Final round of Pyre enablement (#1330)

Summary:
Pull Request resolved: #1330

Primarily use infer to annotate lots of easy case.

`pyre -n --target fbcode//pytorch/captum/captum/... infer -i pytorch/captum/captum/`

`pyre -n --target fbcode//pytorch/captum/tests/... infer -i pytorch/captum/tests/`

Reviewed By: vivekmig

Differential Revision: D61053870

fbshipit-source-id: 37f96e7e5590dc57ddbcaa3c1da106149f5e09bf

Author: yucu

captum/_utils/common.py

@@ -363,10 +363,9 @@ def _expand_target(
     return target
 
 
-# pyre-fixme[3]: Return type must be annotated.
 def _expand_feature_mask(
     feature_mask: Union[Tensor, Tuple[Tensor, ...]], n_samples: int
-):
+) -> Tuple[Tensor, ...]:
     # pyre-fixme[6]: For 1st argument expected `Tensor` but got `Union[Tensor,
     #  typing.Tuple[Tensor, ...]]`.
     is_feature_mask_tuple = _is_tuple(feature_mask)
@@ -379,18 +378,17 @@ def _expand_feature_mask(
         )
         for feature_mask_elem in feature_mask
     )
-    return _format_output(is_feature_mask_tuple, feature_mask_new)
+    return _format_output(is_feature_mask_tuple, feature_mask_new)  # type: ignore
 
 
-# pyre-fixme[3]: Return type must be annotated.
 def _expand_and_update_baselines(
     inputs: Tuple[Tensor, ...],
     n_samples: int,
     # pyre-fixme[24]: Generic type `dict` expects 2 type parameters, use
     #  `typing.Dict[<key type>, <value type>]` to avoid runtime subscripting errors.
     kwargs: dict,
     draw_baseline_from_distrib: bool = False,
-):
+) -> None:
     # pyre-fixme[3]: Return type must be annotated.
     # pyre-fixme[2]: Parameter must be annotated.
     def get_random_baseline_indices(bsz, baseline):
@@ -432,10 +430,9 @@ def get_random_baseline_indices(bsz, baseline):
     kwargs["baselines"] = baselines
 
 
-# pyre-fixme[3]: Return type must be annotated.
 # pyre-fixme[24]: Generic type `dict` expects 2 type parameters, use
 #  `typing.Dict[<key type>, <value type>]` to avoid runtime subscripting errors.
-def _expand_and_update_additional_forward_args(n_samples: int, kwargs: dict):
+def _expand_and_update_additional_forward_args(n_samples: int, kwargs: dict) -> None:
     if "additional_forward_args" not in kwargs:
         return
     additional_forward_args = kwargs["additional_forward_args"]
@@ -451,10 +448,9 @@ def _expand_and_update_additional_forward_args(n_samples: int, kwargs: dict):
     kwargs["additional_forward_args"] = additional_forward_args
 
 
-# pyre-fixme[3]: Return type must be annotated.
 # pyre-fixme[24]: Generic type `dict` expects 2 type parameters, use
 #  `typing.Dict[<key type>, <value type>]` to avoid runtime subscripting errors.
-def _expand_and_update_target(n_samples: int, kwargs: dict):
+def _expand_and_update_target(n_samples: int, kwargs: dict) -> None:
     if "target" not in kwargs:
         return
     target = kwargs["target"]
@@ -465,10 +461,9 @@ def _expand_and_update_target(n_samples: int, kwargs: dict):
     kwargs["target"] = target
 
 
-# pyre-fixme[3]: Return type must be annotated.
 # pyre-fixme[24]: Generic type `dict` expects 2 type parameters, use
 #  `typing.Dict[<key type>, <value type>]` to avoid runtime subscripting errors.
-def _expand_and_update_feature_mask(n_samples: int, kwargs: dict):
+def _expand_and_update_feature_mask(n_samples: int, kwargs: dict) -> None:
     if "feature_mask" not in kwargs:
         return
 
@@ -573,10 +568,9 @@ def _format_outputs(
 # pyre-fixme[24] Callable requires 2 arguments
 def _construct_future_forward(original_forward: Callable) -> Callable:
     # pyre-fixme[3] return type not specified
-    # pyre-ignore
-    def future_forward(*args, **kwargs):
-        # pyre-ignore
-        fut = torch.futures.Future()
+    def future_forward(*args: Any, **kwargs: Any):
+        # pyre-fixme[29]: `typing.Type[torch.futures.Future]` is not a function.
+        fut: torch.futures.Future[Tensor] = torch.futures.Future()
         fut.set_result(original_forward(*args, **kwargs))
         return fut
 
