[CherryPick r2.3] Fix a critical breakage in training argument default value in inference for layers with a default of training=True called in e.g. a Sequential container.

tensorflow/python/keras/engine/base_layer.py

@@ -2891,9 +2891,9 @@ def _init_call_fn_args(self):
     self._expects_training_arg = ('training' in call_fn_args or
                                   self._call_accepts_kwargs)
     # The default training arg will be any (non-None) default specified in the
-    # method signature, or `False` if no non-None default is specified.
+    # method signature, or None if no value is specified.
     self._default_training_arg = self._call_fn_arg_defaults.get(
-        'training') or False
+        'training')
     self._expects_mask_arg = ('mask' in call_fn_args or
                               self._call_accepts_kwargs)
 

tensorflow/python/keras/engine/base_layer_test.py

@@ -650,6 +650,17 @@ def call(self, inputs, training):
         else:
           return self._nested_layer(inputs) * 0.5
 
+    class CustomLayerDefaultTrainingNone(base_layer.Layer):
+
+      def __init__(self, nested_layer=None):
+        self._nested_layer = nested_layer or array_ops.identity
+
+      def call(self, inputs, training=None):
+        if training:
+          return self._nested_layer(inputs)
+        else:
+          return self._nested_layer(inputs) * 0.5
+
     class CustomLayerDefaultTrainingFalse(base_layer.Layer):
 
       def __init__(self, nested_layer=None):
@@ -701,21 +712,30 @@ def call(self, inputs, training=True):
     # Outer layers/models should set the training context implicitly for all
     # nested layers, respecting whatever mode the outer layer was run with.
     layer = CustomLayerDefaultTrainingTrue(CustomLayerDefaultTrainingFalse())
-    self.assertAllEqual(layer(x), x)
+    # No outer value passed: use local defaults
+    self.assertAllEqual(layer(x), x * 0.25)  # Use local default False
+    # Outer value passed: override local defaults
     self.assertAllEqual(layer(x, training=False), x * 0.25)
     self.assertAllEqual(layer(x, training=True), x)
 
     layer = CustomLayerDefaultTrainingFalse(CustomLayerDefaultTrainingTrue())
-    self.assertAllEqual(layer(x), x * 0.25)
+    # No outer value passed: use local defaults
+    self.assertAllEqual(layer(x), x)  # Use local default True
+    # Outer value passed: override local defaults
     self.assertAllEqual(layer(x, training=False), x * 0.25)
     self.assertAllEqual(layer(x, training=True), x)
 
     # If the outer layer `call` doesn't take a training argument at all,
-    # it'll set the nested scope as inference when no training arg is passed in.
+    # it'll set the nested scope as None when no training arg is passed in.
     # If a training arg is passed in it won't use it directly in `call`, but
     # it will set the nested training mode.
     layer = CustomLayerNoTrainingArg(CustomLayerDefaultTrainingTrue())
-    self.assertAllEqual(layer(x), x * 0.5)
+    self.assertAllEqual(layer(x), x)  # Use local default True
+    self.assertAllEqual(layer(x, training=False), x * 0.5)
+    self.assertAllEqual(layer(x, training=True), x)
+
+    layer = CustomLayerDefaultTrainingNone(CustomLayerDefaultTrainingTrue())
+    self.assertAllEqual(layer(x), x)  # Use local default True
     self.assertAllEqual(layer(x, training=False), x * 0.5)
     self.assertAllEqual(layer(x, training=True), x)
 

tensorflow/python/keras/engine/functional_test.py

@@ -2116,13 +2116,13 @@ def call(self, inputs, training=True):
 
     if context.executing_eagerly():
       # In v2, construction still works when no `training` is specified
-      # When no value passed during construction, it uses the runtime value.
+      # When no value passed during construction, it uses the local default.
       inputs = input_layer_lib.Input(10)
       outputs = my_layer(inputs)
       network = functional.Functional(inputs, outputs)
       self.assertAllEqual(network(x, training=True), _call(x, True))
       self.assertAllEqual(network(x, training=False), _call(x, False))
-      self.assertAllEqual(network(x), _call(x, False))
+      self.assertAllEqual(network(x), _call(x, True))  # Use local default
 
     # `None` value passed positionally during construction is ignored at runtime
     inputs = input_layer_lib.Input(10)
@@ -2131,7 +2131,7 @@ def call(self, inputs, training=True):
     self.assertAllEqual(network(x, training=True), _call(x, True))
     self.assertAllEqual(network(x, training=False), _call(x, False))
     if context.executing_eagerly():
-      self.assertAllEqual(network(x), _call(x, False))
+      self.assertAllEqual(network(x), _call(x, True))  # Use local default
     else:
       # in v1 training would have defaulted to using the `None` inside the layer
       # if training is not passed at runtime
@@ -2144,7 +2144,7 @@ def call(self, inputs, training=True):
     self.assertAllEqual(network(x, training=True), _call(x, True))
     self.assertAllEqual(network(x, training=False), _call(x, False))
     if context.executing_eagerly():
-      self.assertAllEqual(network(x), _call(x, False))
+      self.assertAllEqual(network(x), _call(x, True))  # Use local default
     else:
       # in v1 training would have defaulted to using the `None` inside the layer
       # if training is not passed at runtime

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

@@ -26,6 +26,7 @@
 from tensorflow.python.framework import test_util as tf_test_util
 from tensorflow.python.keras import keras_parameterized
 from tensorflow.python.keras import testing_utils
+from tensorflow.python.keras.engine import sequential
 from tensorflow.python.keras.layers.preprocessing import image_preprocessing
 from tensorflow.python.keras.utils.generic_utils import CustomObjectScope
 from tensorflow.python.ops import gen_stateful_random_ops
@@ -1273,5 +1274,38 @@ def test_config_with_custom_name(self):
     self.assertEqual(layer_1.name, layer.name)
 
 
+@keras_parameterized.run_all_keras_modes(always_skip_v1=True)
+class LearningPhaseTest(keras_parameterized.TestCase):
+
+  def test_plain_call(self):
+    layer = image_preprocessing.RandomWidth(.5, seed=123)
+    shape = (12, 12, 3)
+    img = np.random.random((12,) + shape)
+    out = layer(img)  # Default to training=True
+    self.assertNotEqual(tuple(int(i) for i in out.shape[1:]), shape)
+
+    out = layer(img, training=True)
+    self.assertNotEqual(tuple(int(i) for i in out.shape[1:]), shape)
+
+    out = layer(img, training=False)
+    self.assertEqual(tuple(int(i) for i in out.shape[1:]), shape)
+
+  def test_call_in_container(self):
+    layer1 = image_preprocessing.RandomWidth(.5, seed=123)
+    layer2 = image_preprocessing.RandomHeight(.5, seed=123)
+    seq = sequential.Sequential([layer1, layer2])
+
+    shape = (12, 12, 3)
+    img = np.random.random((12,) + shape)
+    out = seq(img)  # Default to training=True
+    self.assertNotEqual(tuple(int(i) for i in out.shape[1:]), shape)
+
+    out = seq(img, training=True)
+    self.assertNotEqual(tuple(int(i) for i in out.shape[1:]), shape)
+
+    out = seq(img, training=False)
+    self.assertEqual(tuple(int(i) for i in out.shape[1:]), shape)
+
+
 if __name__ == '__main__':
   test.main()