DataLoader validation

src/refiners/training_utils/data_loader.py

@@ -1,22 +1,34 @@
 from typing import Callable, TypeVar
 
-from pydantic import BaseModel, ConfigDict, PositiveInt
+from pydantic import BaseModel, ConfigDict, NonNegativeInt, PositiveInt, model_validator
 from torch.utils.data import DataLoader, Dataset
+from typing_extensions import Self
 
 BatchT = TypeVar("BatchT")
 
 
 class DataLoaderConfig(BaseModel):
     batch_size: PositiveInt = 1
-    num_workers: int = 0
+    num_workers: NonNegativeInt = 0
     pin_memory: bool = False
-    prefetch_factor: int | None = None
+    prefetch_factor: PositiveInt | None = None
     persistent_workers: bool = False
     drop_last: bool = False
     shuffle: bool = True
 
     model_config = ConfigDict(extra="forbid")
-    # TODO: Add more validation to the config
+
+    @model_validator(mode="after")
+    def check_prefetch_factor(self) -> Self:
+        if self.prefetch_factor is not None and self.num_workers == 0:
+            raise ValueError(f"prefetch_factor={self.prefetch_factor} requires num_workers > 0")
+        return self
+
+    @model_validator(mode="after")
+    def check_num_workers(self) -> Self:
+        if self.num_workers == 0 and self.persistent_workers is True:
+            raise ValueError(f"persistent_workers={self.persistent_workers} option needs num_workers > 0")
+        return self
 
 
 class DatasetFromCallable(Dataset[BatchT]):

tests/training_utils/test_data_loader.py

@@ -0,0 +1,48 @@
+import pytest
+from pydantic import ValidationError
+from torch.utils.data import DataLoader
+
+from refiners.training_utils.data_loader import DataLoaderConfig, DatasetFromCallable, create_data_loader
+
+
+def get_item(index: int) -> int:
+    return index * 2
+
+
+@pytest.fixture
+def config() -> DataLoaderConfig:
+    return DataLoaderConfig(batch_size=2, num_workers=2, persistent_workers=True)
+
+
+def test_dataloader_config_valid(config: DataLoaderConfig) -> None:
+    assert config.batch_size == 2
+    assert config.num_workers == 2
+    assert config.persistent_workers == True
+
+
+def test_dataloader_config_invalid() -> None:
+    with pytest.raises(ValidationError):
+        DataLoaderConfig(num_workers=0, prefetch_factor=2)
+
+    with pytest.raises(ValidationError):
+        DataLoaderConfig(num_workers=0, persistent_workers=True)
+
+
+def test_dataset_from_callable():
+    dataset = DatasetFromCallable(get_item, 200)
+    assert len(dataset) == 200
+    assert dataset[0] == 0
+    assert dataset[5] == 10
+
+
+def test_create_data_loader(config: DataLoaderConfig) -> None:
+    data_loader = create_data_loader(get_item, 100, config)
+    assert isinstance(data_loader, DataLoader)
+
+
+def test_create_data_loader_with_collate_fn(config: DataLoaderConfig) -> None:
+    def collate_fn(batch: list[int]):
+        return sum(batch)
+
+    data_loader = create_data_loader(get_item, 20, config=config, collate_fn=collate_fn)
+    assert isinstance(data_loader, DataLoader)