last tweak to vat

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "vit-pytorch"
-version = "1.14.2"
+version = "1.14.4"
 description = "Vision Transformer (ViT) - Pytorch"
 readme = { file = "README.md", content-type = "text/markdown" }
 license = { file = "LICENSE" }

vit_pytorch/vat.py

@@ -329,6 +329,7 @@ class VAT(Module):
         extra = None,     # (b d)           - batch, dim extra     
         tasks = None,     # (b)
         actions = None,   # (b k d)         - batch, action chunk length, action dimension
+        return_hiddens = False
     ):
         batch = video_or_image.shape[0]
         return_loss = exists(actions)
@@ -408,6 +409,8 @@ class VAT(Module):
 
         # cross attention
 
+        hiddens = [action_tokens]
+
         for (maybe_film, maybe_self_attn, cross_attn, ff), layer_context in zip(self.layers, context):
 
             if exists(tasks):
@@ -420,6 +423,8 @@ class VAT(Module):
 
             action_tokens = ff(action_tokens) + action_tokens
 
+            hiddens.append(action_tokens)
+
         # maybe unpack extra
 
         if has_extra:
@@ -432,7 +437,10 @@ class VAT(Module):
         pred_action = self.to_pred_action(action_tokens)
 
         if not return_loss:
-            return pred_action
+            if not return_hiddens:
+                return pred_action
+
+            return pred_action, stack(hiddens)
 
         assert pred_action.shape[1] == actions.shape[1]
 
@@ -484,6 +492,6 @@ if __name__ == '__main__':
 
     # after much training
 
-    pred_actions = vat(images, tasks = tasks, extra = extra)
+    pred_actions, hiddens = vat(images, tasks = tasks, extra = extra, return_hiddens = True)
 
     assert pred_actions.shape == (2, 7, 20)