Squeeze about mir HOT 11 CLOSED

np.squeeze(x, axis=(2,)).shape is equal to x.transposed!2[0]. Would it be OK?

from mir.

You can create PR for squeeze if you want)

from mir.

The problem with squeeze is that you don't know how many dimensions it will return on compile-time, so we need to create a Slice with dynamic dimensions ...
@9il any idea how to do this?

auto squeeze(size_t N, Range)(auto ref Slice!(N, Range) slice)
{
    size_t nonSqueezeDims = 0;
    foreach (i; Iota!(0, N))
        if (slice._lengths[i] > 1)
            nonSqueezeDims += 1;

    Slice!(1, Range) ret; // how to make a slice of a dynamic dimension?
    size_t dims = 0;
    foreach (i; Iota!(0, N))
    {
        if (slice._lengths[i] > 1)
        {
            ret._lengths[dims] = slice._lengths[i];
            ret._strides[dims] = slice._strides[i];
            dims++;
        }
    }
    ret._ptr = slice._ptr;
    return ret;
}
unittest
{
    import mir.ndslice.slice;
    import std.range: iota;
    auto s = iota(3).sliced(1, 3, 1);
    s.squeeze.writeln;
    // [0, 1, 2]
}

from mir.

@9il any idea how to do this?

No idea is good. This is major difference between numpy and ndslice

from mir.

No idea is good. This is major difference between numpy and ndslice

Hmm while compile-time dimensions are really awesome, this seems to be a major limitation that we should think about if we ever want to challenge NumPy (or similar libs).

Pinging reviewers and participants of the merge into phobos for their opinions and ideas:

@John-Colvin @burner @TurkeyMan @nordlow @SFrijters @JackStouffer @klickverbot

from mir.

@greenify This is not very good idea to ping all at once =)

Could you provide a real world example where this feature is required?

from mir.

@absolutelyEveryoneEver I'll look at this tomorrow

from mir.

This is not very good idea to ping all at once

I am sorry - acted to rash. I will remember it. Sorry for spamming you @ pinged people.

Could you provide a real world example where this feature is required?

How you would implement loadtxt or loadmat if you don't know the dimensions before compile-time?

from mir.

How you would implement loadtxt or loadmat if you don't know the dimensions before compile-time?

You always can use 2D matrix

from mir.

Or you can use unions

from mir.

Could you provide a real world example where this feature is required?

I have thought long about this and no I can't.
The best example that I could find is here:

http://northstar-www.dartmouth.edu/doc/idl/html_6.2/REFORM.html

You always can use 2D matrix

Well let's assume you want to load a file from MatLab/R which is unknown to one, e.g.

VarSlice s = loadmat("foo.m"); // loads a 4-dimensional matrix, dimension come from file
s.sum;

Of course one could write a wrapper around such format to determine the dimensions, but ultimately we need to its dimensions and I guess that any application that wants to do more based on dynamic dimensions is rare and well better off to switch to a dynamic language.

Maybe it was just too many years of programming in dynamically typed languages (Python, JavaScript, NodeJs, ...) that I saw this to be problematic.

==> a big sorry to all the people that I spammed, but at least know you know that development for mir is happening here ;-)

Just for complete here's a list of squeeze implementations in dynamic languages that I found during a short search:

Octave:https://www.gnu.org/software/octave/doc/interpreter/Object-Sizes.html
Matlab: https://se.mathworks.com/help/matlab/ref/squeeze.html#bum54rg
R: https://stat.ethz.ch/R-manual/R-devel/library/base/html/drop.html
JS: http://mathjs.org/docs/datatypes/matrices.html
FreeMat: http://freemat.sourceforge.net/help/array_squeeze.html

from mir.