Fix hotflip bug where vocab items were not re-encoded correctly (alle…

…nai#4759) * Fix hotflip bug where vocab items were not re-encoded correctly * changelog Co-authored-by: Evan Pete Walsh <[email protected]>
ulmewennberg · Oct 29, 2020 · 812ac57 · 812ac57
1 parent aeb6d36
commit 812ac57
Show file tree

Hide file tree

Showing 3 changed files with 14 additions and 1 deletion.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -22,6 +22,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - Made it possible to instantiate `TrainerCallback` from config files.
 - Fixed the remaining broken internal links in the API docs.
+- Fixed a bug where Hotflip would crash with a model that had multiple TokenIndexers and the input
+  used rare vocabulary items.
 - Fixed a bug where `BeamSearch` would fail if `max_steps` was equal to 1.
 
 ## [v1.2.0rc1](https://github.com/allenai/allennlp/releases/tag/v1.2.0rc1) - 2020-10-22

diff --git a/allennlp/interpret/attackers/hotflip.py b/allennlp/interpret/attackers/hotflip.py
@@ -360,7 +360,7 @@ def _first_order_taylor(self, grad: numpy.ndarray, token_idx: torch.Tensor, sign
             # This happens when we've truncated our fake embedding matrix.  We need to do a dot
             # product with the word vector of the current token; if that token is out of
             # vocabulary for our truncated matrix, we need to run it through the embedding layer.
-            inputs = self._make_embedder_input([self.vocab.get_token_from_index(token_idx)])
+            inputs = self._make_embedder_input([self.vocab.get_token_from_index(token_idx.item())])
             word_embedding = self.embedding_layer(inputs)[0]
         else:
             word_embedding = torch.nn.functional.embedding(

diff --git a/tests/interpret/hotflip_test.py b/tests/interpret/hotflip_test.py
@@ -1,3 +1,5 @@
+import torch
+
 from allennlp.common.testing import AllenNlpTestCase
 from allennlp.data.token_indexers import TokenCharactersIndexer
 from allennlp.interpret.attackers import Hotflip
@@ -49,3 +51,12 @@ def test_with_token_characters_indexer(self):
         assert len(attack["final"][0]) == len(
             attack["original"]
         )  # hotflip replaces words without removing
+
+        # This checks for a bug that arose with a change in the pytorch API.  We want to be sure we
+        # can handle the case where we have to re-encode a vocab item because we didn't save it in
+        # our fake embedding matrix (see Hotflip docstring for more info).
+        hotflipper = Hotflip(predictor, max_tokens=50)
+        hotflipper.initialize()
+        hotflipper._first_order_taylor(
+            grad=torch.rand((10,)).numpy(), token_idx=torch.tensor(60), sign=1
+        )