_step(self, data):
|
| 962 |
-
current_bs = data['gt'].shape[0]
|
| 963 |
-
micro_bs = self.configs.train.microbatch
|
| 964 |
-
num_grad_accumulate = math.ceil(current_bs / micro_bs)
|
| 965 |
-
|
| 966 |
-
# grad zero
|
| 967 |
-
self.model.zero_grad()
|
| 968 |
-
|
| 969 |
-
# update generator
|
| 970 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 971 |
-
self.freeze_model(self.discriminator) # freeze discriminator
|
| 972 |
-
z0_pred_list = []
|
| 973 |
-
tt_list = []
|
| 974 |
-
prompt_embeds_list = []
|
| 975 |
-
for jj in range(0, current_bs, micro_bs):
|
| 976 |
-
micro_data = {key:value[jj:jj+micro_bs] for key, value in data.items()}
|
| 977 |
-
last_batch = (jj+micro_bs >= current_bs)
|
| 978 |
-
if last_batch or self.num_gpus <= 1:
|
| 979 |
-
losses, z0_pred, zt_noisy, tt = self.backward_step(micro_data, num_grad_accumulate)
|
| 980 |
-
else:
|
| 981 |
-
with self.model.no_sync():
|
| 982 |
-
losses, z0_pred, zt_noisy, tt = self.backward_step(micro_data, num_grad_accumulate)
|
| 983 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 984 |
-
z0_pred_list.append(z0_pred.detach())
|
| 985 |
-
tt_list.append(tt)
|
| 986 |
-
prompt_embeds_list.append(self.prompt_embeds.detach())
|
| 987 |
-
|
| 988 |
-
if self.configs.train.use_amp:
|
| 989 |
-
self.amp_scaler.step(self.optimizer)
|
| 990 |
-
self.amp_scaler.update()
|
| 991 |
-
else:
|
| 992 |
-
self.optimizer.step()
|
| 993 |
-
|
| 994 |
-
# update discriminator
|
| 995 |
-
if (self.configs.train.loss_coef.get('ldis', 0) > 0 and
|
| 996 |
-
(self.current_iters < self.configs.train.dis_init_iterations
|
| 997 |
-
or self.current_iters % self.configs.train.dis_update_freq == 0)
|
| 998 |
-
):
|
| 999 |
-
# grad zero
|
| 1000 |
-
self.unfreeze_model(self.discriminator) # update discriminator
|
| 1001 |
-
self.discriminator.zero_grad()
|
| 1002 |
-
for ii, jj in enumerate(range(0, current_bs, micro_bs)):
|
| 1003 |
-
micro_data = {key:value[jj:jj+micro_bs] for key, value in data.items()}
|
| 1004 |
-
last_batch = (jj+micro_bs >= current_bs)
|
| 1005 |
-
target = micro_data['gt_latent']
|
| 1006 |
-
inputs = z0_pred_list[ii]
|
| 1007 |
-
if last_batch or self.num_gpus <= 1:
|
| 1008 |
-
logits = self.dis_backward_step(target, inputs, tt_list[ii], prompt_embeds_list[ii])
|
| 1009 |
-
else:
|
| 1010 |
-
with self.discriminator.no_sync():
|
| 1011 |
-
logits = self.dis_backward_step(
|
| 1012 |
-
target, inputs, tt_list[ii], prompt_embeds_list[ii]
|
| 1013 |
-
)
|
| 1014 |
-
|
| 1015 |
-
# make logging
|
| 1016 |
-
if self.current_iters % self.configs.train.dis_update_freq == 0 and self.rank == 0:
|
| 1017 |
-
ndim = logits[0].ndim
|
| 1018 |
-
losses['real'] = logits[0].detach().mean(dim=list(range(1, ndim)))
|
| 1019 |
-
losses['fake'] = logits[1].detach().mean(dim=list(range(1, ndim)))
|
| 1020 |
-
|
| 1021 |
-
if self.configs.train.use_amp:
|
| 1022 |
-
self.amp_scaler_dis.step(self.optimizer_dis)
|
| 1023 |
-
self.amp_scaler_dis.update()
|
| 1024 |
-
else:
|
| 1025 |
-
self.optimizer_dis.step()
|
| 1026 |
-
|
| 1027 |
-
# make logging
|
| 1028 |
-
if self.rank == 0:
|
| 1029 |
-
self.log_step_train(
|
| 1030 |
-
losses, tt, micro_data, z0_pred, zt_noisy, z0_gt=micro_data['gt_latent'],
|
| 1031 |
-
)
|
| 1032 |
-
|
| 1033 |
-
@torch.no_grad()
|
| 1034 |
-
def log_step_train(self, losses, tt, micro_data, z0_pred, zt_noisy, z0_gt=None, phase='train'):
|
| 1035 |
-
'''
|
| 1036 |
-
param losses: a dict recording the loss informations
|
| 1037 |
-
'''
|
| 1038 |
-
'''
|
| 1039 |
-
param loss: a dict recording the loss informations
|
| 1040 |
-
param micro_data: batch data
|
| 1041 |
-
param tt: 1-D tensor, time steps
|
| 1042 |
-
'''
|
| 1043 |
-
if hasattr(self.configs.train, 'timesteps'):
|
| 1044 |
-
if len(self.configs.train.timesteps) < 3:
|
| 1045 |
-
record_steps = sorted(self.configs.train.timesteps)
|
| 1046 |
-
else:
|
| 1047 |
-
record_steps = [min(self.configs.train.timesteps),
|
| 1048 |
-
max(self.configs.train.timesteps)]
|
| 1049 |
-
else:
|
| 1050 |
-
