class DataLoaderCollectionConfig(cfg.BaseConfig):
transform: TransformConfig = cfg.Option(required=True, type=TransformConfig)
train: DatasetConfig = cfg.Option(required=True, type=DatasetConfig)
test: DatasetConfig = cfg.Option(required=True, type=DatasetConfig)
valid: Optional[DatasetConfig] = cfg.Option(nullable=True, type=DatasetConfig)
@cached_property
def data_loader_collection(self):
self.train.set_transforms(
self.transform.transform_factory.train_transform,
self.transform.transform_factory.train_target_transform
)
self.test.set_transforms(
self.transform.transform_factory.test_transform,
self.transform.transform_factory.test_target_transform
)
if self.valid is not None:
self.valid.set_transforms(
self.transform.transform_factory.valid_transform,
self.transform.transform_factory.valid_target_transform
)
return DataLoaderCollection(
self.train.loader,
self.test.loader,
self.valid.loader if self.valid is not None else None
)
class MetricsCollectionConfig(cfg.BaseConfig):
_train: List[str] = cfg.Option(name="train",
default=[],
type=cfg.config_list(str))
_test: List[str] = cfg.Option(name="test",
default=[],
type=cfg.config_list(str))
_valid: List[str] = cfg.Option(name="valid",
default=[],
type=cfg.config_list(str))
@cached_property
def metrics_collection(self):
return MetricsCollection(self._get_metric_list(self._train),
self._get_metric_list(self._test),
self._get_metric_list(self._valid))
def _get_metric_list(self, metric_names: List[str]) -> List[MetricBase]:
return [
self._get_metric_cls(metric_name)() for metric_name in metric_names
]
def _get_metric_cls(self, metric_name: str) -> Type[MetricBase]:
cls = vars(metrics).get(metric_name, NOT_FOUND)
if cls is NOT_FOUND:
raise KeyError(f"the class {metric_name} was not found.")
return cls
class MyConfig(cfg.BaseConfig):
_a: int = cfg.Option(name="a", required=True, description="hi", type=int)
b: float = cfg.Option(required=False, type=float)
c: bool = cfg.Option(required=True, type=bool)
li: List[MyConfigEmb] = cfg.Option(
type=cfg.config_list(cfg.config_list(MyConfigEmb)))
def post_load(self):
print(self.a)
@property
def a(self):
return self._a + 1
class TrainerConfig(cfg.BaseConfig):
out: Path = cfg.Option(required=True, type=Path)
train_rate: float = cfg.Option(default=0.1, type=float)
max_epochs: int = cfg.Option(required=True, type=int)
save_checkpoint_epochs: int = cfg.Option(default=20, type=int)
device: torch.device = cfg.Option(default=DEVICE_AUTO, type=process_device)
half: bool = cfg.Option(default=True, type=bool)
ddp: bool = cfg.Option(default=True, type=bool)
test_no_grad: bool = cfg.Option(default=True, type=bool)
# TODO callback
def post_load(self):
self.half = self.half and self._is_half_available()
self.ddp = self.ddp and torch.cuda.is_available(
) and "cuda" in self.device.type
def _is_half_available(self):
try:
import apex
return torch.cuda.is_available()
except ImportError as _:
return False
class TransformConfig(ImporterConfig):
param: dict = cfg.Option(default={}, type=cfg.process.flag_container)
@cached_property
def transform_factory(self) -> TransformsFactory:
if issubclass(self.imported, TransformsFactory):
return self.imported(**self.param)
raise TypeError("imported must be TransformFactory")
class LossFunctionConfig(ImporterConfig):
param: dict = cfg.Option(default={}, type=cfg.process.flag_container)
@cached_property
def loss_func(self):
if isinstance(self.imported, type):
return self.imported(**self.param)
return self.imported
def eval_loss(self, data: Tensor, target: Tensor):
return self.loss_func(data, target)
class TrainingConfig(cfg.BaseConfig):
profile_name: str = cfg.Option(required=True, type=str)
_model: ModelConfig = cfg.Option(name="model",
required=True,
type=ModelConfig)
_optimizer: OptimizerConfig = cfg.Option(name="optimizer",
required=True,
type=OptimizerConfig)
dataset: DataLoaderCollectionConfig = cfg.Option(
required=True, type=DataLoaderCollectionConfig)
metrics: MetricsCollectionConfig = cfg.Option(default={},
type=MetricsCollectionConfig)
_trainer: TrainerConfig = cfg.Option(name="trainer",
required=True,
type=TrainerConfig)
_loss_func: LossFunctionConfig = cfg.Option(name="loss_func",
required=True,
type=LossFunctionConfig)
@property
