tensor / tensorfs design issues

[rcoreilly]

This discussion is for random tensor and tensorfs design issues.

tensorfs

Working through the basic tensor access API.

The basic workhorse function is:

tsr := tensorfs.Value[float64](dir, "filename", 5, 5)

Key features:

• avoid tricky issues with timing of creation / init vs. access by using one function for both, with type, name, and optional shape sizes.
• don't duplicate the tensor access / setting api -- just get the tensor and call tensor methods on it.
• Get prefix is an anti-pattern.
• also, as usual, want to make things as compact as possible: provide functions for high-frequency basic types.
• one annoying challenge is avoiding the String method for string tensors -- taken by Stringer. tensor.StringValue is one example.

Added type-specific accessors, which is significantly more compact in real code:

• Float64, Float32, Int, StringValue functions on Node

Another question: should table and tensorfs have the same api? A column is effectively the same as a tensorfs Node. Currently, table just has dt.Column("name") that gives you the tensor with row indexes -- don't want / need to be passing create api around here, right?