tokenizing from the autoencoder should always no grad and eval mode, …

…readying it for use in multimodal network in some final multimodal repo

meshgpt_pytorch/meshgpt_pytorch.py

@@ -361,8 +361,10 @@ def decode_from_codes_to_faces(
 
         return continuous_coors, pred_face_coords
 
+    @torch.no_grad()
     def tokenize(self, *args, **kwargs):
         assert 'return_codes' not in kwargs
+        self.eval()
         return self.forward(*args, return_codes = True, **kwargs)
 
     @beartype
@@ -545,13 +547,11 @@ def forward_from_raw_face_data(
         face_edges: TensorType['b', 2, 'e', int],
         **kwargs
     ):
-        with torch.no_grad():
-            self.autoencoder.eval()
-            codes = self.autoencoder.tokenize(
-                vertices = vertices,
-                faces = faces,
-                face_edges = face_edges
-            )
+        codes = self.autoencoder.tokenize(
+            vertices = vertices,
+            faces = faces,
+            face_edges = face_edges
+        )
 
         return self.forward(codes, **kwargs)
 

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'meshgpt-pytorch',
   packages = find_packages(exclude=[]),
-  version = '0.0.10',
+  version = '0.0.11',
   license='MIT',
   description = 'MeshGPT Pytorch',
   author = 'Phil Wang',