max_inference_steps = self.configs.train.max_inference_steps
|
| 1051 |
-
record_steps = [1, max_inference_steps//2, max_inference_steps]
|
| 1052 |
-
if ((self.current_iters // self.configs.train.dis_update_freq) %
|
| 1053 |
-
(self.configs.train.log_freq[0] // self.configs.train.dis_update_freq) == 1):
|
| 1054 |
-
self.loss_mean = {key:torch.zeros(size=(len(record_steps),), dtype=torch.float64)
|
| 1055 |
-
for key in losses.keys() if key not in ['real', 'fake']}
|
| 1056 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 1057 |
-
self.logit_mean = {key:torch.zeros(size=(len(record_steps),), dtype=torch.float64)
|
| 1058 |
-
for key in ['real', 'fake']}
|
| 1059 |
-
self.loss_count = torch.zeros(size=(len(record_steps),), dtype=torch.float64)
|
| 1060 |
-
for jj in range(len(record_steps)):
|
| 1061 |
-
for key, value in losses.items():
|
| 1062 |
-
index = record_steps[jj] - 1
|
| 1063 |
-
mask = torch.where(tt == index, torch.ones_like(tt), torch.zeros_like(tt))
|
| 1064 |
-
assert value.shape == mask.shape
|
| 1065 |
-
current_loss = torch.sum(value.detach() * mask)
|
| 1066 |
-
if key in ['real', 'fake']:
|
| 1067 |
-
self.logit_mean[key][jj] += current_loss.item()
|
| 1068 |
-
else:
|
| 1069 |
-
self.loss_mean[key][jj] += current_loss.item()
|
| 1070 |
-
self.loss_count[jj] += mask.sum().item()
|
| 1071 |
-
|
| 1072 |
-
if ((self.current_iters // self.configs.train.dis_update_freq) %
|
| 1073 |
-
(self.configs.train.log_freq[0] // self.configs.train.dis_update_freq) == 0):
|
| 1074 |
-
if torch.any(self.loss_count == 0):
|
| 1075 |
-
self.loss_count += 1e-4
|
| 1076 |
-
for key in losses.keys():
|
| 1077 |
-
if key in ['real', 'fake']:
|
| 1078 |
-
self.logit_mean[key] /= self.loss_count
|
| 1079 |
-
else:
|
| 1080 |
-
self.loss_mean[key] /= self.loss_count
|
| 1081 |
-
log_str = f"Train: {self.current_iters:06d}/{self.configs.train.iterations:06d}, "
|
| 1082 |
-
valid_keys = sorted([key for key in losses.keys() if key not in ['loss', 'real', 'fake']])
|
| 1083 |
-
for ii, key in enumerate(valid_keys):
|
| 1084 |
-
if ii == 0:
|
| 1085 |
-
log_str += f"{key}"
|
| 1086 |
-
else:
|
| 1087 |
-
log_str += f"/{key}"
|
| 1088 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 1089 |
-
log_str += "/real/fake:"
|
| 1090 |
-
else:
|
| 1091 |
-
log_str += ":"
|
| 1092 |
-
for jj, current_record in enumerate(record_steps):
|
| 1093 |
-
for ii, key in enumerate(valid_keys):
|
| 1094 |
-
if ii == 0:
|
| 1095 |
-
if key in ['dis', 'ldis']:
|
| 1096 |
-
log_str += 't({:d}):{:+6.4f}'.format(
|
| 1097 |
-
current_record,
|
| 1098 |
-
self.loss_mean[key][jj].item(),
|
| 1099 |
-
)
|
| 1100 |
-
elif key in ['lpips', 'ldif']:
|
| 1101 |
-
log_str += 't({:d}):{:4.2f}'.format(
|
| 1102 |
-
current_record,
|
| 1103 |
-
self.loss_mean[key][jj].item(),
|
| 1104 |
-
)
|
| 1105 |
-
elif key == 'llpips':
|
| 1106 |
-
log_str += 't({:d}):{:5.3f}'.format(
|
| 1107 |
-
current_record,
|
| 1108 |
-
self.loss_mean[key][jj].item(),
|
| 1109 |
-
)
|
| 1110 |
-
else:
|
| 1111 |
-
log_str += 't({:d}):{:.1e}'.format(
|
| 1112 |
-
current_record,
|
| 1113 |
-
self.loss_mean[key][jj].item(),
|
| 1114 |
-
)
|
| 1115 |
-
else:
|
| 1116 |
-
if key in ['dis', 'ldis']:
|
| 1117 |
-
log_str += f"/{self.loss_mean[key][jj].item():+6.4f}"
|
| 1118 |
-
elif key in ['lpips', 'ldif']:
|
| 1119 |
-
log_str += f"/{self.loss_mean[key][jj].item():4.2f}"
|
| 1120 |
-
elif key == 'llpips':
|
| 1121 |
-
log_str += f"/{self.loss_mean[key][jj].item():5.3f}"
|
| 1122 |
-
else:
|
| 1123 |
-
log_str += f"/{self.loss_mean[key][jj].item():.1e}"
|
| 1124 |
-
if self.configs.train.loss_coef.get('ldis', 0) > 0:
|
| 1125 |
-
log_str += f"/{self.logit_mean['real'][jj].item():+4.2f}"
|
| 1126 |
-
log_str += f"/{self.logit_mean['fake'][jj].item():+4.2f}, "
|
| 1127 |
-
else:
|
| 1128 |
-
log_str += f", "
|
| 1129 |
-
log_str += 'lr:{:.1e}'.format(self.optimizer.param_groups[0]['lr'])
|
| 1130 |
-
self.logger.info(log_str)
|
| 1131 |
-
self.logging_metric(self.loss_mean, tag='Loss', phase=phase, add_global_step=True)
|
| 1132 |
