Be explicit that this is config for the transformers package (as these
layers may coexist with other custom stuff in a Keras model, plus the
Keras container itself is called config, and config["config"] is not
great)
Add explicit error handling for initializer calls that have neither
the `config` nor the `transformers_config` argument, or have both.
This was the beginnings of an attempt to address the test failure on
this layer, and instead I backed out of making this layer
keras-serializable at all ... so it was a mistake to commit this.
When supplied by Keras deserialization, the config parameter to initializers
will be a dict. So intercept it and convert to PretrainedConfig object (and
store in instance attribute for get_config to get at it) before passing to the
actual initializer. To accomplish this, and repeat as little code as possible,
use a class decorator on TF*MainLayer classes.
Gunnlaugur Thor Briem