Is there a way to "clear" registered layers?

System information

• TensorFlow.js version (you are using): 4.0.0
• Are you willing to contribute it (Yes/No): Yes

Describe the feature and the current behavior/state. We can register a custom layer with tf.serialization.registerClass.

As far as I can tell, there is no way to deregister a custom layer, or otherwise clear the backend of registered layers.

Will this change the current api? How? Unclear.

Who will benefit with this feature? This seems like a rather niche use case, but I’ll give my specific use case.

I have a model that requires custom layers, and these custom layers are instantiated with some specific parameters.

I unit test a number of model instances in Node.

If I try to test, say, three different model instances with foo, bar, and baz parameters passed to a custom layer, the layer with baz overwrites the rest (as it is instantiated last).

It would be nice if a user could check if any layers have been globally registered, and if so, deregister those layers.

Any Other info.

[pyu10055]

@thekevinscott You can propose the API change first in this thread, it should be more efficient.

[pyu10055]

@mattsoulanille @thekevinscott Thank you for the great discussion, I like the idea of a global registry + model level namespace. The loadOption sounds good to me, it should also apply to custom ops for GraphModel, which is also currently a global registry.