Add stacklevel to across captum to satisfy flake8

Summary:
Add stacklevel to silence flake8:

Warning  (FLAKE8) B028
    No explicit stacklevel argument found. The warn method from the warnings

stacklevel=1 default, stacklevel=2 in some spots that user might need more context

Differential Revision: D64518463

captum/_utils/av.py

@@ -330,7 +330,8 @@ def _manage_loading_layers(
                 "Overwriting activations: load_from_disk is set to False. Removing all "
                 f"activations matching specified parameters {{path: {path}, "
                 f"model_id: {model_id}, layers: {layers}, identifier: {identifier}}} "
-                "before generating new activations."
+                "before generating new activations.",
+                stacklevel=1,
             )
             for layer in layers:
                 files = glob.glob(

captum/_utils/progress.py

@@ -214,7 +214,8 @@ def progress(
             warnings.warn(
                 "Tried to show progress with tqdm "
                 "but tqdm is not installed. "
-                "Fall back to simply print out the progress."
+                "Fall back to simply print out the progress.",
+                stacklevel=1,
             )
         return SimpleProgress(
             iterable, desc=desc, total=total, file=file, mininterval=mininterval

captum/attr/_core/guided_backprop_deconvnet.py

@@ -72,7 +72,8 @@ def attribute(
         # set hooks for overriding ReLU gradients
         warnings.warn(
             "Setting backward hooks on ReLU activations."
-            "The hooks will be removed after the attribution is finished"
+            "The hooks will be removed after the attribution is finished",
+            stacklevel=1,
         )
         try:
             self.model.apply(self._register_hooks)

captum/attr/_core/guided_grad_cam.py

@@ -225,7 +225,8 @@ def attribute(
                 warnings.warn(
                     "Couldn't appropriately interpolate GradCAM attributions for some "
                     "input tensors, returning empty tensor for corresponding "
-                    "attributions."
+                    "attributions.",
+                    stacklevel=1,
                 )
                 output_attr.append(torch.empty(0))
 

captum/attr/_core/lime.py

@@ -715,7 +715,8 @@ def construct_feature_mask(
         if min_interp_features != 0:
             warnings.warn(
                 "Minimum element in feature mask is not 0, shifting indices to"
-                " start at 0."
+                " start at 0.",
+                stacklevel=2,
             )
             feature_mask = tuple(
                 single_mask - min_interp_features for single_mask in feature_mask
@@ -1157,7 +1158,8 @@ def _attribute_kwargs(  # type: ignore
                 "Attempting to construct interpretable model with > 10000 features."
                 "This can be very slow or lead to OOM issues. Please provide a feature"
                 "mask which groups input features to reduce the number of interpretable"
-                "features. "
+                "features. ",
+                stacklevel=1,
             )
 
         coefs: Tensor
@@ -1171,7 +1173,9 @@ def _attribute_kwargs(  # type: ignore
                         "You are providing multiple inputs for Lime / Kernel SHAP "
                         "attributions. This trains a separate interpretable model "
                         "for each example, which can be time consuming. It is "
-                        "recommended to compute attributions for one example at a time."
+                        "recommended to compute attributions for one example at a "
+                        "time.",
+                        stacklevel=1,
                     )
                     output_list = []
                     for (

captum/attr/_core/neuron/neuron_conductance.py

@@ -283,7 +283,8 @@ def attribute(
                 "The neuron_selector provided is a callable. Please ensure that this"
                 " function only selects neurons from the given layer; aggregating"
                 " or performing other operations on the tensor may lead to inaccurate"
-                " results."
+                " results.",
+                stacklevel=1,
             )
         # pyre-fixme[6]: For 1st argument expected `Tensor` but got
         #  `TensorOrTupleOfTensorsGeneric`.

