Model subclassing compatibility about cvnn HOT 4 OPEN

Did you add the ComplexInput layer at the start? If not TF will automatically use the tensorflow Input and cast it to float.

from cvnn.

How would you do that in a subclassed model? I have a custom train and test step so I can't just do:

inputs = ComplexInput((1,2,3))
outputs = SubClassedModel(inputs)
model = tf.keras.Model(inputs, outputs)

from cvnn.

It is possible... did it work? I have never worked with SubCalssedModels, I don't know how they worked. What is the difference from a normal model?

from cvnn.

I couldn't get it to work. In order to use the approach above, I would have needed too big a refactor. Subclassed models are just a pytorch like interface. You just inherit from the normal model and then build the layers in the constructor, and the implementation in call. It's the same as a custom layer, but with a model instead.

from cvnn.