adhoc change layers of existing model

[LeonhardFeiner]

In TensorFlow, there exists tf.keras.models.clone_model which together with the clone_function can be used to ad-hoc change/replace layers of an existing model object.
https://www.tensorflow.org/api_docs/python/tf/keras/models/clone_model

Do you have an idea, how I could generically replace every e.g. conv layer in a model?

I looked into the implementation and in my opinion the model's submodules are initialized mostly on the fly while calling nn.compact and therefore don't exist in the model object at all before calling / compiling the model.

[cgarciae]

Hey @LeonhardFeiner, as you point out there is no good way to do this, even setup is called lazily so you don't even have access to sub-modules defined there from the outside. This probably means that for now you probably have to fork the whole model architecture and make changes yourself.