captum/attr/_core/shapley_value.py

@@ -411,7 +411,8 @@ def attribute(
                         warnings.warn(
                             "Feature mask is missing some integers between 0 and "
                             "num_features, for optimal performance, make sure each"
-                            " consecutive integer corresponds to a feature."
+                            " consecutive integer corresponds to a feature.",
+                            stacklevel=1,
                         )
                     # modified_eval dimensions: 1D tensor with length
                     # equal to #num_examples * #features in batch
@@ -858,7 +859,8 @@ def attribute(
             warnings.warn(
                 "You are attempting to compute Shapley Values with at least 10 "
                 "features, which will likely be very computationally expensive."
-                "Consider using Shapley Value Sampling instead."
+                "Consider using Shapley Value Sampling instead.",
+                stacklevel=1,
             )
 
         return super().attribute.__wrapped__(

captum/attr/_models/base.py

@@ -211,7 +211,8 @@ def configure_interpretable_embedding_layer(
         "embeddings and compute attributions for each embedding dimension. "
         "The original embedding layer must be set "
         "back by calling `remove_interpretable_embedding_layer` function "
-        "after model interpretation is finished. "
+        "after model interpretation is finished. ",
+        stacklevel=1,
     )
     interpretable_emb = InterpretableEmbeddingBase(
         embedding_layer, embedding_layer_name

captum/attr/_utils/batching.py

@@ -51,7 +51,8 @@ def _batch_attribution(
         warnings.warn(
             "Internal batch size cannot be less than the number of input examples. "
             "Defaulting to internal batch size of %d equal to the number of examples."
-            % num_examples
+            % num_examples,
+            stacklevel=1,
         )
     # Number of steps for each batch
     step_count = max(1, internal_batch_size // num_examples)
@@ -62,7 +63,8 @@ def _batch_attribution(
                 "This method computes finite differences between evaluations at "
                 "consecutive steps, so internal batch size must be at least twice "
                 "the number of examples. Defaulting to internal batch size of %d"
-                " equal to twice the number of examples." % (2 * num_examples)
+                " equal to twice the number of examples." % (2 * num_examples),
+                stacklevel=1,
             )
 
     total_attr = None
@@ -161,7 +163,8 @@ def _batched_generator(
         warnings.warn(
             """It looks like that the attribution for a gradient-based method is
             computed in a `torch.no_grad` block or perhaps the inputs have no
-            requires_grad."""
+            requires_grad.""",
+            stacklevel=1,
         )
     if internal_batch_size is None:
         # pyre-fixme[7]: Expected `Iterator[Tuple[typing.Tuple[Tensor, ...], typing.A...

captum/influence/_core/similarity_influence.py

@@ -297,7 +297,11 @@ def influence(  # type: ignore[override]
                     "returned as a tensor with [inputs_idx, src_dataset_idx] pairs "
                     "which may have corrupted similarity scores."
                 )
-                warnings.warn(zero_warning, RuntimeWarning)
+                warnings.warn(
+                    zero_warning,
+                    RuntimeWarning,
+                    stacklevel=1,
+                )
                 key = "-".join(["zero_acts", layer])
                 influences[key] = zero_acts
 

captum/influence/_core/tracincp.py

@@ -178,7 +178,8 @@ def __init__(
                 "Unable to determine the number of batches in training dataset "
                 "`train_dataset`. Therefore, if showing the progress of computations, "
                 "only the number of batches processed can be displayed, and not the "
-                "percentage completion of the computation, nor any time estimates."
+                "percentage completion of the computation, nor any time estimates.",
+                stacklevel=1,
             )
 
     @property
@@ -1232,7 +1233,8 @@ def _self_influence_by_checkpoints(
                     "Therefore, if showing the progress of the computation of self "
                     "influence scores, only the number of batches processed can be "
                     "displayed, and not the percentage completion of the computation, "
-                    "nor any time estimates."
+                    "nor any time estimates.",
+                    stacklevel=1,
                 )
 
         # pyre-fixme[3]: Return type must be annotated.

captum/influence/_core/tracincp_fast_rand_proj.py

@@ -579,7 +579,8 @@ def _self_influence_by_checkpoints(
                     "Therefore, if showing the progress of the computation of self "
                     "influence scores, only the number of batches processed can be "
                     "displayed, and not the percentage completion of the computation, "
-                    "nor any time estimates."
+                    "nor any time estimates.",
+                    stacklevel=1,
                 )
 
         # pyre-fixme[53]: Captured variable `inputs_len` is not annotated.
@@ -1040,7 +1041,8 @@ def __init__(
                 "each call to `influence` to compute influence scores or proponents "
                 "will be slower, but may avoid running out of memory."
             )
-            % "`TracInCPFast`"
+            % "`TracInCPFast`",
+            stacklevel=1,
         )
 