-
if ((self.current_iters // self.configs.train.dis_update_freq) %
|
| 1133 |
-
(self.configs.train.log_freq[1] // self.configs.train.dis_update_freq) == 0):
|
| 1134 |
-
if zt_noisy is not None:
|
| 1135 |
-
xt_pred = self.decode_first_stage(zt_noisy.detach())
|
| 1136 |
-
self.logging_image(xt_pred, tag='xt-noisy', phase=phase, add_global_step=False)
|
| 1137 |
-
if z0_pred is not None:
|
| 1138 |
-
x0_pred = self.decode_first_stage(z0_pred.detach())
|
| 1139 |
-
self.logging_image(x0_pred, tag='x0-pred', phase=phase, add_global_step=False)
|
| 1140 |
-
if z0_gt is not None:
|
| 1141 |
-
x0_recon = self.decode_first_stage(z0_gt.detach())
|
| 1142 |
-
self.logging_image(x0_recon, tag='x0-recons', phase=phase, add_global_step=False)
|
| 1143 |
-
if 'txt' in micro_data:
|
| 1144 |
-
self.logging_text(micro_data['txt'], phase=phase)
|
| 1145 |
-
self.logging_image(micro_data['lq'], tag='LQ', phase=phase, add_global_step=False)
|
| 1146 |
-
self.logging_image(micro_data['gt'], tag='GT', phase=phase, add_global_step=True)
|
| 1147 |
-
|
| 1148 |
-
if ((self.current_iters // self.configs.train.dis_update_freq) %
|
| 1149 |
-
(self.configs.train.save_freq // self.configs.train.dis_update_freq) == 1):
|
| 1150 |
-
self.tic = time.time()
|
| 1151 |
-
if ((self.current_iters // self.configs.train.dis_update_freq) %
|
| 1152 |
-
(self.configs.train.save_freq // self.configs.train.dis_update_freq) == 0):
|
| 1153 |
-
self.toc = time.time()
|
| 1154 |
-
elaplsed = (self.toc - self.tic)
|
| 1155 |
-
self.logger.info(f"Elapsed time: {elaplsed:.2f}s")
|
| 1156 |
-
self.logger.info("="*100)
|
| 1157 |
-
|
| 1158 |
-
@torch.no_grad()
|
| 1159 |
-
def validation(self, phase='val'):
|
| 1160 |
-
torch.cuda.empty_cache()
|
| 1161 |
-
if not (self.configs.validate.use_ema and hasattr(self.configs.train, 'ema_rate')):
|
| 1162 |
-
self.model.eval()
|
| 1163 |
-
|
| 1164 |
-
if self.configs.train.start_mode:
|
| 1165 |
-
start_noise_predictor = self.ema_model if self.configs.validate.use_ema else self.model
|
| 1166 |
-
intermediate_noise_predictor = None
|
| 1167 |
-
else:
|
| 1168 |
-
start_noise_predictor = self.start_model
|
| 1169 |
-
intermediate_noise_predictor = self.ema_model if self.configs.validate.use_ema else self.model
|
| 1170 |
-
num_iters_epoch = math.ceil(len(self.datasets[phase]) / self.configs.validate.batch)
|
| 1171 |
-
mean_psnr = mean_lpips = 0
|
| 1172 |
-
for jj, data in enumerate(self.dataloaders[phase]):
|
| 1173 |
-
data = self.prepare_data(data, phase='val')
|
| 1174 |
-
with torch.amp.autocast('cuda'):
|
| 1175 |
-
xt_progressive, x0_progressive = self.sample(
|
| 1176 |
-
image_lq=data['lq'],
|
| 1177 |
-
prompt=[_positive,]*data['lq'].shape[0],
|
| 1178 |
-
target_size=tuple(data['gt'].shape[-2:]),
|
| 1179 |
-
start_noise_predictor=start_noise_predictor,
|
| 1180 |
-
intermediate_noise_predictor=intermediate_noise_predictor,
|
| 1181 |
-
)
|
| 1182 |
-
x0 = xt_progressive[-1]
|
| 1183 |
-
num_inference_steps = len(xt_progressive)
|
| 1184 |
-
|
| 1185 |
-
if 'gt' in data:
|
| 1186 |
-
if not hasattr(self, 'psnr_metric'):
|
| 1187 |
-
self.psnr_metric = pyiqa.create_metric(
|
| 1188 |
-
'psnr',
|
| 1189 |
-
test_y_channel=self.configs.train.get('val_y_channel', True),
|
| 1190 |
-
color_space='ycbcr',
|
| 1191 |
-
device=torch.device("cuda"),
|
| 1192 |
-
)
|
| 1193 |
-
if not hasattr(self, 'lpips_metric'):
|
| 1194 |
-
self.lpips_metric = pyiqa.create_metric(
|
| 1195 |
-
'lpips-vgg',
|
| 1196 |
-
device=torch.device("cuda"),
|
| 1197 |
-
as_loss=False,
|
| 1198 |
-
)
|
| 1199 |
-
x0_normalize = util_image.normalize_th(x0, mean=0.5, std=0.5, reverse=True)
|
| 1200 |
-
mean_psnr += self.psnr_metric(x0_normalize, data['gt']).sum().item()
|
| 1201 |
-
with torch.amp.autocast('cuda'), torch.no_grad():
|
| 1202 |
-
mean_lpips += self.lpips_metric(x0_normalize, data['gt']).sum().item()
|
| 1203 |
-
|
| 1204 |
-
if (jj + 1) % self.configs.validate.log_freq == 0:
|
| 1205 |
-
self.logger.info(f'Validation: {jj+1:02d}/{num_iters_epoch:02d}...')
