add depth anything

Signed-off-by: Phillip Kuznetsov <[email protected]>

Author: philkuz

src/transformers/models/depth_anything/modeling_depth_anything.py

@@ -224,16 +224,16 @@ def forward(self, hidden_states, size=None):
         hidden_states = hidden_states[::-1]
 
         fused_hidden_states = []
-        # first layer only uses the last hidden_state
-        size = hidden_states[1].shape[2:]
-        fused_hidden_state = self.layers[0](hidden_states[0], size=size)
-        fused_hidden_states.append(fused_hidden_state)
+        fused_hidden_state = None
 
-        # looping from the last layer to the second
-        for idx, (hidden_state, layer) in enumerate(zip(hidden_states[1:], self.layers[1:])):
-            size = hidden_states[1:][idx + 1].shape[2:] if idx != (len(hidden_states[1:]) - 1) else None
+        for idx, (hidden_state, layer) in enumerate(zip(hidden_states, self.layers)):
+            size = hidden_states[idx + 1].shape[2:] if idx != (len(hidden_states) - 1) else None
 
-            fused_hidden_state = layer(fused_hidden_state, hidden_state, size=size)
+            if fused_hidden_state is None:
+                # first layer only uses the last hidden_state
+                fused_hidden_state = layer(hidden_state, size=size)
+            else:
+                fused_hidden_state = layer(fused_hidden_state, hidden_state, size=size)
 
             fused_hidden_states.append(fused_hidden_state)
 

tests/models/depth_anything/test_modeling_depth_anything.py

@@ -18,6 +18,7 @@
 
 from transformers import DepthAnythingConfig, Dinov2Config
 from transformers.file_utils import is_torch_available, is_vision_available
+from transformers.pytorch_utils import is_torch_greater_or_equal_than_2_4
 from transformers.testing_utils import require_torch, require_vision, slow, torch_device
 
 from ...test_configuration_common import ConfigTester
@@ -290,3 +291,30 @@ def test_inference(self):
         ).to(torch_device)
 
         self.assertTrue(torch.allclose(predicted_depth[0, :3, :3], expected_slice, atol=1e-4))
+
+    def test_export(self):
+        for strict in [True, False]:
+            with self.subTest(strict=strict):
+                if not is_torch_greater_or_equal_than_2_4:
+                    self.skipTest(reason="This test requires torch >= 2.4 to run.")
+                model = (
+                    DepthAnythingForDepthEstimation.from_pretrained("LiheYoung/depth-anything-small-hf")
+                    .to(torch_device)
+                    .eval()
+                )
+                image_processor = DPTImageProcessor.from_pretrained("LiheYoung/depth-anything-small-hf")
+                image = prepare_img()
+                inputs = image_processor(images=image, return_tensors="pt").to(torch_device)
+
+                exported_program = torch.export.export(
+                    model,
+                    args=(inputs["pixel_values"],),
+                    strict=strict,
+                )
+                with torch.no_grad():
+                    eager_outputs = model(**inputs)
+                    exported_outputs = exported_program.module().forward(inputs["pixel_values"])
+                self.assertEqual(eager_outputs.predicted_depth.shape, exported_outputs.predicted_depth.shape)
+                self.assertTrue(
+                    torch.allclose(eager_outputs.predicted_depth, exported_outputs.predicted_depth, atol=1e-4)
+                )