fix

Author: Saurabh Singh

keras/src/backend/openvino/numpy.py

@@ -988,83 +988,69 @@ def logspace(start, stop, num=50, endpoint=True, base=10, dtype=None, axis=0):
     )
 
 
-def maximum(x1, x2):
-    x1 = get_ov_output(x1)
-    x2 = get_ov_output(x2)
-    x1, x2 = _align_operand_types(x1, x2, "maximum()")
-    return OpenVINOKerasTensor(ov_opset.maximum(x1, x2).output(0))
-
-
 def median(x, axis=None, keepdims=False):
     x = get_ov_output(x)
     original_type = x.get_element_type()
     is_bool = original_type == Type.boolean
-
     if is_bool:
         x = ov_opset.convert(x, Type.i32).output(0)
     elif original_type not in (Type.f32, Type.f64):
         x = ov_opset.convert(x, Type.f32).output(0)
-
     if axis is None:
         x = ov_opset.reshape(x, ov_opset.constant([-1], Type.i32).output(0), False).output(0)
         axis = 0
-
     shape = ov_opset.convert(ov_opset.shape_of(x).output(0), Type.i64).output(0)
+    rank = x.get_partial_shape().rank.get_length()
     axis_const = ov_opset.constant([axis], Type.i32).output(0)
     axis_length = ov_opset.reshape(
         ov_opset.gather(shape, axis_const, 0).output(0),
         ov_opset.constant([], Type.i32).output(0),
         False
     ).output(0)
-
     const_zero = ov_opset.constant(0, Type.i64).output(0)
     is_empty = ov_opset.equal(axis_length, const_zero).output(0)
     zero_value = ov_opset.constant(0.0 if not is_bool else 0, x.get_element_type()).output(0)
-
-    result_shape = shape
-    if keepdims:
-        result_shape = ov_opset.select(
-            is_empty,
-            shape,
-            ov_opset.scatter_elements_update(shape, axis_const, ov_opset.constant([1], Type.i64).output(0), 0).output(0)
-        ).output(0)
-    elif axis is None and x.get_partial_shape().rank.get_length() > 1:
-        result_shape = ov_opset.constant([], Type.i32).output(0)
-
+    if axis is None:
+        if keepdims:
+            result_shape = ov_opset.constant([1] * rank, Type.i32).output(0)
+        else:
+            result_shape = ov_opset.constant([], Type.i32).output(0)
+    else:
+        if keepdims:
+            one_i64 = ov_opset.constant([1], Type.i64).output(0)
+            result_shape = ov_opset.scatter_elements_update(shape, axis_const, one_i64, 0).output(0)
+        else:
+            kept_axes = [i for i in range(rank) if i != axis]
+            kept_const = ov_opset.constant(kept_axes, Type.i32).output(0)
+            result_shape = ov_opset.gather(shape, kept_const, 0).output(0)
     empty_result = ov_opset.reshape(zero_value, result_shape, False).output(0)
-
     sorted_values = ov_opset.topk(x, axis_length, axis, "min", "value").output(0)
-
     const_one = ov_opset.constant(1, Type.i64).output(0)
-    is_odd = ov_opset.equal(ov_opset.floor_mod(axis_length, ov_opset.constant(2, Type.i64).output(0)).output(0), const_one).output(0)
-
+    mod_two = ov_opset.floor_mod(axis_length, ov_opset.constant(2, Type.i64).output(0)).output(0)
+    is_odd = ov_opset.equal(mod_two, const_one).output(0)
     half = ov_opset.floor(ov_opset.divide(axis_length, ov_opset.constant(2, Type.i64).output(0)).output(0)).output(0)
     half = ov_opset.convert(half, Type.i64).output(0)
-
     mid_index = ov_opset.convert(half, Type.i32).output(0)
     prev_index = ov_opset.convert(ov_opset.subtract(half, const_one).output(0), Type.i32).output(0)
-
     mid_elem = ov_opset.gather(sorted_values, mid_index, axis).output(0)
     prev_elem = ov_opset.gather(sorted_values, prev_index, axis).output(0)
-
     if is_bool:
         sum_middle = ov_opset.add(mid_elem, prev_elem).output(0)
         is_two = ov_opset.equal(sum_middle, ov_opset.constant(2, Type.i32).output(0)).output(0)
         is_one = ov_opset.equal(sum_middle, ov_opset.constant(1, Type.i32).output(0)).output(0)
-        even_result = ov_opset.select(is_two, ov_opset.constant(1, Type.i32).output(0),
-                                      ov_opset.select(is_one, ov_opset.constant(1, Type.i32).output(0),
-                                                      ov_opset.constant(0, Type.i32).output(0))).output(0)
+        even_result = ov_opset.select(
+            is_two,
+            ov_opset.constant(1, Type.i32).output(0),
+            ov_opset.select(is_one, ov_opset.constant(1, Type.i32).output(0), ov_opset.constant(0, Type.i32).output(0))
+        ).output(0)
     else:
         even_result = ov_opset.divide(
             ov_opset.add(mid_elem, prev_elem).output(0),
             ov_opset.constant(2.0, x.get_element_type()).output(0)
         ).output(0)
-
     median_result = ov_opset.select(is_odd, mid_elem, even_result).output(0)
-
-    if keepdims or (axis is None and x.get_partial_shape().rank.get_length() > 1):
+    if keepdims or (axis is None and rank > 1):
         median_result = ov_opset.reshape(median_result, result_shape, False).output(0)
-
     final_result = ov_opset.select(is_empty, empty_result, median_result).output(0)
     return OpenVINOKerasTensor(ov_opset.convert(final_result, original_type).output(0))