|
| 1206 |
-
|
| 1207 |
-
self.logging_image(data['gt'], tag='GT', phase=phase, add_global_step=False)
|
| 1208 |
-
xt_progressive = rearrange(torch.cat(xt_progressive, dim=1), 'b (k c) h w -> (b k) c h w', c=3)
|
| 1209 |
-
self.logging_image(
|
| 1210 |
-
xt_progressive,
|
| 1211 |
-
tag='sample-progress',
|
| 1212 |
-
phase=phase,
|
| 1213 |
-
add_global_step=False,
|
| 1214 |
-
nrow=num_inference_steps,
|
| 1215 |
-
)
|
| 1216 |
-
x0_progressive = rearrange(torch.cat(x0_progressive, dim=1), 'b (k c) h w -> (b k) c h w', c=3)
|
| 1217 |
-
self.logging_image(
|
| 1218 |
-
x0_progressive,
|
| 1219 |
-
tag='x0-progress',
|
| 1220 |
-
phase=phase,
|
| 1221 |
-
add_global_step=False,
|
| 1222 |
-
nrow=num_inference_steps,
|
| 1223 |
-
)
|
| 1224 |
-
self.logging_image(data['lq'], tag='LQ', phase=phase, add_global_step=True)
|
| 1225 |
-
|
| 1226 |
-
if 'gt' in data:
|
| 1227 |
-
mean_psnr /= len(self.datasets[phase])
|
| 1228 |
-
mean_lpips /= len(self.datasets[phase])
|
| 1229 |
-
self.logger.info(f'Validation Metric: PSNR={mean_psnr:5.2f}, LPIPS={mean_lpips:6.4f}...')
|
| 1230 |
-
self.logging_metric(mean_psnr, tag='PSNR', phase=phase, add_global_step=False)
|
| 1231 |
-
self.logging_metric(mean_lpips, tag='LPIPS', phase=phase, add_global_step=True)
|
| 1232 |
-
|
| 1233 |
-
self.logger.info("="*100)
|
| 1234 |
-
|
| 1235 |
-
if not (self.configs.validate.use_ema and hasattr(self.configs.train, 'ema_rate')):
|
| 1236 |
-
self.model.train()
|
| 1237 |
-
torch.cuda.empty_cache()
|
| 1238 |
-
|
| 1239 |
-
def backward_step(self, micro_data, num_grad_accumulate):
|
| 1240 |
-
loss_coef = self.configs.train.loss_coef
|
| 1241 |
-
|
| 1242 |
-
losses = {}
|
| 1243 |
-
z0_gt = micro_data['gt_latent']
|
| 1244 |
-
tt = torch.tensor(
|
| 1245 |
-
random.choices(self.configs.train.timesteps, k=z0_gt.shape[0]),
|
| 1246 |
-
dtype=torch.int64,
|
| 1247 |
-
device=f"cuda:{self.rank}",
|
| 1248 |
-
) - 1
|
| 1249 |
-
|
| 1250 |
-
with torch.autocast(device_type="cuda", enabled=self.configs.train.use_amp):
|
| 1251 |
-
model_pred = self.model(
|
| 1252 |
-
micro_data['lq'], tt, sample_posterior=False, center_input_sample=True,
|
| 1253 |
-
)
|
| 1254 |
-
z0_pred, zt_noisy_pred, z0_lq = self.sd_forward_step(
|
| 1255 |
-
prompt=micro_data['txt'],
|
| 1256 |
-
latents_hq=micro_data['gt_latent'],
|
| 1257 |
-
image_lq=micro_data['lq'],
|
| 1258 |
-
image_hq=micro_data['gt'],
|
| 1259 |
-
model_pred=model_pred,
|
| 1260 |
-
timesteps=tt,
|
| 1261 |
-
)
|
| 1262 |
-
# diffusion loss
|
| 1263 |
-
if loss_coef.get('ldif', 0) > 0:
|
| 1264 |
-
if self.configs.train.loss_type == 'L2':
|
| 1265 |
-
ldif_loss = F.mse_loss(z0_pred, z0_gt, reduction='none')
|
| 1266 |
-
elif self.configs.train.loss_type == 'L1':
|
| 1267 |
-
ldif_loss = F.l1_loss(z0_pred, z0_gt, reduction='none')
|
| 1268 |
-
else:
|
| 1269 |
-
raise TypeError(f"Unsupported Loss type for Diffusion: {self.configs.train.loss_type}")
|
| 1270 |
-
ldif_loss = torch.mean(ldif_loss, dim=list(range(1, z0_gt.ndim)))
|
| 1271 |
-
losses['ldif'] = ldif_loss * loss_coef['ldif']
|
| 1272 |
-
# Gaussian constraints
|
| 1273 |
-
if loss_coef.get('kl', 0) > 0:
|
| 1274 |
-
losses['kl'] = model_pred.kl() * loss_coef['kl']
|
| 1275 |
-
if loss_coef.get('pkl', 0) > 0:
|
| 1276 |
-
losses['pkl'] = model_pred.partial_kl() * loss_coef['pkl']
|
| 1277 |
-
if loss_coef.get('rkl', 0) > 0:
|
| 1278 |
-
other = Box(
|
| 1279 |
-
{'mean': z0_gt-z0_lq,
|
| 1280 |
-
'var':torch.ones_like(z0_gt),
|
| 1281 |
-
'logvar':torch.zeros_like(z0_gt)}
|
| 1282 |
-
)
|
| 1283 |
-
losses['rkl'] = model_pred.kl(other) * loss_coef['rkl']
|
| 1284 |
-
# discriminator loss
|
| 1285 |
-
if loss_coef.get('ldis', 0) > 0:
|
| 1286 |
-
if self.current_iters > self.configs.train.dis_init_iterations:
|
| 1287 |
-
logits_fake = self.discriminator(
|
| 1288 |
-
torch.clamp(z0_pred, min=_Latent_bound['min'], max=_Latent_bound['max']),
|
| 1289 |
-
timestep=tt,
