torch optional (#59)

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "smashed"
-version = "0.20.0"
+version = "0.21.0"
 description = """\
 SMASHED is a toolkit designed to apply transformations to samples in \
 datasets, such as fields extraction, tokenization, prompting, batching, \
@@ -11,7 +11,6 @@ license = {text = "Apache-2.0"}
 readme = "README.md"
 requires-python = ">=3.8"
 dependencies = [
-    "torch>=1.9",
     "necessary>=0.4.1",
     "trouting>=0.3.3",
     "ftfy>=6.1.1",
@@ -103,12 +102,17 @@ remote = [
     "smart-open>=5.2.1",
     "boto3>=1.25.5",
 ]
+torch = [
+    "torch>=1.9",
+]
 datasets = [
+    "smashed[torch]",
     "transformers>=4.5",
     "datasets>=2.8.0",
     "dill>=0.3.0",
 ]
 prompting = [
+    "smashed[torch]",
     "transformers>=4.5",
     "promptsource>=0.2.3",
     "blingfire>=0.1.8",
@@ -119,6 +123,7 @@ torchdata = [
 ]
 all = [
     "smashed[dev]",
+    "smashed[torch]",
     "smashed[datasets]",
     "smashed[torchdata]",
     "smashed[remote]",

src/smashed/base/interfaces.py

@@ -15,7 +15,6 @@
 )
 
 from necessary import necessary
-from torch._utils import classproperty
 from trouting import trouting
 
 from .abstract import (
@@ -52,7 +51,7 @@ class MapMethodInterfaceMixIn(AbstractBaseMapper):
     and various interfaces. Do not inherit from this class directly,
     but use SingleBaseMapper/BatchedBaseMapper instead."""
 
-    @classproperty
+    @classmethod
     def always_remove_columns(cls) -> bool:
         """Whether this mapper should always remove its input columns
         from the dataset. If False, the mapper will only remove columns
@@ -191,7 +190,7 @@ def _map_list_of_dicts(
         # TODO[lucas]: maybe support specifying which fields to keep?
         remove_columns = (
             bool(map_kwargs.get("remove_columns", False))
-            or self.always_remove_columns
+            or self.always_remove_columns()
         )
 
         if isinstance(dataset, abc.Sequence):
@@ -258,7 +257,7 @@ def _map_huggingface_dataset(
 
             print_fingerprint = map_kwargs.pop("print_fingerprint", False)
 
-            if self.always_remove_columns:
+            if self.always_remove_columns():
                 remove_columns = list(dataset.features.keys())
             else:
                 remove_columns = map_kwargs.get("remove_columns", [])
@@ -320,7 +319,7 @@ def _map_huggingface_dataset_batch(
             # TODO[lucas]: maybe support specifying which fields to keep?
             remove_columns = (
                 bool(map_kwargs.get("remove_columns", False))
-                or self.always_remove_columns
+                or self.always_remove_columns()
             )
 
             dtview: DataBatchView[LazyBatch, str, Any] = DataBatchView(dataset)

src/smashed/mappers/collators.py

@@ -14,7 +14,6 @@
     Union,
 )
 
-import torch
 from necessary import necessary
 
 from ..base import SingleBaseMapper, TransformElementType
@@ -26,6 +25,10 @@
             PreTrainedTokenizerBase,
         )
 
+with necessary("torch", soft=True) as PYTORCH_AVAILABLE:
+    if PYTORCH_AVAILABLE or TYPE_CHECKING:
+        import torch
+
 
 __all__ = [
     "ListCollatorMapper",
@@ -170,15 +173,21 @@ class TensorCollatorMapper(BaseCollator, SingleBaseMapper):
     >>> data_loader = DataLoader(..., collate_fn=collator.transform)
     """
 
+    def __init__(self, *args, **kwargs):
+        if not PYTORCH_AVAILABLE:
+            cls_name = self.__class__.__name__
+            raise ImportError(f"Pytorch is required to use {cls_name}")
+        super().__init__(*args, **kwargs)
+
     @staticmethod
     def _pad(
-        sequence: Sequence[torch.Tensor],
+        sequence: Sequence["torch.Tensor"],
         pad_value: Union[int, float],
         dim: int = 0,
         pad_to_length: Optional[Union[int, Sequence[int]]] = None,
         pad_to_multiple_of: Optional[int] = None,
         right_pad: bool = True,
-    ) -> torch.Tensor:
+    ) -> "torch.Tensor":
         """Pad a sequence of tensors to the same length.
 
