Feature request: ability to read heterogeneous types in FortranFile

[astrofrog]

At the moment, the three read_* methods of scipy.io.FortranFile can only read records with homogeneous data types, but it would be nice to allow some more flexibiity, for cases where the data is written e.g. as:

write(10) some_int, some_float

I was thinking that we could add a new method read_mixed that could take a list of dtypes, and returns a Python list with scalars inside.

I could look into it if there was interest in having such a feature in scipy.

[rgommers]

@astrofrog FortranFile is quite new so I don't think it has many users yet, so they may not express their interest here. Sounds like a reasonable addition though. If it's useful for you and can be implemented in a clean way (meaning: it doesn't start to look like the numpy.loadtxt/genfromtxt mess), then I'd say go for it.

[astrofrog]

@rgommers - ok, I'll see what I can do! I won't have time this week, but will look into it at the weekend.

[WarrenWeckesser]

Definitely use the numpy dtype! dtype has quite a few ways to spell a data format, including the following, which generates field names automatically:

In [30]: dtype("i4,d,(3,3)f")
Out[30]: dtype([('f0', '<i4'), ('f1', '<f8'), ('f2', '<f4', (3, 3))])

[astrofrog]

@WarrenWeckesser - thanks! The reason I deleted my comment is because I realized that in fact I could maybe even use the dtype way with the existing read_record and it may work without changes! (then the only thing left to do for this PR would be to improve the docs). I'm testing it out!

[astrofrog]

@WarrenWeckesser @rgommers - ok, so read_record does work with mixed type dtypes from Numpy, so no need for a new method! However, I will open a pull request with an improved docstring for read_record :)

[astrofrog]

I can also add some tests for that mode to make sure if doesn't break in future.

[astrofrog]

@rgommers @WarrenWeckesser - I've now attached code to this issue, but I have one question - as you can see if the tests, if one has an array with shape (2,) then read_record will return a numpy structured array with a length of 1 so that record['f3'] returns a (1,2) array. I could make it so that if the dtype is mixed and matches the size of the record, we slice the structured array with [0] automatically? What do you think?

[astrofrog]

Presumably this doesn't need a changelog entry since there is no change in behavior?

[coveralls]

Changes Unknown when pulling eb17f81 on astrofrog:fortran-mixed-records into * on scipy:master*.

[pv]

LGTM, merged