Fix ViT-MAE decoder interpolate

src/transformers/models/vit_mae/modeling_vit_mae.py

@@ -841,21 +841,20 @@ def __init__(self, config, num_patches):
 
     def interpolate_pos_encoding(self, embeddings: torch.Tensor) -> torch.Tensor:
         """
-        This method is a modified version of the interpolation function for ViT-mae model at the deocder, that
+        This method is a modified version of the interpolation function for ViT-mae model at the decoder, that
         allows to interpolate the pre-trained decoder position encodings, to be able to use the model on higher
         resolution images.
 
-        Source:
+        Adapted from:
         https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
         """
 
         # -1 removes the class dimension since we later append it without interpolation
         embeddings_positions = embeddings.shape[1] - 1
-        num_positions = self.decoder_pos_embed.shape[1] - 1
 
         # Separation of class token and patch tokens
-        class_pos_embed = self.decoder_pos_embed[:, 0, :]
-        patch_pos_embed = self.decoder_pos_embed[:, 1:, :]
+        class_pos_embed = self.decoder_pos_embed[:, :1]
+        patch_pos_embed = self.decoder_pos_embed[:, 1:]
 
         # To retain the final 3d tensor with the required dimensions
         dim = self.decoder_pos_embed.shape[-1]
@@ -867,18 +866,18 @@ def interpolate_pos_encoding(self, embeddings: torch.Tensor) -> torch.Tensor:
         patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
 
         # Interpolating the decoder position embeddings shape wrt embeddings shape i.e (x).
-        # 1 keeps the other dimension constant
+        # we keep the second last dimension constant
         patch_pos_embed = nn.functional.interpolate(
             patch_pos_embed,
-            scale_factor=(1, embeddings_positions / num_positions),
+            size=(patch_pos_embed.shape[-2], embeddings_positions),
             mode="bicubic",
             align_corners=False,
         )
 
         # Converting back to the original shape
         patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
         # Adding the class token back
-        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)
+        return torch.cat((class_pos_embed, patch_pos_embed), dim=1)
 
     def initialize_weights(self, num_patches):
         # initialize (and freeze) position embeddings by sin-cos embedding

tests/models/vit_mae/test_modeling_vit_mae.py

@@ -298,12 +298,16 @@ class ViTMAEModelIntegrationTest(unittest.TestCase):
     def default_image_processor(self):
         return ViTImageProcessor.from_pretrained("facebook/vit-mae-base")
 
+    @cached_property
+    def default_model(self):
+        return ViTMAEForPreTraining.from_pretrained("facebook/vit-mae-base").to(torch_device)
+
     @slow
     def test_inference_for_pretraining(self):
         # make random mask reproducible across the PT and TF model
         np.random.seed(2)
 
-        model = ViTMAEForPreTraining.from_pretrained("facebook/vit-mae-base").to(torch_device)
+        model = self.default_model
 
         image_processor = self.default_image_processor
         image = prepare_img()
@@ -313,11 +317,11 @@ def test_inference_for_pretraining(self):
         # (this way we can ensure that the PT and TF models operate on the same inputs)
         vit_mae_config = ViTMAEConfig()
         num_patches = int((vit_mae_config.image_size // vit_mae_config.patch_size) ** 2)
-        noise = np.random.uniform(size=(1, num_patches))
+        noise = torch.from_numpy(np.random.uniform(size=(1, num_patches))).to(device=torch_device)
 
         # forward pass
         with torch.no_grad():
-            outputs = model(**inputs, noise=torch.from_numpy(noise).to(device=torch_device))
+            outputs = model(**inputs, noise=noise)
 
         # verify the logits
         expected_shape = torch.Size((1, 196, 768))
@@ -339,7 +343,7 @@ def test_inference_interpolate_pos_encoding(self):
         # make random mask reproducible across the PT and TF model
         np.random.seed(2)
 
-        model = ViTMAEForPreTraining.from_pretrained("facebook/vit-mae-base").to(torch_device)
+        model = self.default_model
 
         image_processor = self.default_image_processor
         image = prepare_img()
@@ -349,14 +353,38 @@ def test_inference_interpolate_pos_encoding(self):
         # (this way we can ensure that the PT and TF models operate on the same inputs)
         vit_mae_config = ViTMAEConfig()
         num_patches = (image.height // vit_mae_config.patch_size) * (image.width // vit_mae_config.patch_size)
-        noise = np.random.uniform(size=(1, num_patches))
+        noise = torch.from_numpy(np.random.uniform(size=(1, num_patches))).to(device=torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs, noise=noise, interpolate_pos_encoding=True)
+
+        # verify the logits
+        expected_shape = torch.Size((1, 1200, 768))
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+    @slow
+    def test_inference_interpolate_pos_encoding_custom_sizes(self):
+        # Ensure custom sizes are correctly handled when interpolating the position embeddings
+
+        # make random mask reproducible across the PT and TF model
+        np.random.seed(2)
+
+        model = self.default_model
+        image_processor = self.default_image_processor
+
+        image = prepare_img()
+        inputs = image_processor(images=image, return_tensors="pt", size={"height": 256, "width": 256}).to(
+            torch_device
+        )
 
         # forward pass
         with torch.no_grad():
             outputs = model(
-                **inputs, noise=torch.from_numpy(noise).to(device=torch_device), interpolate_pos_encoding=True
+                **inputs,
+                interpolate_pos_encoding=True,
             )
 
         # verify the logits
-        expected_shape = torch.Size((1, 1200, 768))
+        expected_shape = torch.Size((1, 256, 768))
         self.assertEqual(outputs.logits.shape, expected_shape)