Avoid unnecessary tensor construction when creating input masks for permutation/ablation (#1527)

Summary:

Study: https://docs.google.com/spreadsheets/d/1GyNJJBrNkazGOyJQLv00QV4phX2R3488oNgVPT17qzU/edit?gid=0#gid=0
Saw a regression in the new logic introduced in D69531512 with one of the models for both permutation and ablation methods, potentially due to large sparse features. vivekmig suggested we can avoid creating all these zero tensors

Differential Revision: D71057703

Author: sarahtranfb

captum/attr/_core/feature_ablation.py

@@ -3,7 +3,18 @@
 # pyre-strict
 
 import math
-from typing import Any, Callable, cast, Generator, List, Optional, Tuple, TypeVar, Union
+from typing import (
+    Any,
+    Callable,
+    cast,
+    Dict,
+    Generator,
+    List,
+    Optional,
+    Tuple,
+    TypeVar,
+    Union,
+)
 
 import torch
 from captum._utils.common import (
@@ -465,13 +476,21 @@ def _attribute_with_cross_tensor_feature_masks(
         attrib_type: dtype,
         **kwargs: Any,
     ) -> Tuple[List[Tensor], List[Tensor]]:
+        feature_idx_to_tensor_idx: Dict[int, List[int]] = {}
+        for i, mask in enumerate(formatted_feature_mask):
+            for feature_idx in torch.unique(mask):
+                if feature_idx.item() not in feature_idx_to_tensor_idx.keys():
+                    feature_idx_to_tensor_idx[feature_idx.item()] = []
+                feature_idx_to_tensor_idx[feature_idx.item()].append(i)
+
         for (
             current_inputs,
             current_mask,
         ) in self._ablation_generator(
             formatted_inputs,
             baselines,
             formatted_feature_mask,
+            feature_idx_to_tensor_idx,
             **kwargs,
         ):
             # modified_eval has (n_feature_perturbed * n_outputs) elements
@@ -511,11 +530,12 @@ def _ablation_generator(
         inputs: Tuple[Tensor, ...],
         baselines: BaselineType,
         input_mask: Tuple[Tensor, ...],
+        feature_idx_to_tensor_idx: Dict[int, List[int]],
         **kwargs: Any,
     ) -> Generator[
         Tuple[
             Tuple[Tensor, ...],
-            Tuple[Tensor, ...],
+            Tuple[Optional[Tensor], ...],
         ],
         None,
         None,
@@ -531,7 +551,11 @@ def _ablation_generator(
         for feature_idx in unique_feature_ids:
             ablated_inputs, current_masks = (
                 self._construct_ablated_input_across_tensors(
-                    inputs, input_mask, baselines, feature_idx
+                    inputs,
+                    input_mask,
+                    baselines,
+                    feature_idx,
+                    feature_idx_to_tensor_idx[feature_idx],
                 )
             )
             yield ablated_inputs, current_masks
@@ -542,18 +566,17 @@ def _construct_ablated_input_across_tensors(
         input_mask: Tuple[Tensor, ...],
         baselines: BaselineType,
         feature_idx: int,
-    ) -> Tuple[Tuple[Tensor, ...], Tuple[Tensor, ...]]:
+        tensor_idxs: List[int],
+    ) -> Tuple[Tuple[Tensor, ...], Tuple[Optional[Tensor], ...]]:
 