|
| 1290 |
-
encoder_hidden_states=self.prompt_embeds,
|
| 1291 |
-
)
|
| 1292 |
-
losses['ldis'] = self.get_loss_from_discrimnator(logits_fake) * loss_coef['ldis']
|
| 1293 |
-
else:
|
| 1294 |
-
losses['ldis'] = torch.zeros((z0_gt.shape[0], ), dtype=torch.float32).cuda()
|
| 1295 |
-
# perceptual loss
|
| 1296 |
-
if loss_coef.get('llpips', 0) > 0:
|
| 1297 |
-
losses['llpips'] = self.llpips_loss(z0_pred, z0_gt).view(-1) * loss_coef['llpips']
|
| 1298 |
-
|
| 1299 |
-
for key in ['ldif', 'kl', 'rkl', 'pkl', 'ldis', 'llpips']:
|
| 1300 |
-
if loss_coef.get(key, 0) > 0:
|
| 1301 |
-
if not 'loss' in losses:
|
| 1302 |
-
losses['loss'] = losses[key]
|
| 1303 |
-
else:
|
| 1304 |
-
losses['loss'] = losses['loss'] + losses[key]
|
| 1305 |
-
loss = losses['loss'].mean() / num_grad_accumulate
|
| 1306 |
-
|
| 1307 |
-
if self.amp_scaler is None:
|
| 1308 |
-
loss.backward()
|
| 1309 |
-
else:
|
| 1310 |
-
self.amp_scaler.scale(loss).backward()
|
| 1311 |
-
|
| 1312 |
-
return losses, z0_pred, zt_noisy_pred, tt
|
| 1313 |
-
|
| 1314 |
-
def dis_backward_step(self, target, inputs, tt, prompt_embeds):
|
| 1315 |
-
with torch.autocast(device_type="cuda", enabled=self.configs.train.use_amp):
|
| 1316 |
-
logits_real = self.discriminator(target, tt, prompt_embeds)
|
| 1317 |
-
inputs = inputs.clamp(min=_Latent_bound['min'], max=_Latent_bound['max'])
|
| 1318 |
-
logits_fake = self.discriminator(inputs, tt, prompt_embeds)
|
| 1319 |
-
|
| 1320 |
-
loss = hinge_d_loss(logits_real, logits_fake).mean()
|
| 1321 |
-
|
| 1322 |
-
if self.amp_scaler_dis is None:
|
| 1323 |
-
loss.backward()
|
| 1324 |
-
else:
|
| 1325 |
-
self.amp_scaler_dis.scale(loss).backward()
|
| 1326 |
-
|
| 1327 |
-
return logits_real[-1], logits_fake[-1]
|
| 1328 |
-
|
| 1329 |
-
def scale_sd_input(
|
| 1330 |
-
self,
|
| 1331 |
-
x:torch.Tensor,
|
| 1332 |
-
sigmas: torch.Tensor = None,
|
| 1333 |
-
timestep: torch.Tensor = None,
|
| 1334 |
-
) :
|
| 1335 |
-
if sigmas is None:
|
| 1336 |
-
if not self.sd_pipe.scheduler.sigmas.numel() == (self.configs.sd_pipe.num_train_steps + 1):
|
| 1337 |
-
self.sd_pipe.scheduler = EulerDiscreteScheduler.from_pipe(
|
| 1338 |
-
self.configs.sd_pipe.params.pretrained_model_name_or_path,
|
| 1339 |
-
cache_dir=self.configs.sd_pipe.params.cache_dir,
|
| 1340 |
-
subfolder='scheduler',
|
| 1341 |
-
)
|
| 1342 |
-
assert self.sd_pipe.scheduler.sigmas.numel() == (self.configs.sd_pipe.num_train_steps + 1)
|
| 1343 |
-
sigmas = self.sd_pipe.scheduler.sigmas.flip(0).to(x.device)[timestep] # (b,)
|
| 1344 |
-
sigmas = append_dims(sigmas, x.ndim)
|
| 1345 |
-
|
| 1346 |
-
if sigmas.ndim < x.ndim:
|
| 1347 |
-
sigmas = append_dims(sigmas, x.ndim)
|
| 1348 |
-
out = x / ((sigmas**2 + 1) ** 0.5)
|
| 1349 |
-
return out
|
| 1350 |
-
|
| 1351 |
-
def prepare_lq_latents(
|
| 1352 |
-
self,
|
| 1353 |
-
image_lq: torch.Tensor,
|
| 1354 |
-
timestep: torch.Tensor,
|
| 1355 |
-
height: int = 512,
|
| 1356 |
-
width: int = 512,
|
| 1357 |
-
start_noise_predictor: torch.nn.Module = None,
|
| 1358 |
-
):
|
| 1359 |
-
"""
|
| 1360 |
-
Input:
|
| 1361 |
-
image_lq: low-quality image, torch.Tensor, range in [0, 1]
|
| 1362 |
-
hight, width: resolution for high-quality image
|
| 1363 |
-
|
| 1364 |
-
"""
|
| 1365 |
-
image_lq_up = F.interpolate(image_lq, size=(height, width), mode='bicubic')
|
| 1366 |
-
init_latents = self.encode_first_stage(
|
| 1367 |
-
image_lq_up, deterministic=False, center_input_sample=True,
|
| 1368 |
-
)
|
| 1369 |
-
|
| 1370 |
-
if start_noise_predictor is None:
|
| 1371 |
-
model_pred = None
|
| 1372 |
-
else:
|
| 1373 |
-
model_pred = start_noise_predictor(
|
| 1374 |
-
image_lq, timestep, sample_posterior=False, center_input_sample=True,
|
| 1375 |
-
)
|
| 1376 |
-
|
| 1377 |
-
# get latents
|
| 1378 |
-
sigmas = self.sigmas_cache[timestep]
|
| 1379 |
-
sigmas = append_dims(sigmas, init_latents.ndim)
|
| 1380 |
-
latents = self.add_noise(init_latents, sigmas, model_pred)
