그냥 공부용 정리
import importlib
from omegaconf import OmegaConf
Json, yaml 등 각종 설정 파일 읽기 및 생성, 변경 작업 지원
from pytorch_lightning import seed_everything
from ldm.data.base import Txt2ImgIterableBaseDataset
미리 정의해 둔 데이터셋 형식
class Txt2ImgIterableBaseDataset(IterableDataset):
'''
Define an interface to make the IterableDatasets for text2img data chainable
'''
def __init__(self, num_records=0, valid_ids=None, size=256):
super().__init__()
self.num_records = num_records
self.valid_ids = valid_ids
self.sample_ids = valid_ids
self.size = size
print(f'{self.__class__.__name__} dataset contains {self.__len__()} examples.')
def __len__(self):
return self.num_records
@abstractmethod
def __iter__(self):
pass
from ldm.util import instantiate_from_config
명시되어 있는 모듈과 피라미터로 인스턴스를 생성
def instantiate_from_config(config):
if not "target" in config:
if config == '__is_first_stage__':
return None
elif config == "__is_unconditional__":
return None
raise KeyError("Expected key `target` to instantiate.")
return get_obj_from_str(config["target"])(**config.get("params", dict()))
def get_obj_from_str(string, reload=False):
module, cls = string.rsplit(".", 1)
if reload:
module_imp = importlib.import_module(module)
importlib.reload(module_imp)
return getattr(importlib.import_module(module, package=None), cls)
예시 :
def nondefault_trainer_args(opt):
parser = argparse.ArgumentParser()
parser = Trainer.add_argparse_args(parser)
args = parser.parse_args([])
return sorted(k for k in vars(args) if getattr(opt, k) != getattr(args, k))
Trainer의 default 값이 opt와 다르면 해당 key를 반환
def worker_init_fn(_):
worker_info = torch.utils.data.get_worker_info()
dataset = worker_info.dataset
worker_id = worker_info.id
if isinstance(dataset, Txt2ImgIterableBaseDataset):
split_size = dataset.num_records // worker_info.num_workers
# reset num_records to the true number to retain reliable length information
dataset.sample_ids = dataset.valid_ids[worker_id * split_size:(worker_id + 1) * split_size]
current_id = np.random.choice(len(np.random.get_state()[1]), 1)
return np.random.seed(np.random.get_state()[1][current_id] + worker_id)
else:
return np.random.seed(np.random.get_state()[1][0] + worker_id)
데이터셋 초기화를 위한 정보 얻기.
Txt2ImgIterableBaseDataset일 경우 내부값까지 초기화.
class DataModuleFromConfig(pl.LightningDataModule):
def __init__(self, batch_size, train=None, validation=None, test=None, predict=None,
wrap=False, num_workers=None, shuffle_test_loader=False, use_worker_init_fn=False,
shuffle_val_dataloader=False):
super().__init__()
self.batch_size = batch_size
self.dataset_configs = dict()
self.num_workers = num_workers if num_workers is not None else batch_size * 2
self.use_worker_init_fn = use_worker_init_fn
if train is not None:
self.dataset_configs["train"] = train
self.train_dataloader = self._train_dataloader
if validation is not None:
...
if test is not None:
...
if predict is not None:
...
self.wrap = wrap
def prepare_data(self):
for data_cfg in self.dataset_configs.values():
instantiate_from_config(data_cfg)
def setup(self, stage=None):
self.datasets = dict(
(k, instantiate_from_config(self.dataset_configs[k]))
for k in self.dataset_configs)
if self.wrap:
for k in self.datasets:
self.datasets[k] = WrappedDataset(self.datasets[k])
def _train_dataloader(self):
is_iterable_dataset = isinstance(self.datasets['train'], Txt2ImgIterableBaseDataset)
if is_iterable_dataset or self.use_worker_init_fn:
init_fn = worker_init_fn
else:
init_fn = None
return DataLoader(self.datasets["train"], batch_size=self.batch_size,
num_workers=self.num_workers, shuffle=False if is_iterable_dataset else True,
worker_init_fn=init_fn)
def _val_dataloader(self, shuffle=False):
if isinstance(self.datasets['validation'], Txt2ImgIterableBaseDataset) or self.use_worker_init_fn:
init_fn = worker_init_fn
else:
init_fn = None
return DataLoader(self.datasets["validation"],
batch_size=self.batch_size,
num_workers=self.num_workers,
worker_init_fn=init_fn,
shuffle=shuffle)
def _test_dataloader(self, shuffle=False):
...
return DataLoader(...)
def _predict_dataloader(self, shuffle=False):
...
return DataLoader(...)
Pytorch lightning의 LightningDataModule로 Dataloader 정의
class SetupCallback(Callback):
def __init__(self, resume, now, logdir, ckptdir, cfgdir, config, lightning_config):
super().__init__()
self.resume = resume
self.now = now
self.logdir = logdir
self.ckptdir = ckptdir
self.cfgdir = cfgdir
self.config = config
self.lightning_config = lightning_config
def on_keyboard_interrupt(self, trainer, pl_module):
if trainer.global_rank == 0:
...
def on_pretrain_routine_start(self, trainer, pl_module):
if trainer.global_rank == 0:
...