         ablated_inputs = []
-        current_masks = []
+        current_masks: List[Optional[Tensor]] = []
         for i, input_tensor in enumerate(inputs):
-            mask = input_mask[i]
-            tensor_mask = mask == feature_idx
-            if not tensor_mask.any():
+            if i not in tensor_idxs:
                 ablated_inputs.append(input_tensor)
-                current_masks.append(torch.zeros_like(tensor_mask))
+                current_masks.append(None)
                 continue
-            tensor_mask = tensor_mask.to(input_tensor.device).long()
+            tensor_mask = (input_mask[i] == feature_idx).to(input_tensor.device).long()
             baseline = baselines[i] if isinstance(baselines, tuple) else baselines
             if isinstance(baseline, torch.Tensor):
                 baseline = baseline.reshape(
@@ -1173,7 +1196,7 @@ def _process_ablated_out(
     def _process_ablated_out_full(
         self,
         modified_eval: Tensor,
-        current_mask: Tuple[Tensor, ...],
+        current_mask: Tuple[Optional[Tensor], ...],
         flattened_initial_eval: Tensor,
         inputs: TensorOrTupleOfTensorsGeneric,
         n_outputs: int,
@@ -1195,9 +1218,10 @@ def _process_ablated_out_full(
 
         if self.use_weights:
             for weight, mask in zip(weights, current_mask):
-                weight += mask.float().sum(dim=0)
+                if mask is not None:
+                    weight += mask.float().sum(dim=0)
         for i, mask in enumerate(current_mask):
-            if inputs[i].numel() == 0:
+            if mask is None or inputs[i].numel() == 0:
                 continue
             eval_diff = eval_diff.reshape(
                 eval_diff_shape + (inputs[i].dim() - 1) * (1,)

captum/attr/_core/feature_permutation.py

@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 
 # pyre-strict
-from typing import Any, Callable, Optional, Tuple, Union
+from typing import Any, Callable, List, Optional, Tuple, Union
 
 import torch
 from captum._utils.typing import BaselineType, TargetType, TensorOrTupleOfTensorsGeneric
@@ -26,15 +26,15 @@ def _permute_feature(x: Tensor, feature_mask: Tensor) -> Tensor:
 
 
 def _permute_features_across_tensors(
-    inputs: Tuple[Tensor, ...], feature_masks: Tuple[Tensor, ...]
+    inputs: Tuple[Tensor, ...], feature_masks: Tuple[Optional[Tensor], ...]
 ) -> Tuple[Tensor, ...]:
     """
     Permutes features across multiple input tensors using the corresponding
     feature masks.
     """
     permuted_outputs = []
     for input_tensor, feature_mask in zip(inputs, feature_masks):
-        if not feature_mask.any():
+        if feature_mask is None or not feature_mask.any():
             permuted_outputs.append(input_tensor)
             continue
         n = input_tensor.size(0)
@@ -103,7 +103,7 @@ def __init__(
         forward_func: Callable[..., Union[int, float, Tensor, Future[Tensor]]],
         perm_func: Callable[[Tensor, Tensor], Tensor] = _permute_feature,
         perm_func_cross_tensor: Callable[
-            [Tuple[Tensor, ...], Tuple[Tensor, ...]], Tuple[Tensor, ...]
+            [Tuple[Tensor, ...], Tuple[Optional[Tensor], ...]], Tuple[Tensor, ...]
         ] = _permute_features_across_tensors,
     ) -> None:
         r"""
@@ -392,9 +392,14 @@ def _construct_ablated_input_across_tensors(
         input_mask: Tuple[Tensor, ...],
         baselines: BaselineType,
         feature_idx: int,
-    ) -> Tuple[Tuple[Tensor, ...], Tuple[Tensor, ...]]:
-        feature_masks = tuple(
-            (mask == feature_idx).to(inputs[0].device) for mask in input_mask
-        )
+        tensor_idxs: List[int],
+    ) -> Tuple[Tuple[Tensor, ...], Tuple[Optional[Tensor], ...]]:
+        current_masks: List[Optional[Tensor]] = []
+        for i, mask in enumerate(input_mask):
+            if i in tensor_idxs:
+                current_masks.append((mask == feature_idx).to(inputs[0].device))
+            else:
+                current_masks.append(None)
+        feature_masks = tuple(current_masks)
         permuted_outputs = self.perm_func_cross_tensor(inputs, feature_masks)
         return permuted_outputs, feature_masks