|
| 1381 |
-
|
| 1382 |
-
return latents
|
| 1383 |
-
|
| 1384 |
-
def add_noise(self, latents, sigmas, model_pred=None):
|
| 1385 |
-
if sigmas.ndim < latents.ndim:
|
| 1386 |
-
sigmas = append_dims(sigmas, latents.ndim)
|
| 1387 |
-
|
| 1388 |
-
if model_pred is None:
|
| 1389 |
-
noise = torch.randn_like(latents)
|
| 1390 |
-
zt_noisy = latents + sigmas * noise
|
| 1391 |
-
else:
|
| 1392 |
-
if self.configs.train.loss_coef.get('rkl', 0) > 0:
|
| 1393 |
-
mean, std = model_pred.mean, model_pred.std
|
| 1394 |
-
zt_noisy = latents + mean + sigmas * std * torch.randn_like(latents)
|
| 1395 |
-
else:
|
| 1396 |
-
zt_noisy = latents + sigmas * model_pred.sample()
|
| 1397 |
-
|
| 1398 |
-
return zt_noisy
|
| 1399 |
-
|
| 1400 |
-
def retrieve_timesteps(self):
|
| 1401 |
-
device=torch.device(f"cuda:{self.rank}")
|
| 1402 |
-
|
| 1403 |
-
num_inference_steps = self.configs.train.get('num_inference_steps', 5)
|
| 1404 |
-
timesteps = np.linspace(
|
| 1405 |
-
max(self.configs.train.timesteps), 0, num_inference_steps,
|
| 1406 |
-
endpoint=False, dtype=np.int64,
|
| 1407 |
-
) - 1
|
| 1408 |
-
timesteps = torch.from_numpy(timesteps).to(device)
|
| 1409 |
-
self.sd_pipe.scheduler.timesteps = timesteps
|
| 1410 |
-
|
| 1411 |
-
sigmas = self.sigmas_cache[timesteps.long()]
|
| 1412 |
-
sigma_last = torch.tensor([0,], dtype=torch.float32).to(device=sigmas.device)
|
| 1413 |
-
sigmas = torch.cat([sigmas, sigma_last]).type(torch.float32)
|
| 1414 |
-
self.sd_pipe.scheduler.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication
|
| 1415 |
-
|
| 1416 |
-
self.sd_pipe.scheduler._step_index = None
|
| 1417 |
-
self.sd_pipe.scheduler._begin_index = None
|
| 1418 |
-
|
| 1419 |
-
return self.sd_pipe.scheduler.timesteps, num_inference_steps
|
| 1420 |
-
|
| 1421 |
-
class TrainerSDTurboSR(TrainerBaseSR):
|
| 1422 |
-
def sd_forward_step(
|
| 1423 |
-
self,
|
| 1424 |
-
prompt: Union[str, List[str]] = None,
|
| 1425 |
-
latents_hq: Optional[torch.Tensor] = None,
|
| 1426 |
-
image_lq: torch.Tensor = None,
|
| 1427 |
-
image_hq: torch.Tensor = None,
|
| 1428 |
-
model_pred: DiagonalGaussianDistribution = None,
|
| 1429 |
-
timesteps: List[int] = None,
|
| 1430 |
-
**kwargs,
|
| 1431 |
-
):
|
| 1432 |
-
r"""
|
| 1433 |
-
Function invoked when calling the pipeline for generation.
|
| 1434 |
-
|
| 1435 |
-
Args:
|
| 1436 |
-
prompt (`str` or `List[str]`, *optional*):
|
| 1437 |
-
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
| 1438 |
-
instead.
|
| 1439 |
-
image_lq (`torch.Tensor`): The low-quality image(s) for enhancement, range in [0, 1].
|
| 1440 |
-
image_hq (`torch.Tensor`): The high-quality image(s) for enhancement, range in [0, 1].
|
| 1441 |
-
noise_pred (`torch.Tensor`): Predicted noise by the noise prediction model
|
| 1442 |
-
latents_hq (`torch.Tensor`, *optional*):
|
| 1443 |
-
Pre-generated high-quality latents, sampled from a Gaussian distribution, to be used as inputs for image
|
| 1444 |
-
generation. If not provided, a latents tensor will be generated by sampling using vae .
|
| 1445 |
-
timesteps (`List[int]`, *optional*):
|
| 1446 |
-
Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
|
| 1447 |
-
in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
|
| 1448 |
-
passed will be used. Must be in descending order.
|
| 1449 |
-
aesthetic_score (`float`, *optional*, defaults to 6.0):
|
| 1450 |
-
Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
|
| 1451 |
-
Part of SDXL's micro-conditioning as explained in section 2.2 of
|
| 1452 |
-
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
|
| 1453 |
-
negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
|
| 1454 |
-
Part of SDXL's micro-conditioning as explained in section 2.2 of
|
| 1455 |
-
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to
|
| 1456 |
-
simulate an aesthetic score of the generated image by influencing the negative text condition.
