chore: adding a wrapper to numpy sum, started fixing layer tests

keras_core/backend/numpy/numpy.py

@@ -560,7 +560,12 @@ def var(x, axis=None, keepdims=False):
 
 def sum(x, axis=None, keepdims=False):
     axis = tuple(axis) if isinstance(axis, list) else axis
-    return np.sum(x, axis=axis, keepdims=keepdims)
+    # `np.sum()` function, when applied to an array, returns a scalar
+    # value that represents the sum of all the elements in the array.
+    # This operation reduces the dimensionality of the input and hence
+    # return objects of type `np.<DTYPE>`. To fix this behaviour, we
+    # wrap the result in a list and explicitly reutrns an `np.ndarray`.
+    return np.array([np.sum(x, axis=axis, keepdims=keepdims)])
 
 
 def eye(N, M=None, k=0, dtype="float32"):

keras_core/layers/layer_test.py

@@ -264,6 +264,10 @@ def call(self, x):
         layer(layers.Input(batch_shape=(2, 2)))
         self.assertLen(layer.losses, 0)
 
+    @pytest.mark.skipif(
+        backend.backend() == "numpy",
+        reason="Trainer not implemented for NumPy backend.",
+    )
     def test_add_loss(self):
         class LossLayer(layers.Layer):
             def call(self, x):
@@ -378,6 +382,10 @@ def test_mixed_precision(self):
         self.assertEqual(backend.standardize_dtype(y.dtype), "float16")
         self.assertEqual(layer.kernel.dtype, "float32")
 
+    @pytest.mark.skipif(
+        backend.backend() == "numpy",
+        reason="Numpy backend does not support masking.",
+    )
     def test_masking(self):
         class BasicMaskedLayer(layers.Layer):
             def __init__(self):