def model(self) -> Module:
return self._model.model
@cached_property
def optimizer(self) -> Optimizer:
self._optimizer.set_model(self.model)
return self._optimizer.optimizer
@property
def loss_func(self) -> Callable:
return self._loss_func.eval_loss
@property
def data_loader_collection(self) -> DataLoaderCollection:
return self.dataset.data_loader_collection
@property
def metrics_collection(self) -> MetricsCollection:
return self.metrics.metrics_collection
@cached_property
def trainer(self) -> Trainer:
return Trainer(self._get_out_prefix(), self.data_loader_collection,
self.metrics_collection, self.loss_func, self.model,
self.optimizer, self._trainer.out,
self._trainer.train_rate, self._trainer.max_epochs,
self._trainer.save_checkpoint_epochs,
self._trainer.device, self._trainer.half,
self._trainer.ddp)
def _get_out_prefix(self) -> str:
cls_name = self._model.imported.__name__
return self.profile_name + "_" + cls_name
class ModelConfig(ImporterConfig):
param: dict = cfg.Option(default={}, type=cfg.process.flag_container)
@cached_property
def model(self) -> Module:
if not isinstance(self.param, dict):
raise TypeError("params must be a dict")
model = self.imported(**self.param)
if not isinstance(model, Module):
raise TypeError("imported_cls must be subclass of torch.nn.Module or a Callable returned it.")
return model
class DatasetConfig(ImporterConfig):
param: dict = cfg.Option(default={}, type=cfg.process.flag_container)
batch_size: int = cfg.Option(default=1, type=int)
shuffle: bool = cfg.Option(default=False, type=bool)
pin_memory: bool = cfg.Option(default=True, type=bool)
num_workers: int = cfg.Option(default=0, type=int)
transform: Callable = None
target_transform: Callable = None
def set_transforms(self, transform: Callable, target_transform: Callable):
self.transform = transform
self.target_transform = target_transform
@property
def loader(self) -> DataLoader:
loader = DataLoader(
self.dataset,
shuffle=self.shuffle,
pin_memory=self.pin_memory,
batch_size=self.batch_size,
num_workers=self.num_workers
)
return loader
# TODO transform and target_transform
@property
def dataset(self) -> Dataset:
cls = self.imported
if not isinstance(cls, type):
raise ValueError("Dataset type should be a class method_name.")
if not issubclass(cls, Dataset):
raise ValueError(f"Incorrect dataset type: {cls}.")
if not isinstance(self.param, MutableMapping):
raise ValueError('param must be instance of MutableMapping')
dataset = cls(transform=self.transform, target_transform=self.target_transform, **self.param)
return dataset
class OptimizerConfig(ImporterConfig):
param: dict = cfg.Option(default={}, type=cfg.process.flag_container)
model: Module = None
def set_model(self, model: Module) -> None:
self.model = model
@cached_property
def optimizer(self) -> Optimizer:
if self.model is None:
raise ValueError("model has not loaded.")
if not isinstance(self.param, dict):
raise TypeError("params must be a dict")
optimizer = self.imported(self.model.parameters(), **self.param)
if not isinstance(optimizer, Optimizer):
raise TypeError(
"imported_cls must be subclass of torch.optim.Optimizer or a Callable returned it."
)
return optimizer
class ImporterConfig(cfg.BaseConfig):
_import_statement: str = cfg.Option(required=True,
name="import_statement",
type=str)
statement_re = r"from (\.*((\w+)\.)*\w+) import (\w+)"
@cached_property
def imported(self):
_from, _import = self._split_from_and_import()
return vars(importlib.import_module(_from))[_import]
def _split_from_and_import(self):
import_statement = self._import_statement.strip()
result = re.fullmatch(self.statement_re, import_statement)
if result is None:
raise ValueError(
f"Incorrect import_statement: {self._import_statement} ({self.statement_re})"
)
g = result.groups()
return g[0], g[3]
class PylonCameraConfig(cfg.BaseConfig):
exposure_time: int = cfg.Option(default=100, type=int)
shutter_mode: str = cfg.Option(default="GlobalResetRelease", type=str)
width: int = cfg.Option(required=True, type=int)
height: int = cfg.Option(required=True, type=int)
delay: float = cfg.Option(required=True, type=float)
class MyConfigEmb(cfg.BaseConfig):
x: int = cfg.Option(type=int)
y: int = cfg.Option(type=int)
i: bool = cfg.Option(type=bool)