         # pyre-fixme[4]: Attribute must be annotated.
@@ -1230,7 +1232,8 @@ def self_influence(
                 "random projections results only in approximate self influence "
                 "scores, there is no reason to use `TracInCPFastRandProj` when "
                 "calculating self influence scores."
-            )
+            ),
+            stacklevel=1,
         )
         raise NotImplementedError
 

captum/influence/_utils/common.py

@@ -420,7 +420,8 @@ def _self_influence_by_batches_helper(
                 "Therefore, if showing the progress of the computation of self "
                 "influence scores, only the number of batches processed can be "
                 "displayed, and not the percentage completion of the computation, "
-                "nor any time estimates."
+                "nor any time estimates.",
+                stacklevel=1,
             )
         # then create the progress bar
         inputs_dataset = progress(
@@ -501,7 +502,8 @@ def _check_loss_fn(
                 f'please set the reduction attribute of `{loss_fn_name}` to "mean", '
                 f'i.e. `{loss_fn_name}.reduction = "mean"`. Note that if '
                 "`sample_wise_grads_per_batch` is True, the implementation "
-                "assumes the reduction is either a sum or mean reduction."
+                "assumes the reduction is either a sum or mean reduction.",
+                stacklevel=1,
             )
             reduction_type = "sum"
         else:
@@ -510,7 +512,8 @@ def _check_loss_fn(
                 "`sample_wise_grads_per_batch` is False, the implementation "
                 f'assumes that `{loss_fn_name}` is a "per-example" loss function (see '
                 f"documentation for `{loss_fn_name}` for details).  Please ensure "
-                "that this is the case."
+                "that this is the case.",
+                stacklevel=1,
             )
 
     return reduction_type
@@ -531,7 +534,8 @@ def _set_active_parameters(model: Module, layers: List[str]) -> List[Module]:
                 warnings.warn(
                     "Setting required grads for layer: {}, name: {}".format(
                         ".".join(layer), name
-                    )
+                    ),
+                    stacklevel=1,
                 )
                 param.requires_grad = True
     return layer_modules
@@ -556,7 +560,8 @@ def _progress_bar_constructor(
             f"of the computation of {quantities_name}, "
             "only the number of batches processed can be "
             "displayed, and not the percentage completion of the computation, "
-            "nor any time estimates."
+            "nor any time estimates.",
+            stacklevel=1,
         )
 
     return progress(
@@ -989,7 +994,10 @@ def _compute_batch_loss_influence_function_base(
                 "`reduction='sum'` loss function, or a `reduction='none'` "
                 "and set `sample_grads_per_batch` to false."
             )
-            warnings.warn(msg)
+            warnings.warn(
+                msg,
+                stacklevel=1,
+            )
         return _loss * multiplier
     elif reduction_type == "sum":
         return _loss

captum/insights/example.py

@@ -7,7 +7,8 @@
 from captum.insights.attr_vis.example import *  # noqa
 
 warnings.warn(
-    "Deprecated. Please import from captum.insights.attr_vis.example instead."
+    "Deprecated. Please import from captum.insights.attr_vis.example instead.",
+    stacklevel=1,
 )
 
 

captum/metrics/_utils/batching.py

@@ -61,7 +61,8 @@ def _divide_and_aggregate_metrics(
                 "to compute the metrics, contains at least an instance of "
                 "the original example and doesn't exceed the number of "
                 "expanded n_perturb_samples."
-            ).format(max_examples_per_batch, bsz)
+            ).format(max_examples_per_batch, bsz),
+            stacklevel=1,
         )
 
     max_inps_per_batch = (

captum/robust/_core/metrics/attack_comparator.py

@@ -399,7 +399,8 @@ def _check_and_evaluate(input_list, key_list):
             for key in attack.additional_args:
                 if key not in kwargs:
                     warnings.warn(
-                        f"Additional sample arg {key} not provided for {attack_key}"
+                        f"Additional sample arg {key} not provided for {attack_key}",
+                        stacklevel=1,
                     )
                 else:
                     additional_attack_args[key] = kwargs[key]