Document that hub.KerasLayer does not implement precision policy.

This answers issue tensorflow#466

PiperOrigin-RevId: 288445293

docs/common_issues.md

@@ -98,3 +98,11 @@ session.run(init_op)
 ```python
 result = session.run(embedded_text, feed_dict={text_input: ["Hello world"]})
 ```
+
+## Cannot change a model's dype (e.g., float32 to bfloat16)
+
+TensorFlow's SavedModels (shared on TF Hub or otherwise) contain
+operations that work on fixed data types (often, float32 for the weights
+and intermediate activations of neural networks). These cannot be
+changed after the fact when lodaing the SavedModel (but model publishers
+can choose to publish different models with different data types).

tensorflow_hub/keras_layer.py

@@ -84,6 +84,10 @@ class KerasLayer(tf.keras.layers.Layer):
   explicitly from Keras objects instead of relying on graph collections.
   This layer class does not support graph collections.
 
+  Note: The data types used by a saved model have been fixed at saving time.
+  Using tf.keras.mixed_precision etc. has no effect on the saved model
+  that gets loaded by a hub.KerasLayer.
+
   Attributes:
     handle: A callable object (subject to the conventions above), or a
       Python string to load a saved model via hub.load().