Fix KeyError when validation_data was given as a dict

tensorflow/python/keras/engine/training_arrays.py

@@ -35,6 +35,7 @@
 from tensorflow.python.keras.utils.generic_utils import slice_arrays
 from tensorflow.python.keras.utils.mode_keys import ModeKeys
 from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.util import nest
 
 try:
   from scipy.sparse import issparse  # pylint: disable=g-import-not-at-top
@@ -207,7 +208,8 @@ def model_iteration(model,
     val_samples_or_steps = validation_steps
   else:
     # Get num samples for printing.
-    val_samples_or_steps = val_inputs and val_inputs[0].shape[0] or None
+    val_samples_or_steps = val_inputs and nest.flatten(
+        val_inputs)[0].shape[0] or None
 
   if mode == ModeKeys.TRAIN and verbose:
     _print_train_info(num_samples_or_steps, val_samples_or_steps, is_dataset)

tensorflow/python/keras/engine/training_arrays_test.py

@@ -110,6 +110,41 @@ def test_print_info_with_numpy(self, do_validation):
     if do_validation:
       self.assertIn(", validate on 50 samples", mock_stdout.getvalue())
 
+  def test_dict_validation_input(self):
+    """Test case for GitHub issue 30122."""
+    train_input_0 = np.random.rand(1000, 1)
+    train_input_1 = np.random.rand(1000, 1)
+    train_labels = np.random.rand(1000, 1)
+    val_input_0 = np.random.rand(1000, 1)
+    val_input_1 = np.random.rand(1000, 1)
+    val_labels = np.random.rand(1000, 1)
+
+    input_0 = keras.Input(shape=(None,), name='input_0')
+    input_1 = keras.Input(shape=(None,), name='input_1')
+
+    class my_model(keras.Model):
+      def __init__(self):
+        super(my_model, self).__init__(self)
+        self.hidden_layer_0 = keras.layers.Dense(100, activation="relu")
+        self.hidden_layer_1 = keras.layers.Dense(100, activation="relu")
+        self.concat = keras.layers.Concatenate()
+        self.out_layer = keras.layers.Dense(1, activation="sigmoid")
+
+      def call(self, inputs=[input_0, input_1]):
+        activation_0 = self.hidden_layer_0(inputs['input_0'])
+        activation_1 = self.hidden_layer_1(inputs['input_1'])
+        concat = self.concat([activation_0, activation_1])
+        return self.out_layer(concat)
+
+    model = my_model()
+    model.compile(loss="mae", optimizer="adam")
+
+    model.fit(
+        x={'input_0': train_input_0, 'input_1': train_input_1},
+        y=train_labels,
+        validation_data=(
+            {'input_0': val_input_0, 'input_1': val_input_1}, val_labels))
+
 
 if __name__ == "__main__":
   test.main()