각종 설정을 기록하고 저장하는 pytorch lightning callback.
class ImageLogger(Callback):
def __init__(self, batch_frequency, max_images, clamp=True, increase_log_steps=True,
rescale=True, disabled=False, log_on_batch_idx=False, log_first_step=False,
log_images_kwargs=None):
super().__init__()
self.rescale = rescale
self.batch_freq = batch_frequency
self.max_images = max_images
self.logger_log_images = {
pl.loggers.TestTubeLogger: self._testtube,
}
self.log_steps = [2 ** n for n in range(int(np.log2(self.batch_freq)) + 1)]
if not increase_log_steps:
self.log_steps = [self.batch_freq]
self.clamp = clamp
self.disabled = disabled
self.log_on_batch_idx = log_on_batch_idx
self.log_images_kwargs = log_images_kwargs if log_images_kwargs else {}
self.log_first_step = log_first_step
@rank_zero_only
#이미지 기록
def _testtube(self, pl_module, images, batch_idx, split):
for k in images:
grid = torchvision.utils.make_grid(images[k])
grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w
tag = f"{split}/{k}"
pl_module.logger.experiment.add_image(
tag, grid,
global_step=pl_module.global_step)
@rank_zero_only
#이미지 저장
def log_local(self, save_dir, split, images,
global_step, current_epoch, batch_idx):
root = os.path.join(save_dir, "images", split)
for k in images:
grid = torchvision.utils.make_grid(images[k], nrow=4)
if self.rescale:
grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w
grid = grid.transpose(0, 1).transpose(1, 2).squeeze(-1)
grid = grid.numpy()
grid = (grid * 255).astype(np.uint8)
filename = "{}_gs-{:06}_e-{:06}_b-{:06}.png".format(
k,
global_step,
current_epoch,
batch_idx)
path = os.path.join(root, filename)
os.makedirs(os.path.split(path)[0], exist_ok=True)
Image.fromarray(grid).save(path)
#eval -> 로그 기록, 저장 -> train
def log_img(self, pl_module, batch, batch_idx, split="train"):
check_idx = batch_idx if self.log_on_batch_idx else pl_module.global_step
if (self.check_frequency(check_idx) and # batch_idx % self.batch_freq == 0
hasattr(pl_module, "log_images") and
callable(pl_module.log_images) and
self.max_images > 0):
logger = type(pl_module.logger)
is_train = pl_module.training
if is_train:
pl_module.eval()
with torch.no_grad():
images = pl_module.log_images(batch, split=split, **self.log_images_kwargs)
for k in images:
N = min(images[k].shape[0], self.max_images)
images[k] = images[k][:N]
if isinstance(images[k], torch.Tensor):
images[k] = images[k].detach().cpu()
if self.clamp:
images[k] = torch.clamp(images[k], -1., 1.)
self.log_local(pl_module.logger.save_dir, split, images,
pl_module.global_step, pl_module.current_epoch, batch_idx)
logger_log_images = self.logger_log_images.get(logger, lambda *args, **kwargs: None)
logger_log_images(pl_module, images, pl_module.global_step, split)
if is_train:
pl_module.train()
def check_frequency(self, check_idx):
if ((check_idx % self.batch_freq) == 0 or (check_idx in self.log_steps)) and (
check_idx > 0 or self.log_first_step):
try:
self.log_steps.pop(0)
except IndexError as e:
print(e)
pass
return True
return False
#배치가 끝날때마다 로깅
def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx):
if not self.disabled and (pl_module.global_step > 0 or self.log_first_step):
self.log_img(pl_module, batch, batch_idx, split="train")
def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx):
if not self.disabled and pl_module.global_step > 0:
self.log_img(pl_module, batch, batch_idx, split="val")
if hasattr(pl_module, 'calibrate_grad_norm'):
if (pl_module.calibrate_grad_norm and batch_idx % 25 == 0) and batch_idx > 0:
self.log_gradients(trainer, pl_module, batch_idx=batch_idx)
opt, unknown = parser.parse_known_args()
def divein(*args, **kwargs):
if trainer.global_rank == 0:
import pudb;
pudb.set_trace()
pdb = 파이썬 디버거, pudb = 콘솔 기반 파이썬 디버거
try:
...
# run
if opt.train:
try:
trainer.fit(model, data)
except Exception:
melk()
raise
if not opt.no_test and not trainer.interrupted:
trainer.test(model, data)
except Exception:
if opt.debug and trainer.global_rank == 0:
try:
import pudb as debugger
except ImportError:
import pdb as debugger
debugger.post_mortem()
raise
학습 중 에러 발생 시 체크포인트 저장하고 디버깅
'코드 리뷰 > Diffusion' 카테고리의 다른 글
DiffStyler 코드 리뷰 (0) | 2023.01.16 |
---|---|
Paint by Example 코드 리뷰 (1) | 2023.01.15 |
DAAM 코드 리뷰 (0) | 2023.01.12 |
Latent Diffusion (0) | 2022.12.28 |
Classifier-Guidance Diffusion (1) | 2022.12.07 |
Improved DDPM (0) | 2022.09.28 |