         Args:
@@ -272,8 +281,8 @@ def _pad(
         return torch.cat(to_stack, dim=dim)
 
     def transform(  # type: ignore
-        self: "TensorCollatorMapper", data: Dict[str, Sequence[torch.Tensor]]
-    ) -> Dict[str, torch.Tensor]:
+        self: "TensorCollatorMapper", data: Dict[str, Sequence["torch.Tensor"]]
+    ) -> Dict[str, "torch.Tensor"]:
         collated_data = {
             field_name: self._pad(
                 sequence=list_of_tensors,

src/smashed/mappers/converters.py

@@ -1,6 +1,5 @@
 from typing import TYPE_CHECKING, Any, Dict, Mapping, Optional, TypeVar, Union
 
-import torch
 from necessary import necessary
 from trouting import trouting
 
@@ -15,16 +14,22 @@
             "HuggingFaceDataset", Dataset, IterableDataset
         )
 
+with necessary("torch", soft=True) as PYTORCH_AVAILABLE:
+    if PYTORCH_AVAILABLE or TYPE_CHECKING:
+        import torch
+
 
 class Python2TorchMapper(SingleBaseMapper):
     __slots__ = ["field_cast_map", "device"]
-    field_cast_map: Dict[str, torch.dtype]
-    device: Union[torch.device, None]
+    field_cast_map: Dict[str, "torch.dtype"]
+    device: Union["torch.device", None]
 
     def __init__(
         self: "Python2TorchMapper",
-        field_cast_map: Optional[Mapping[str, Union[str, torch.dtype]]] = None,
-        device: Optional[Union[torch.device, str]] = None,
+        field_cast_map: Optional[
+            Mapping[str, Union[str, "torch.dtype"]]
+        ] = None,
+        device: Optional[Union["torch.device", str]] = None,
     ) -> None:
         """Mapper that converts Python types to Torch types. It can optionally
         cast the values of a field to a specific type, and move to a specific
@@ -37,6 +42,10 @@ def __init__(
             device (Union[torch.device, str], optional): Device to move the
                 tensors to. Defaults to None, which means no moving occurs.
         """
+        if not PYTORCH_AVAILABLE:
+            cls_name = self.__class__.__name__
+            raise ImportError(f"{cls_name} requires PyTorch to be installed")
+
         self.device = torch.device(device) if device else None
 
         self.field_cast_map = {
@@ -49,7 +58,7 @@ def __init__(
         )
 
     @staticmethod
-    def _get_dtype(dtype: Any) -> torch.dtype:
+    def _get_dtype(dtype: Any) -> "torch.dtype":
         if isinstance(dtype, str):
             dtype = getattr(torch, dtype, None)
             if dtype is None:
@@ -102,7 +111,7 @@ def __init__(self: "Torch2PythonMapper") -> None:
         super().__init__()
 
     def transform(  # type: ignore
-        self: "Torch2PythonMapper", data: Dict[str, torch.Tensor]
+        self: "Torch2PythonMapper", data: Dict[str, "torch.Tensor"]
     ) -> TransformElementType:
         return {
             field_name: field_value.cpu().tolist()

src/smashed/mappers/fields.py

@@ -1,7 +1,6 @@
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, TypeVar
 
 from necessary import necessary
-from torch._utils import classproperty
 
 from ..base import SingleBaseMapper, TransformElementType
 
@@ -19,7 +18,7 @@ class ChangeFieldsMapper(SingleBaseMapper):
     """Mapper that removes some of the fields in a dataset.
     Either `keep_fields` or `drop_fields` must be specified, but not both."""
 
-    @classproperty
+    @classmethod
     def always_remove_columns(cls) -> bool:
         return True
 
@@ -71,7 +70,7 @@ def transform(self, data: TransformElementType) -> TransformElementType:
 class RenameFieldsMapper(SingleBaseMapper):
     """Mapper that renames some of the fields batch"""
 
-    @classproperty
+    @classmethod
     def always_remove_columns(cls) -> bool:
         return True
 

src/smashed/mappers/glom.py

@@ -23,7 +23,7 @@ class ExtendGlommerMixin:
 
     def __getstate__(self):
         state = super().__getstate__()  # pyright: ignore
-        state["__dict__"].pop("glommer", None)
+        state["__dict__"].pop("glommer", None)  # pyright: ignore
         return state
 
     @cached_property

src/smashed/utils/io_utils/compression.py

@@ -48,7 +48,6 @@ def compress_stream(
     errors: str = "strict",
     gzip: bool = True,
 ) -> Iterator[IO]:
-
     assert gzip, "Only gzip compression is supported at this time"
 
     if mode == "wb" or mode == "w":