[AL-2368] Optimization for saving shallow views

[FayazRahman]

🚀 🚀 Pull Request

Impact

• Bug fix (non-breaking change which fixes expected existing functionality)
• Enhancement/New feature (adds functionality without impacting existing logic)
• Breaking change (fix or feature that would cause existing functionality to change)

Description

Extending with a numpy array of indices is much faster.

Things to be aware of

Things to worry about

Additional Context

[xiaohanhuang]

Why not handle this inside extend implementation? Also, is the tuple conversion necessary here?

[FayazRahman]

Extend can handle lists, but the extend method has optimizations for when the inputs are numpy arrays.

The tuple conversion is necessary because .indices() is a generator.

[xiaohanhuang]

I meant moving the optimization to extend method so if we have other places call this method, the optimization will be applied automatically.

[FayazRahman]

We would have to loop over all the input samples and check their type to achieve that and it would be better not to have such loops. There might also be issues deciding which dtype to cast the samples to.

[xiaohanhuang]

I assume extend method accept sequence-like object as first argument. Isn't in the implementation just wrap the first argument with np.asarray() would do the same as calling np.array() here? Of course, you still need the code to make sure the input is sequence-like object, like tuple here or list from the original code.

[FayazRahman]

The first argument of extend can be a sequence of a great many things, not just integers. This is why we can't simply wrap it in a np.asarray and leverage the optimizations. We will need to loop over every sample, make sure it is integer, and then cast the whole sequence to a numpy array of proper dtype. But that is not good design.