Proper way to read/write fortran order array with numpy

Question

I'm writing an application that reads an ascii file with fortran-ordered arrays, modifies values, then writes that data back out (in fortran order) in ascii. What is the correct way to read this array into numpy, denote the array is in fortran order, then write the data back out in fortran order?

Assume I have a file with the following ascii text:

0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0

These numbers represent a 2x2x2 array written in fortran order.

The ascii format is a bit more complicated than above. However, suffice it to say the format is not really well suited for using any of the automatic numpy ascii loaders like numpy.loadtxt, etc.

I'm executing a line similar to the following to create the array:

x = numpy.array([0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0], order='F')

I know this is in-efficient and does a lot of extra data copying, etc. I'm more worried about the ordering though.

So, at this point x is ordered in memory like a fortran array, I think. Now, when I export this array should I use numpy.nditer(x, order='F')?

If x is already ordered like a Fortran array, order is not needed in nditer (which, by default, iterates in memory order). — nneonneo
– nneonneo, Commented Jul 22, 2013 at 20:04

NPE · Accepted Answer · 2013-07-22 20:13:55Z

1

Consider the following:

In [11]: x = numpy.array([0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0])

In [12]: xf = x.reshape((2, 2, 2), order='F')

Here, xf is a 2x2x2 Fortran-ordered view onto x. You could modify it and x will change accordingly:

In [22]: xf[0,:,1] = [11, 12]

In [23]: x
Out[23]: array([  0.,   1.,   2.,   3.,  11.,   5.,  12.,   7.])

Exporting x would preserve the original ordering.

answered Jul 22, 2013 at 20:13

NPE

503k114 gold badges970 silver badges1k bronze badges

Sign up to request clarification or add additional context in comments.

3 Comments

durden2.0 Over a year ago

This looks correct. My main problem was I didn't realize doing x.shape = (2,2,2) actually changes the order. I created it with order 'F'. However, when I first created the array it was 1-dimensional so the order doesn't take effect. Thus, changing the shape with .shape = (2,2,2) didn't maintain the fortran ordering.

durden2.0 Over a year ago

Is there a way to create an array with a shape AND order at the same time? I don't see this in the API. Probably not a huge deal because I'll just be using the view everywhere, not the original array. So, the data isn't copied twice in this case, just once when creating it and copying from python list.

Sven Over a year ago

Don't use numpy tofile though. It always writes C order no matter what order the array is in memory (see manual).

Sven · Accepted Answer · 2016-02-17 14:54:42Z

1

Here is what works. The transposition works around the fact that numpy tofile() only knows C order.

import numpy as np
file_path = 'C:/Temp/arr.txt'
x = np.array([0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0])
xf = x.reshape((2, 2, 2), order='F')
xf.T.tofile(file_path, sep=' ') # T is transpose, so it gets written in correct order
xn = np.fromfile(file_path, sep=' ')
xnr = np.reshape(xn, (2,2,2),order='F')
assert (xf==xnr).all()

# show the transpose is necessary
xf.tofile(file_path, sep=' ') # no .T, write in wrong order
xn = np.fromfile(file_path, sep=' ')
xnr = np.reshape(xn, (2,2,2),order='F')
assert (xf==xnr).all()==False

answered Feb 17, 2016 at 14:54

Sven

9951 gold badge11 silver badges29 bronze badges

Collectives™ on Stack Overflow

Proper way to read/write fortran order array with numpy

2 Answers 2

3 Comments

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

3 Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Related