@@ -921,8 +915,7 @@ def input_tensor_hook(input_grad: Tensor):
     ]
 
 
-# pyre-fixme[3]: Return type must be annotated.
-def _get_max_feature_index(feature_mask: Tuple[Tensor, ...]):
+def _get_max_feature_index(feature_mask: Tuple[Tensor, ...]) -> int:
     """
     Returns the max feature mask index
     The feature mask should be formatted to tuple of tensors at first.

captum/_utils/gradient.py

@@ -986,7 +986,11 @@ def _compute_jacobian_wrt_params_with_sample_wise_trick(
                 out = loss
 
             sample_grad_wrapper.compute_param_sample_gradients(
-                out, loss_mode=reduction_type
+                out,
+                # pyre-fixme[6]: In call `SampleGradientWrapper.
+                # compute_param_sample_gradients`, for argument `loss_mode`,
+                # expected `str` but got `Optional[str]`.
+                loss_mode=reduction_type,  # type: ignore
             )
             if layer_modules is not None:
                 layer_parameters = _extract_parameters_from_layers(layer_modules)

captum/_utils/models/linear_model/model.py

@@ -41,7 +41,6 @@ def __init__(self, train_fn: Callable, **kwargs) -> None:
         # pyre-fixme[4]: Attribute must be annotated.
         self.construct_kwargs = kwargs
 
-    # pyre-fixme[3]: Return type must be annotated.
     def _construct_model_params(
         self,
         in_features: Optional[int] = None,
@@ -52,7 +51,7 @@ def _construct_model_params(
         weight_values: Optional[Tensor] = None,
         bias_value: Optional[Tensor] = None,
         classes: Optional[Tensor] = None,
-    ):
+    ) -> None:
         r"""
         Lazily initializes a linear model. This will be called for you in a
         train method.

captum/_utils/models/linear_model/train.py

@@ -9,9 +9,8 @@
 from torch.utils.data import DataLoader
 
 
-# pyre-fixme[3]: Return type must be annotated.
 # pyre-fixme[2]: Parameter must be annotated.
-def l2_loss(x1, x2, weights=None):
+def l2_loss(x1, x2, weights=None) -> torch.Tensor:
     if weights is None:
         return torch.mean((x1 - x2) ** 2) / 2.0
     else:
@@ -236,7 +235,7 @@ def get_point(datapoint):
 
 class NormLayer(nn.Module):
     # pyre-fixme[2]: Parameter must be annotated.
-    def __init__(self, mean, std, n=None, eps=1e-8) -> None:
+    def __init__(self, mean, std, n=None, eps: float = 1e-8) -> None:
         super().__init__()
         # pyre-fixme[4]: Attribute must be annotated.
         self.mean = mean
@@ -251,7 +250,6 @@ def forward(self, x):
         return (x - self.mean) / (self.std + self.eps)
 
 
-# pyre-fixme[3]: Return type must be annotated.
 def sklearn_train_linear_model(
     model: LinearModel,
     dataloader: DataLoader,
@@ -260,7 +258,7 @@ def sklearn_train_linear_model(
     norm_input: bool = False,
     # pyre-fixme[2]: Parameter must be annotated.
     **fit_kwargs,
-):
+) -> Dict[str, float]:
     r"""
     Alternative method to train with sklearn. This does introduce some slight
     overhead as we convert the tensors to numpy and then convert the resulting

captum/_utils/progress.py

@@ -5,7 +5,7 @@
 import sys
 import warnings
 from time import time
-from typing import cast, Iterable, Optional, Sized, TextIO
+from typing import Any, cast, Iterable, Optional, Sized, TextIO
 
 from captum._utils.typing import Literal
 
@@ -61,15 +61,17 @@ class NullProgress:
     progress bars.
     """
 
-    # pyre-fixme[3]: Return type must be annotated.
-    # pyre-fixme[24]: Generic type `Iterable` expects 1 type parameter.
-    # pyre-fixme[2]: Parameter must be annotated.
-    def __init__(self, iterable: Optional[Iterable] = None, *args, **kwargs):
+    def __init__(
+        self,
+        # pyre-fixme[24]: Generic type `Iterable` expects 1 type parameter.
+        iterable: Optional[Iterable] = None,
+        *args: Any,
+        **kwargs: Any,
+    ) -> None:
         del args, kwargs
         self.iterable = iterable
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def __enter__(self):
+    def __enter__(self) -> "NullProgress":
         return self
 
