translate upsample_bicubic2d into bilinear layer

Author: jakesabathia2

coremltools/converters/mil/frontend/torch/ops.py

@@ -3937,6 +3937,9 @@ def _translate_torch_args(x, output_size, align_corners, scales) -> Var:
         "upsample_bilinear2d.vec",
         "_upsample_bilinear2d_aa",
         "_upsample_bilinear2d_aa.vec",
+        # Note that in CoreML, there is no bicubic2d layer,
+        # hence we use the bilinear layer for the approximation.
+        "upsample_bicubic2d",
     ],
 )
 def upsample_bilinear2d(context, node):
@@ -4003,6 +4006,13 @@ def _translate_torch_args(x, output_size, align_corners, scales_h, scales_w) ->
                 scales_h, scales_w = scales
         return scales_h, scales_w
 
+    if node.kind == "upsample_bicubic2d":
+        logger.warning(
+            "upsample_bicubic2d is not supported in CoreML. " 
+            "Hence be approximated by the upsample_bilinear layer. "
+            "This could lead to incorrect inference results!"
+        )
+
     x, output_size, align_corners, scales_h, scales_w = _parse_positional_args(context, node)
     scales_h, scales_w = _parse_keyword_args(context, node, scales_h, scales_w)
     scales_h, scales_w = _translate_torch_args(x, output_size, align_corners, scales_h, scales_w)

coremltools/converters/mil/frontend/torch/test/test_torch_ops.py

@@ -2872,6 +2872,33 @@ def forward(self, args):
                 if layer.WhichOneof("layer") == "upsample":
                     assert len(layer.upsample.fractionalScalingFactor) == 0
 
+    @pytest.mark.parametrize(
+        "compute_unit, backend, frontend, scales_h, scales_w",
+        itertools.product(compute_units, backends, frontends, [2, 3, 4.5], [4, 5, 5.5]),
+    )
+    def test_upsample_upsample_bicubic2d_with_scales(
+        self, compute_unit, backend, frontend, scales_h, scales_w
+    ):
+        if backend[0] == "neuralnetwork":
+            if isinstance(scales_h, float) or isinstance(scales_w, float):
+                return  # Skip fractional scale factors tests for neuralnetwork
+
+        input_shape = (1, 3, 10, 10)
+        class Model(nn.Module):
+            def forward(self, x):
+                y = nn.functional.interpolate(
+                    x,
+                    scale_factor=(scales_h, scales_w),
+                    mode="bicubic"
+                )
+                # since we approximate bicubic2d with bilinear,
+                # we only check the output shape.
+                return torch.tensor(y.shape)
+
+        self.run_compare_torch(
+            input_shape, Model().eval(), frontend=frontend, backend=backend, compute_unit=compute_unit
+        )
+
 
 class TestEmpty(TorchBaseTest):
     @pytest.mark.parametrize(