François CholletVerified account

These functional API advantages seem like they could also be implemented in a sub-classing style (though they haven’t been yet.) Classes are perfectly capable of holding/serializing/being reconstructed from data structures, including graphs and input shapes. What am I missing?https://twitter.com/fchollet/status/1414667159326715908 …

You mean, implementing a one-off version of the Functional Model class every time you make a new model? Doesn't seem very scalable. If your model is defined in the body of the `call` method, then it's made of bytecode. Turning it into an actual graph will require extra work.

Well, no, I mean that in concept at least, a base class that provides functionality for building layers up into graphs that isn’t defined in the `call` method (even if it runs there) seems possible.

And building up a serializable data structure like a graph of layers doesn’t seem to have an inherent relationship with either functional or sub-classing *styles* (even if it is currently substantially easier to do with one API than the other.)

I mean, maybe I’m wrong here and there is a dependency I’m not thinking of. I’m just trying to peel apart how much the current advantages of the functional API are inherent to the style and how much they’re just a reflection of the current state of the world (which could change.)