Request: currying `numpy.einsum` for repeated calls with the same subscript string?

Hi,

During my experience with the numpy.einsum function, I discovered that I frequently use it in the following pattern: The zeroth argument to einsum(), the string describing the subscripts, is a fixed one, but the array arguments will vary during runtime. That is to say, I apply the same summation rule to many different array arguments repeatedly.

I don't know if einsum() keeps its subscript-parsing results in a cache, but if not, this could be a bit wasteful, because the same subscript description need to be unnecessarily parsed again and again.

So I wonder if it is possible to have a "curriable" version of einsum that, when called with the subscript description string, returns a callable that has already parsed the string. This returned callable is then ready to be applied to the array arguments.

On the other hand, I'm not sure about the performance impact of parsing. Is this kind of optimization worth the effort?

Just my 2c.

Start at the numpy.einsum entry point and investigate whether repeated calls re-parse the subscript string. Measure the performance impact of parsing across repeated calls, then determine whether a callable curried interface is justified and what tests would demonstrate that it works.