|
| 1457 |
-
"""
|
| 1458 |
-
device=torch.device(f"cuda:{self.rank}")
|
| 1459 |
-
# Encode input prompt
|
| 1460 |
-
prompt_embeds, negative_prompt_embeds = self.sd_pipe.encode_prompt(
|
| 1461 |
-
prompt=prompt,
|
| 1462 |
-
device=device,
|
| 1463 |
-
num_images_per_prompt=1,
|
| 1464 |
-
do_classifier_free_guidance=False,
|
| 1465 |
-
)
|
| 1466 |
-
self.prompt_embeds = prompt_embeds
|
| 1467 |
-
|
| 1468 |
-
# select the noise level, self.scheduler.sigmas, [1001,], descending
|
| 1469 |
-
if not hasattr(self, 'sigmas_cache'):
|
| 1470 |
-
assert self.sd_pipe.scheduler.sigmas.numel() == (self.configs.sd_pipe.num_train_steps + 1)
|
| 1471 |
-
self.sigmas_cache = self.sd_pipe.scheduler.sigmas.flip(0)[1:].to(device) #ascending,1000
|
| 1472 |
-
sigmas = self.sigmas_cache[timesteps] # (b,)
|
| 1473 |
-
|
| 1474 |
-
# Prepare input for SD
|
| 1475 |
-
height, width = image_hq.shape[-2:]
|
| 1476 |
-
if self.configs.train.start_mode:
|
| 1477 |
-
image_lq_up = F.interpolate(image_lq, size=(height, width), mode='bicubic')
|
| 1478 |
-
zt_clean = self.encode_first_stage(
|
| 1479 |
-
image_lq_up, center_input_sample=True,
|
| 1480 |
-
deterministic=self.configs.train.loss_coef.get('rkl', 0) > 0,
|
| 1481 |
-
)
|
| 1482 |
-
else:
|
| 1483 |
-
if latents_hq is None:
|
| 1484 |
-
zt_clean = self.encode_first_stage(
|
| 1485 |
-
image_hq, center_input_sample=True, deterministic=False,
|
| 1486 |
-
)
|
| 1487 |
-
else:
|
| 1488 |
-
zt_clean = latents_hq
|
| 1489 |
-
|
| 1490 |
-
sigmas = append_dims(sigmas, zt_clean.ndim)
|
| 1491 |
-
zt_noisy = self.add_noise(zt_clean, sigmas, model_pred)
|
| 1492 |
-
|
| 1493 |
-
prompt_embeds = prompt_embeds.to(device)
|
| 1494 |
-
|
| 1495 |
-
zt_noisy_scale = self.scale_sd_input(zt_noisy, sigmas)
|
| 1496 |
-
eps_pred = self.sd_pipe.unet(
|
| 1497 |
-
zt_noisy_scale,
|
| 1498 |
-
timesteps,
|
| 1499 |
-
encoder_hidden_states=prompt_embeds,
|
| 1500 |
-
timestep_cond=None,
|
| 1501 |
-
cross_attention_kwargs=None,
|
| 1502 |
-
added_cond_kwargs=None,
|
| 1503 |
-
return_dict=False,
|
| 1504 |
-
)[0] # eps-mode for sdxl and sdxl-refiner
|
| 1505 |
-
|
| 1506 |
-
if self.configs.train.noise_detach:
|
| 1507 |
-
z0_pred = zt_noisy.detach() - sigmas * eps_pred
|
| 1508 |
-
else:
|
| 1509 |
-
z0_pred = zt_noisy - sigmas * eps_pred
|
| 1510 |
-
|
| 1511 |
-
return z0_pred, zt_noisy, zt_clean
|
| 1512 |
-
|
| 1513 |
-
@torch.no_grad()
|
| 1514 |
-
def sample(
|
| 1515 |
-
self,
|
| 1516 |
-
image_lq: torch.Tensor,
|
| 1517 |
-
prompt: Union[str, List[str]] = None,
|
| 1518 |
-
target_size: Tuple[int, int] = (1024, 1024),
|
| 1519 |
-
start_noise_predictor: torch.nn.Module = None,
|
| 1520 |
-
intermediate_noise_predictor: torch.nn.Module = None,
|
| 1521 |
-
**kwargs,
|
| 1522 |
-
):
|
| 1523 |
-
r"""
|
| 1524 |
-
Function invoked when calling the pipeline for generation.
|
| 1525 |
-
|
| 1526 |
-
Args:
|
| 1527 |
-
prompt (`str` or `List[str]`, *optional*):
|
| 1528 |
-
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
| 1529 |
-
instead.
|
| 1530 |
-
image_lq (`torch.Tensor` or `PIL.Image.Image` or `np.ndarray` or `List[torch.Tensor]` or `List[PIL.Image.Image]` or `List[np.ndarray]`):
|
| 1531 |
-
The image(s) to modify with the pipeline.
|
| 1532 |
-
target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
|
| 1533 |
-
The required height and width of the super-resolved image.
|
| 1534 |
-
strength (`float`, *optional*, defaults to 0.3):
|
| 1535 |
-
Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
|
| 1536 |
-
will be used as a starting point, adding more noise to it the larger the `strength`. The number of
|
| 1537 |
-
denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
|
| 1538 |
-
be maximum and the denoising process will run for the full number of iterations specified in
|
| 1539 |
-
`num_inference_steps`. A value of 1, therefore, essentially ignores `image`. Note that in the case of
|
| 1540 |
-
`denoising_start` being declared as an integer, the value of `strength` will be ignored.
|
| 1541 |
-
num_inference_steps (`int`, *optional*, defaults to 50):
|
| 1542 |
-
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
| 1543 |
-
expense of slower inference.
|
| 1544 |
-
timesteps (`List[int]`, *optional*):
|
| 1545 |
-
Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
|
| 1546 |
-
in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
|
| 1547 |
-
passed will be used. Must be in descending order.
|
| 1548 |
-
negative_prompt (`str` or `List[str]`, *optional*):
|
| 1549 |
-
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
| 1550 |
-
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
| 1551 |
-
less than `1`).