     # pyre-fixme[2]: Parameter must be annotated.
@@ -87,12 +89,10 @@ def __iter__(self):
         for it in self.iterable:
             yield it
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def update(self, amount: int = 1):
+    def update(self, amount: int = 1) -> None:
         pass
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def close(self):
+    def close(self) -> None:
         pass
 
 
@@ -133,8 +133,7 @@ def __init__(
         self.closed = False
         self._is_parent = False
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def __enter__(self):
+    def __enter__(self) -> "SimpleProgress":
         self._is_parent = True
         self._refresh()
         return self
@@ -158,8 +157,7 @@ def __iter__(self):
             self.update()
         self.close()
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def _refresh(self):
+    def _refresh(self) -> None:
         progress_str = self.desc + ": " if self.desc else ""
         if self.total:
             # e.g., progress: 60% 3/5
@@ -172,8 +170,7 @@ def _refresh(self):
         end = "\n" if self._is_parent else ""
         print("\r" + progress_str, end=end, file=self.file)
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def update(self, amount: int = 1):
+    def update(self, amount: int = 1) -> None:
         if self.closed:
             return
         self.cur += amount
@@ -183,8 +180,7 @@ def update(self, amount: int = 1):
             self._refresh()
             self.last_print_t = cur_t
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def close(self):
+    def close(self) -> None:
         if not self.closed and not self._is_parent:
             self._refresh()
             print(file=self.file)  # end with new line
@@ -197,8 +193,7 @@ def progress(
     iterable: Optional[Iterable] = None,
     desc: Optional[str] = None,
     total: Optional[int] = None,
-    # pyre-fixme[2]: Parameter must be annotated.
-    use_tqdm=True,
+    use_tqdm: bool = True,
     file: Optional[TextIO] = None,
     mininterval: float = 0.5,
     # pyre-fixme[2]: Parameter must be annotated.

captum/_utils/sample_gradient.py

@@ -103,7 +103,7 @@ class SampleGradientWrapper:
     """
 
     # pyre-fixme[2]: Parameter must be annotated.
-    def __init__(self, model, layer_modules=None) -> None:
+    def __init__(self, model, layer_modules: Optional[List[Module]] = None) -> None:
         # pyre-fixme[4]: Attribute must be annotated.
         self.model = model
         self.hooks_added = False
@@ -162,8 +162,9 @@ def _reset(self) -> None:
         self.activation_dict = defaultdict(list)
         self.gradient_dict = defaultdict(list)
 
-    # pyre-fixme[2]: Parameter must be annotated.
-    def compute_param_sample_gradients(self, loss_blob, loss_mode="mean") -> None:
+    def compute_param_sample_gradients(
+        self, loss_blob: Tensor, loss_mode: str = "mean"
+    ) -> None:
         assert (
             loss_mode.upper() in LossMode.__members__
         ), f"Provided loss mode {loss_mode} is not valid"

captum/attr/_core/dataloader_attr.py

@@ -30,9 +30,8 @@ class InputRole:
 
 
 # default reducer wehn reduce is None. Simply concat the outputs by the batch dimension
-# pyre-fixme[3]: Return type must be annotated.
 # pyre-fixme[2]: Parameter must be annotated.
-def _concat_tensors(accum, cur_output, _):
+def _concat_tensors(accum, cur_output, _) -> Tensor:
     return cur_output if accum is None else torch.cat([accum, cur_output])
 
 
@@ -185,7 +184,6 @@ def __init__(self, attr_method: Attribution) -> None:
 
         self.attr_method.forward_func = self._forward_with_dataloader
 
-    # pyre-fixme[3]: Return type must be annotated.
     def _forward_with_dataloader(
         self,
         batched_perturbed_feature_indices: Tensor,
@@ -199,7 +197,7 @@ def _forward_with_dataloader(
         to_metric: Optional[Callable],
         show_progress: bool,
         feature_idx_to_mask_idx: Dict[int, List[int]],
-    ):
+    ) -> Tensor:
         """
         Wrapper of the original given forward_func to be used in the attribution method
         It iterates over the dataloader with the given forward_func
@@ -468,10 +466,8 @@ def attribute(
 
             return _format_output(is_inputs_tuple, attr)
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def attribute_future(
-        self,
-    ):
+    # pyre-fixme[24] Generic type `Callable` expects 2 type parameters.
+    def attribute_future(self) -> Callable:
         r"""
         This method is not implemented for DataLoaderAttribution.
         """