Handle trainable and name in the backbone base class #680
Conversation
| @@ -27,6 +27,12 @@ class Backbone(keras.Model): | |||
| def __init__(self, *args, **kwargs): | |||
| super().__init__(*args, **kwargs) | |||
|
|
|||
| def get_config(self): | |||
There was a problem hiding this comment.
Q: Couldn't you technically chain all the way up to keras.Model for this?
There was a problem hiding this comment.
We cannot because of a pain point with functional subclassed models.
By default get_config on a functional model will return a huge nested description of all layers, that can only be used with from_config for the function keras model class itself. There's no way to use the standard functional model get_config and from_config and still get back a BertBackbone, you would just get back a vanilla functional model (with no special properties, methods, docstring etc).
So essentially for functional subclasses, you have to "break the chain" of get_config and from_config chaining to super, which is unusual. Everywhere else in Keras you want to do this! But at least we are able to hide this pain point from our downstream model implementation now. And potentially we could prioritize some work in core Keras at some point to improve the "functional subclass" experience and remove this weird gotcha.
I can leave a (much shorter than this), comment explaining this on the backbone class.
There was a problem hiding this comment.
Thanks for the explanation! Lifelong learner 🤓
No description provided.