Standardized / preferred way to implement blocks and models.

[sebastian-sz]

Given the models requirements are being gathered in the discussion, is there a preferred way to implement them?

There are multiple ways, to implement blocks and models:

1. keras.applications way -> Models and blocks are functional Example.
2. Model Garden way -> Model is functional, but blocks are layer subclasses. (even though the model is a direct subclass, it does not override call method).
3. Blocks are layer subclasses, and models are model subclasses of keras.layers.Layer and keras.Model respectively. Both implement call method.

Each way has it's own benefits and drawbacks. Is one of the above preferred? Or maybe something entirely different?

[chjort]

Personally, I would prefer the keras.applications way, where the model is purely functional. Functional models are validated and built during construction. So if a functional models is constructed, the forward pass is guaranteed to work. This also makes model.summary() available right away.

Some applications will require non-standard keras layers, so these can be implemented by subclassing keras.layers.Layer, and then included in the functional model building.

[sebastian-sz]

@chjort Yes, personally I would also prefer keras.applications way. This works well with Tensorflow Model Optimization (I'm not sure about subclassed model's) and the flat structure allows for easy layer modifications.

[LukeWood]

We will be following the keras.applications standard. For example, ResNetBlock will be a functional model.

[sebastian-sz]

@LukeWood thank you, that answer's the question.