Refactor ops.core tests and update the backend. (#21466)

Author: james77777778

keras/src/backend/openvino/core.py

@@ -584,17 +584,6 @@ def _is_scalar(elem):
     return not isinstance(elem, (list, tuple, set, dict))
 
 
-def _get_first_element(x):
-    if isinstance(x, (tuple, list)):
-        for elem_in_x in x:
-            elem = _get_first_element(elem_in_x)
-            if elem is not None:
-                return elem
-    elif _is_scalar(x):
-        return x
-    return None
-
-
 def convert_to_tensor(x, dtype=None, sparse=None, ragged=None):
     if sparse:
         raise ValueError("`sparse=True` is not supported with openvino backend")
@@ -603,24 +592,29 @@ def convert_to_tensor(x, dtype=None, sparse=None, ragged=None):
     if dtype is not None:
         dtype = standardize_dtype(dtype)
     if isinstance(x, OpenVINOKerasTensor):
+        if dtype and dtype != standardize_dtype(x.dtype):
+            x = cast(x, dtype)
         return x
     elif isinstance(x, np.ndarray):
         if dtype is not None:
             ov_type = OPENVINO_DTYPES[dtype]
-            return OpenVINOKerasTensor(ov_opset.constant(x, ov_type).output(0))
-        return OpenVINOKerasTensor(ov_opset.constant(x).output(0))
+        else:
+            ov_type = OPENVINO_DTYPES[standardize_dtype(x.dtype)]
+        return OpenVINOKerasTensor(ov_opset.constant(x, ov_type).output(0))
     elif isinstance(x, (list, tuple)):
         if dtype is None:
-            # try to properly deduce element type
-            elem = _get_first_element(x)
-            if isinstance(elem, float):
-                dtype = "float32"
-            elif isinstance(elem, int):
-                dtype = "int32"
+            dtype = result_type(
+                *[
+                    getattr(item, "dtype", type(item))
+                    for item in tree.flatten(x)
+                ]
+            )
         x = np.array(x, dtype=dtype)
-        return OpenVINOKerasTensor(ov_opset.constant(x).output(0), x)
+        ov_type = OPENVINO_DTYPES[dtype]
+        return OpenVINOKerasTensor(ov_opset.constant(x, ov_type).output(0), x)
     elif isinstance(x, (float, int, bool)):
-        dtype = standardize_dtype(dtype)
+        if dtype is None:
+            dtype = standardize_dtype(type(x))
         ov_type = OPENVINO_DTYPES[dtype]
         return OpenVINOKerasTensor(ov_opset.constant(x, ov_type).output(0), x)
     if isinstance(x, Variable):

keras/src/backend/openvino/excluded_concrete_tests.txt

@@ -186,13 +186,3 @@ CoreOpsCorrectnessTest::test_slice_update
 CoreOpsCorrectnessTest::test_switch 
 CoreOpsCorrectnessTest::test_unstack
 CoreOpsCorrectnessTest::test_vectorized_map
-CoreOpsDtypeTest::test_convert_to_tensor0
-CoreOpsDtypeTest::test_convert_to_tensor1
-CoreOpsDtypeTest::test_convert_to_tensor2
-CoreOpsDtypeTest::test_convert_to_tensor3
-CoreOpsDtypeTest::test_convert_to_tensor8
-CoreOpsDtypeTest::test_convert_to_tensor11
-CoreOpsDtypeTest::test_convert_to_tensor12
-CoreOpsDtypeTest::test_convert_to_tensor14
-CoreOpsDtypeTest::test_convert_to_tensor25
-CoreOpsDtypeTest::test_convert_to_tensor37

keras/src/backend/torch/core.py

@@ -575,7 +575,7 @@ def scatter(indices, values, shape):
 def scatter_update(inputs, indices, updates):
     inputs = convert_to_tensor(inputs)
     indices = convert_to_tensor(indices, dtype="int64")
-    updates = convert_to_tensor(updates)
+    updates = convert_to_tensor(updates, dtype=inputs.dtype)
     indices = torch.transpose(indices, 0, 1)
 
     outputs = torch.clone(inputs)

keras/src/ops/core.py

@@ -542,7 +542,9 @@ def call(self, loop_vars):
         )
 
     def compute_output_spec(self, loop_vars):
-        return [KerasTensor(v.shape, dtype=v.dtype) for v in loop_vars]
+        return tree.map_structure(
+            lambda v: KerasTensor(v.shape, dtype=v.dtype), loop_vars
+        )
 
 
 @keras_export("keras.ops.while_loop")
@@ -587,6 +589,10 @@ def while_loop(
     >>> keras.ops.while_loop(cond, body, (x, y))
     10, 11
     """
+    if any_symbolic_tensors((loop_vars,)):
+        return WhileLoop(
+            cond, body, maximum_iterations=maximum_iterations
+        ).symbolic_call(loop_vars)
     return backend.core.while_loop(
         cond,
         body,
@@ -808,8 +814,6 @@ def cast(x, dtype):
     >>> x = keras.ops.arange(4)
     >>> x = keras.ops.cast(x, dtype="float16")
     """
-    dtype = backend.standardize_dtype(dtype)
-
     if any_symbolic_tensors((x,)):
         return Cast(dtype=dtype)(x)
     return backend.core.cast(x, dtype)
@@ -874,8 +878,6 @@ def saturate_cast(x, dtype):
     >>> #  [255 255 255 255]]
 
     """
-    dtype = backend.standardize_dtype(dtype)
-
     if any_symbolic_tensors((x,)):
         return SaturateCast(dtype=dtype)(x)
     return _saturate_cast(x, dtype)