Clean way to shuffle rows/columns of 2D array?

[matzhaugen]

Is there a clean way to shuffle or sample rows/columns of an Array with/without replacement?

[bluss]

This example should really go somewhere better and bigger, but I guess it applies. Sorting along an axis and shuffling is a similar problem - if you use a permutation like in the example https://github.com/rust-ndarray/ndarray/blob/d2375956a96347e5d39aaf0328cbde121170324b/examples/sort-axis.rs

and for sampling, look in ndarray-rand

[matzhaugen]

Thanks, this is very helpful. sample_axis is what I'm looking for. Would it make sense to implement a version of this method that returns a view instead of copying the data? https://docs.rs/ndarray-rand/0.13.0/ndarray_rand/trait.RandomExt.html#tymethod.sample_axis . This is not essential, but I'm just curious. Often times the resampling is a means to and end which is just to collapse the resampled data onto a single scalar or a lower dimensional space. Something as simple as let resampled_mean = m.sample_axis(Axis(0), n, SamplingMethod::WithReplacement).iter().axis_mean(Axis(0)); . Thoughts?

This might easily be done by implementing a method called select_view on ArrayBase and use it as a replacement here .

A caveat is that this might mess up the stride order and make it slow to iterate over rows/columns.

[bluss]

A sampling view is not possible - views can't be random access like that unfortunately. Views can only "select" rectangular grids from the underlying data.

The illustrations in this article kind of make the same point, see the text that says The “view” sign means that no copying is actually done when slicing an array. and the image above it. The image shows what kinds of rectangular grids there might be.