Fix batched training on right-padded models (#38)

* Take hidden states from last non-padding token when batching using attention_mask --------- Co-authored-by: Theia Vogel <[email protected]>
vgel · Jul 2, 2024 · 9c1c4c2 · 9c1c4c2
1 parent 6d70970
commit 9c1c4c2
Show file tree

Hide file tree

Showing 2 changed files with 23 additions and 11 deletions.
diff --git a/repeng/extract.py b/repeng/extract.py
@@ -276,15 +276,23 @@ def batched_get_hiddens(
     hidden_states = {layer: [] for layer in hidden_layers}
     with torch.no_grad():
         for batch in tqdm.tqdm(batched_inputs):
-            out = model(
-                **tokenizer(batch, padding=True, return_tensors="pt").to(model.device),
-                output_hidden_states=True,
-            )
-            for layer in hidden_layers:
-                # if not indexing from end, account for embedding hiddens
-                hidden_idx = layer + 1 if layer >= 0 else layer
-                for batch in out.hidden_states[hidden_idx]:
-                    hidden_states[layer].append(batch[-1, :].squeeze().cpu().numpy())
+            # get the last token, handling right padding if present
+            encoded_batch = tokenizer(batch, padding=True, return_tensors="pt")
+            encoded_batch = encoded_batch.to(model.device)
+            out = model(**encoded_batch, output_hidden_states=True)
+            attention_mask = encoded_batch["attention_mask"]
+            for i in range(len(batch)):
+                last_non_padding_index = (
+                    attention_mask[i].nonzero(as_tuple=True)[0][-1].item()
+                )
+                for layer in hidden_layers:
+                    hidden_idx = layer + 1 if layer >= 0 else layer
+                    hidden_state = (
+                        out.hidden_states[hidden_idx][i][last_non_padding_index]
+                        .cpu()
+                        .numpy()
+                    )
+                    hidden_states[layer].append(hidden_state)
             del out
 
     return {k: np.vstack(v) for k, v in hidden_states.items()}

diff --git a/repeng/tests.py b/repeng/tests.py
@@ -68,13 +68,17 @@ def gen(vector: ControlVector | None, strength_coeff: float | None = None):
     assert baseline == gen(happy_vector * 0.0)
     assert baseline == gen(happy_vector - happy_vector)
 
-    assert happy == "You are feeling great and happy. I'm"
+    assert happy == "You are feeling a little more relaxed and enjoying"
     # these should be identical
     assert happy == gen(happy_vector, 20.0)
     assert happy == gen(happy_vector * 20)
     assert happy == gen(-(happy_vector * -20))
 
-    assert sad == "You are feeling the worst,\n—("
+    assert sad == 'You are feeling the fucking damn goddamn worst,"'
+    # these should be identical
+    assert sad == gen(happy_vector, -50.0)
+    assert sad == gen(happy_vector * -50)
+    assert sad == gen(-(happy_vector * 50))
 
 
 def test_train_llama_tinystories():