Add openvino backend support for numpy.hstack

keras/src/backend/openvino/excluded_concrete_tests.txt

@@ -24,7 +24,6 @@ NumpyDtypeTest::test_exp2
 NumpyDtypeTest::test_eye
 NumpyDtypeTest::test_flip
 NumpyDtypeTest::test_floor
-NumpyDtypeTest::test_hstack
 NumpyDtypeTest::test_inner
 NumpyDtypeTest::test_isfinite
 NumpyDtypeTest::test_isinf
@@ -87,7 +86,6 @@ NumpyOneInputOpsCorrectnessTest::test_diagonal
 NumpyOneInputOpsCorrectnessTest::test_exp2
 NumpyOneInputOpsCorrectnessTest::test_flip
 NumpyOneInputOpsCorrectnessTest::test_floor_divide
-NumpyOneInputOpsCorrectnessTest::test_hstack
 NumpyOneInputOpsCorrectnessTest::test_imag
 NumpyOneInputOpsCorrectnessTest::test_isfinite
 NumpyOneInputOpsCorrectnessTest::test_isinf

keras/src/backend/openvino/numpy.py

@@ -12,6 +12,7 @@
 from keras.src.backend.openvino.core import (
     align_operand_types as _align_operand_types,
 )
+from keras.src.backend.openvino.core import convert_to_tensor
 from keras.src.backend.openvino.core import get_ov_output
 from keras.src.backend.openvino.core import ov_to_keras_type
 
@@ -846,7 +847,18 @@ def greater_equal(x1, x2):
 
 
 def hstack(xs):
-    raise NotImplementedError("`hstack` is not supported with openvino backend")
+    if not isinstance(xs, (list, tuple)):
+        xs = (xs,)
+    elems = [convert_to_tensor(elem) for elem in xs]
+    element_type = elems[0].output.get_element_type()
+    elems = [get_ov_output(elem, element_type) for elem in elems]
+    is_1d = elems and len(elems[0].get_partial_shape().to_shape()) == 1
+    axis = 0 if is_1d else 1
+    for i in range(1, len(elems)):
+        elems[0], elems[i] = _align_operand_types(
+            elems[0], elems[i], "hstack()"
+        )
+    return OpenVINOKerasTensor(ov_opset.concat(elems, axis).output(0))
 
 
 def identity(n, dtype=None):