Denormalization layer

keras/layers/preprocessing/normalization.py

@@ -61,6 +61,7 @@ class Normalization(base_preprocessing_layer.PreprocessingLayer):
         value(s) will be broadcast to the shape of the kept axes above; if the
         value(s) cannot be broadcast, an error will be raised when this layer's
         `build()` method is called.
+      invert: If True, this layer will return the denormalized values of inputs. Default to False.
 
   Examples:
 
@@ -96,9 +97,32 @@ class Normalization(base_preprocessing_layer.PreprocessingLayer):
   array([[-1.4142135 ],
          [-0.70710677],
          [ 0.        ]], dtype=float32)>
+
+  Using it in the invert manner for denormalizing the inputs with calculating a mean and variance for each index on the last axis.
+
+  >>> adapt_data = np.array([[0., 7., 4.],
+  ...                        [2., 9., 6.],
+  ...                        [0., 7., 4.],
+  ...                        [2., 9., 6.]], dtype='float32')
+  >>> input_data = np.array([[1., 2., 3.]], dtype='float32')
+  >>> layer = tf.keras.layers.Normalization(axis=-1, invert=True)
+  >>> layer.adapt(adapt_data)
+  >>> layer(input_data)
+  <tf.Tensor: shape=(1, 3), dtype=float32, numpy=
+  array([2., 10., 8.], dtype=float32)>
+
+  Using it in the invert manner for denormalizing the inputs with passing the mean and variance directly.
+
+  >>> input_data = np.array([[-1.4142135], [-0.70710677], [0.]], dtype='float32')
+  >>> layer = tf.keras.layers.Normalization(mean=3., variance=2., invert=True)
+  >>> layer(input_data)
+  <tf.Tensor: shape=(3, 1), dtype=float32, numpy=
+  array([[1. ],
+         [2. ],
+         [3. ]], dtype=float32)>  
   """
 
-  def __init__(self, axis=-1, mean=None, variance=None, **kwargs):
+  def __init__(self, axis=-1, mean=None, variance=None, invert=False, **kwargs):
     super().__init__(**kwargs)
     base_preprocessing_layer.keras_kpl_gauge.get_cell('Normalization').set(True)
 
@@ -124,6 +148,7 @@ def __init__(self, axis=-1, mean=None, variance=None, **kwargs):
           'must be set. Got mean: {} and variance: {}'.format(mean, variance))
     self.input_mean = mean
     self.input_variance = variance
+    self.invert = invert
 
   def build(self, input_shape):
     super().build(input_shape)
@@ -302,7 +327,11 @@ def call(self, inputs):
     # The base layer automatically casts floating-point inputs, but we
     # explicitly cast here to also allow integer inputs to be passed
     inputs = tf.cast(inputs, self.compute_dtype)
-    return ((inputs - self.mean) /
+    if self.invert:
+      return ((inputs + self.mean) *
+            tf.maximum(tf.sqrt(self.variance), backend.epsilon()))
+    else:
+      return ((inputs - self.mean) /
             tf.maximum(tf.sqrt(self.variance), backend.epsilon()))
 
   def compute_output_shape(self, input_shape):

keras/layers/preprocessing/normalization_test.py

@@ -198,6 +198,24 @@ def test_output_dtype(self):
     output = layer(input_data)
     self.assertAllEqual(output.dtype, tf.float64)
 
+  def test_invert(self):
+    data = np.array([0., 2., 0., 2.])
+    norm = normalization.Normalization(mean=1.0, variance=1.0)
+    inv_norm = normalization.Normalization(mean=1.0, variance=1.0, invert=True)
+    output = norm(data)
+    output2 = inv_norm(output)
+    self.assertListEqual(output2.shape.as_list(), [4])
+    self.assertAllClose(output2, [0., 2., 0., 2.])
+
+  def test_invert_adapt(self):
+    input_data = [[0.], [2.], [0.], [2.]]
+    norm = keras.layers.Normalization(axis=-1)
+    norm.adapt(input_data)
+    inv_norm = keras.layers.Normalization(axis=-1, invert=True)
+    inv_norm.adapt(input_data)
+    output = norm(input_data)
+    output2 = inv_norm(output)
+    self.assertAllClose(input_data, output2)
 
 @test_combinations.run_all_keras_modes(always_skip_v1=True)
 class NormalizationAdaptTest(test_combinations.TestCase,