Reconcile TensorShape functionality with tensorflow's Shape

[knok16]

Tensorflow already provides a class that encapsulates the idea of tensor shape. TensorShape class stores the same data and supports the same mental model. Through the code, one or another is used without a notable difference. What do you think about eliminating TensorShape class and exposing any additional functionality that is currently provided by it in form of extension methods to Shape?

Pros:

• Single class to represent shape (to follow "entities should not be multiplied beyond necessity").

Cons:

• Shape has not-obvious equals implementation: it requires all dimensions to be known to declare shapes equal.

Example: knok16@8fa3a91

[zaleslaw]

The main idea which is not implemented yet: have a separate API layer (a module), the API implementation made with TF.
TensorShape is a draft (or lost building block) of this layer of abstraction. Also, we need a KotlinDLTensor which could be implemented with TFTensor/ONNXTensor/PyTorchTensor and so on.

The initial idea was to avoid working with TF Shape class due to some limitations and unuseful conversion to dimensions

I agree that this is one of the ugliest parts of the internal code, to change it, but I disagree with extension of TF class. Better to make TensorShape more clear to the KotlinDL developer and rewrite carefully all places of usage.

[knok16]

Got it, thanks for the explanation. Agree that an additional layer of abstraction will be helpful in that case. Then what do you think about

1. replacing usages of Shape with TensorShape whenever it is possible(for example in Layer::computeOutputShape (or at least moving it as close to TensorFlow calls as possible;
2. making TensorShape immutable?