Fix OSS flake8 and mypy failures (meta-pytorch#1538)

Summary: Pull Request resolved: meta-pytorch#1538 Fix OSS test failures due to updated flake8 and mypy versions. Reviewed By: sarahtranfb Differential Revision: D72193786 fbshipit-source-id: 5567147675307094a18eb95d1e41be0367840cdb

Author: craymichael

captum/_utils/common.py

@@ -863,10 +863,8 @@ def forward_hook(
  inp: Union[Tensor, Tuple[Tensor, ...]],
  out: Union[Tensor, Tuple[Tensor, ...]],
  ) -> None:
- nonlocal grad_out
 
  def output_tensor_hook(output_grad: Tensor) -> None:
- nonlocal grad_out
  grad_out[output_grad.device] = output_grad
 
  if isinstance(out, tuple):
@@ -881,8 +879,6 @@ def pre_hook(module: Module, inp: Union[Tensor, Tuple[Tensor, ...]]) -> Tensor:
  def input_tensor_hook(
  input_grad: Tensor,
  ) -> Union[None, Tensor, Tuple[Tensor, ...]]:
- nonlocal grad_out
-
  if len(grad_out) == 0:
  return None
  hook_out = hook(module, input_grad, grad_out[input_grad.device])

captum/_utils/gradient.py

@@ -306,7 +306,6 @@ def forward_hook(module, inp, out=None):
  if require_layer_grads:
  apply_gradient_requirements(eval_tsrs, warn=False)
  with lock:
- nonlocal saved_layer
  # Note that cloning behaviour of `eval_tsr` is different
  # when `forward_hook_with_return` is set to True. This is because
  # otherwise `backward()` on the last output layer won't execute.

captum/influence/_core/influence_function.py

@@ -1054,7 +1054,6 @@ def compute_intermediate_quantities(
  # define a helper function that returns the embeddings for a batch
  # pyre-fixme[53]: Captured variable `loss_fn` is not annotated.
  def get_batch_embeddings(batch: Tuple[Tensor, ...]) -> Tensor:
- nonlocal loss_fn, reduction_type, return_device
  # if `self.R` is on cpu, and `self.model_device` was not cpu, this implies
  # `self.R` was too large to fit in gpu memory, and we should do the matrix
  # multiplication of the batch jacobians with `self.R` separately for each

captum/influence/_core/tracincp.py

@@ -878,7 +878,6 @@ def compute_intermediate_quantities(
  f_inputs: DataLoader = _format_inputs_dataset(inputs)
 
  def get_checkpoint_contribution(checkpoint: str) -> Tensor:
- nonlocal f_inputs
  assert (
  checkpoint is not None
  ), "None returned from `checkpoints`, cannot load."
@@ -1246,7 +1245,6 @@ def calculate_via_vector_norm(layer_jacobian: Tensor) -> Tensor:
  )
 
  def get_checkpoint_contribution(checkpoint: str) -> Tensor:
- nonlocal inputs_len
  # This function returns a 1D tensor representing the contribution to the
  # self influence score for the given checkpoint, for all batches in
  # `inputs`. The length of the 1D tensor is the total number of

captum/testing/helpers/basic_models.py

@@ -436,7 +436,7 @@ class BasicModel_GradientLayerAttribution(nn.Module):
  def __init__(
  self,
  inplace: bool = False,
- unsupported_layer_output: PassThroughOutputType = None,
+ unsupported_layer_output: Optional[PassThroughOutputType] = None,
  ) -> None:
  super().__init__()
  # Linear 0 is simply identity transform
@@ -464,7 +464,7 @@ def __init__(
  self.linear3.weight = nn.Parameter(torch.ones(2, 4))
  self.linear3.bias = nn.Parameter(torch.tensor([-1.0, 1.0]))
 
- self.int_layer = PassThroughLayerOutput() # sample layer with an int ouput
+ self.int_layer = PassThroughLayerOutput() # sample layer with an int output
 
  @no_type_check
  def forward(

tests/attr/test_input_layer_wrapper.py

@@ -45,7 +45,6 @@
 
 class InputLayerMeta(type):
  def __new__(metacls, name: str, bases: Tuple, attrs: Dict):
- global layer_methods_to_test_with_equiv
  for (
  layer_method,
  equiv_method,