Fix stddev 0 bug in Normalization layer

PiperOrigin-RevId: 322836694
Change-Id: I22a669e19f369cba271e56b63a08ca2763a6eab8

tensorflow/python/keras/layers/preprocessing/normalization.py

@@ -156,7 +156,8 @@ class Normalization(base_preprocessing_layer.CombinerPreprocessingLayer):
     # broadcasts the data correctly.
     mean = array_ops.reshape(self.mean, self._broadcast_shape)
     variance = array_ops.reshape(self.variance, self._broadcast_shape)
-    return (inputs - mean) / math_ops.sqrt(variance)
+    return ((inputs - mean) /
+            math_ops.maximum(math_ops.sqrt(variance), K.epsilon()))
 
   def compute_output_shape(self, input_shape):
     return input_shape

tensorflow/python/keras/layers/preprocessing/normalization_test.py

@@ -97,6 +97,16 @@ def _get_layer_computation_test_cases():
               np.float32),
       "testcase_name":
           "3d_multiple_axis"
+  }, {
+      "adapt_data":
+          np.zeros((3, 4)),
+      "axis": -1,
+      "test_data":
+          np.zeros((3, 4)),
+      "expected":
+          np.zeros((3, 4)),
+      "testcase_name":
+          "zero_variance"
   })
 
   crossed_test_cases = []