将稀疏格式转换为数组numpy

Question

Suppose you have a generator that returns the rows of a table. Something you could use like this:

for (labels, value) in rows:
    pass

"labels" is length n and say it's all strings for simplicity. "value" is something numeric like a float.

Is there a fast, best or built-in way to hash the labels and end up with an n-dimensional array of values and n lists telling you how to map the labels values to indices? I think you could maybe store this in a recarray? I do this all the time but it always ends up being a bit of throw away code. I'd like to find or create something more reusable.

I would be very happy with (('here', 'there', 'nowhere'), 1.234) being mapped to either results['here']['there']['nowhere'] = 1.234 or results[12,3,45] = 1.234 (and having the corresponding lists given the labels down the axes in each dimension.)

I could probably write generator conversion to sparse format and use scipy, but this seems like a nasty way to do something relatively simple.

I see a lot of similar sounding questions but none which exactly answer this question. Maybe I'm missing a search phrase.

Answer 1

You could try using a structured array:

result = np.fromiter(your_generator, dtype=[('labels', '|S10'), ('value', float)])

You'll be able to retrieve a ndarray of labels as result['labels'] (and the values as result['value'] , of course).

Note that you could store several entries in the same label (provided you always have the same number), like in

result = np.array([(('a','b','c'), 1.23),(('a','c','d'), 2.34)],
                  dtype=[('label', ("|S10", 3)), ('value', float)])

where each individual record is given as a tuple, and the three 'labels' of a record as a tuple themselves. You can also name each 'label' individually using a tailored dtype, for example:

 dtype=[('label',[('A','|S10'),('B','|S10'),('C','|S10')]),('value',float)]

This way, you could access all the A through result['label']['A'] ...