|
| 1552 |
-
"""
|
| 1553 |
-
device=torch.device(f"cuda:{self.rank}")
|
| 1554 |
-
batch_size = image_lq.shape[0]
|
| 1555 |
-
|
| 1556 |
-
# Encode input prompt
|
| 1557 |
-
prompt_embeds, negative_prompt_embeds = self.sd_pipe.encode_prompt(
|
| 1558 |
-
prompt=prompt,
|
| 1559 |
-
device=device,
|
| 1560 |
-
num_images_per_prompt=1,
|
| 1561 |
-
do_classifier_free_guidance=False,
|
| 1562 |
-
)
|
| 1563 |
-
|
| 1564 |
-
timesteps, num_inference_steps = self.retrieve_timesteps()
|
| 1565 |
-
latent_timestep = timesteps[:1].repeat(batch_size)
|
| 1566 |
-
|
| 1567 |
-
# Prepare latent variables
|
| 1568 |
-
height, width = target_size
|
| 1569 |
-
latents = self.prepare_lq_latents(image_lq, latent_timestep.long(), height, width, start_noise_predictor)
|
| 1570 |
-
|
| 1571 |
-
# Prepare extra step kwargs.
|
| 1572 |
-
extra_step_kwargs = self.sd_pipe.prepare_extra_step_kwargs(None, 0.0)
|
| 1573 |
-
|
| 1574 |
-
prompt_embeds = prompt_embeds.to(device)
|
| 1575 |
-
|
| 1576 |
-
x0_progressive = []
|
| 1577 |
-
images_progressive = []
|
| 1578 |
-
for i, t in enumerate(timesteps):
|
| 1579 |
-
latents_scaled = self.sd_pipe.scheduler.scale_model_input(latents, t)
|
| 1580 |
-
|
| 1581 |
-
# predict the noise residual
|
| 1582 |
-
eps_pred = self.sd_pipe.unet(
|
| 1583 |
-
latents_scaled,
|
| 1584 |
-
t,
|
| 1585 |
-
encoder_hidden_states=prompt_embeds,
|
| 1586 |
-
timestep_cond=None,
|
| 1587 |
-
added_cond_kwargs=None,
|
| 1588 |
-
return_dict=False,
|
| 1589 |
-
)[0]
|
| 1590 |
-
z0_pred = latents - self.sigmas_cache[t.long()] * eps_pred
|
| 1591 |
-
|
| 1592 |
-
# compute the previous noisy sample x_t -> x_t-1
|
| 1593 |
-
if intermediate_noise_predictor is not None and i + 1 < len(timesteps):
|
| 1594 |
-
t_next = timesteps[i+1]
|
| 1595 |
-
noise = intermediate_noise_predictor(image_lq, t_next, center_input_sample=True)
|
| 1596 |
-
else:
|
| 1597 |
-
noise = None
|
| 1598 |
-
extra_step_kwargs['noise'] = noise
|
| 1599 |
-
latents = self.sd_pipe.scheduler.step(eps_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
| 1600 |
-
|
| 1601 |
-
image = self.decode_first_stage(latents)
|
| 1602 |
-
images_progressive.append(image)
|
| 1603 |
-
|
| 1604 |
-
x0_pred = self.decode_first_stage(z0_pred)
|
| 1605 |
-
x0_progressive.append(x0_pred)
|
| 1606 |
-
|
| 1607 |
-
return images_progressive, x0_progressive
|
| 1608 |
-
|
| 1609 |
-
def my_worker_init_fn(worker_id):
|
| 1610 |
-
np.random.seed(np.random.get_state()[1][0] + worker_id)
|
| 1611 |
-
|
| 1612 |
-
def hinge_d_loss(
|
| 1613 |
-
logits_real: Union[torch.Tensor, List[torch.Tensor,]],
|
| 1614 |
-
logits_fake: Union[torch.Tensor, List[torch.Tensor,]],
|
| 1615 |
-
):
|
| 1616 |
-
def _hinge_d_loss(logits_real, logits_fake):
|
| 1617 |
-
loss_real = F.relu(1.0 - logits_real)
|
| 1618 |
-
loss_fake = F.relu(1.0 + logits_fake)
|
| 1619 |
-
d_loss = 0.5 * (loss_real + loss_fake)
|
| 1620 |
-
loss = d_loss.mean(dim=list(range(1, logits_real.ndim)))
|
| 1621 |
-
|
| 1622 |
-
return loss
|
| 1623 |
-
|
| 1624 |
-
if not (isinstance(logits_real, list) or isinstance(logits_real, tuple)):
|
| 1625 |
-
loss = _hinge_d_loss(logits_real, logits_fake)
|
| 1626 |
-
else:
|
| 1627 |
-
loss = _hinge_d_loss(logits_real[0], logits_fake[0])
|
| 1628 |
-
for xx, yy in zip(logits_real[1:], logits_fake[1:]):
|
| 1629 |
-
loss += _hinge_d_loss(xx, yy)
|
| 1630 |
-
|
| 1631 |
-
loss /= len(logits_real)
|
| 1632 |
-
|
| 1633 |
-
return loss
|
| 1634 |
-
|
| 1635 |
-
def get_torch_dtype(torch_dtype: str):
|
| 1636 |
-
if torch_dtype == 'torch.float16':
|
| 1637 |
-
return torch.float16
|
| 1638 |
-
elif torch_dtype == 'torch.bfloat16':
|
| 1639 |
-
return torch.bfloat16
|
| 1640 |
-
elif torch_dtype == 'torch.float32':
|
| 1641 |
-
return torch.float32
|
| 1642 |
-
else:
|
| 1643 |
-
raise ValueError(f'Unexpected torch dtype:{torch_dtype}')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|