CompVis/stable-diffusion/main.py

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