How to translate matlab code into python (numpy)

Sep 10, 2018 · 6 min read

Just like written languages, different programming languages have different innate styles, which suit different needs. Like Matlab love matrix, love script; python favors tensor, favors subfunction, favors objective-oriented programming; Mathematica are muchly functional; R are better at datatable demonstration.

But if we just want to translate an algorithm from one language to another, we can do so.

Note: An official table of translation could be looked up here. numpy for matlab users .

Objective : If we have a script in matlab, how can we translate that into python. (numpy and matplotlib)

  • General
    • Replace all the comment mark % with #. use '''''' for multiline.
    • Discard all end and corret the indentation by tab
    • Line break ... in mat , \ in python
    • lambda is protected in python, use another var name
  • Operators
    • ^ in matlab = ** in python
    • .* ./ .^ in matlab = default operator in numpy
    • Matrix transpose ' into .T
    • Matrix multiplication * in matlab, . in Mathematica, but not easy in python. Seems for 2 ndarray, A@B do the job in numpy!

  • Indexing

    • Matlab indexing using ( , ), python using [ , ]. Mathematica using [[, ]]

    • Matlab start from 1 end in end, numpy from 0 to ``(nothing)

    • Note one of the super useful thing in matlab is find i.e. find all the none zero entries’ index. Python doesn’t have function with exactly the same functionality.

      • np.non_zero, np.where may be a choice for numpy arrays, but the indices it returns are more complex, you need to extract it from tuple. ii = np.where(values == searchval)[0]

      • [i for i, x in enumerate(my_list) if x == "whatever"] may be a quick way to do so for lists.

  • Number Array

    • Matlab use strt:step:end numpy use range(strt,end,step)
    • And use np.arange for actual array.
    • np.array has a little bit complex control of dimension: all 1-d array (even vertical ones) are neither 1n nor n1 thus has to be reshaped to do matrix multiply!
    • size(), length() in matlab, len() in python
    • [;] can represent break the row in matlab, but cannot do so in python.
    • Append with + or [].append() in Python; list = [list, new_itm] in Matlab

  • Numpy functions

    • min in matlab works element wise, the corresponding function in numpy is minimum not min
    • numpy is more sensitive to data type, sometimes you don’t need to change data type to int to run in matlab. But you have to cast to int to run in python. Esp. when doing indexing!
    • Argument format is quite different for quite a few common functions
      • np.zeros only takes list or tuple of size. if only input one value np.zeros(3) it outputs a 1d array
      • zeros takes both list and individual values, if only input one value zeros(3) it outputs a square mat
    • norm in matlab doesn’t work for vector norm along 1 axis. np.linalg.norm can do this.

  • Dictionary and Struct and Cell

    • In matlab, no dict to use! struct is similar, you can add fields to a struct object
    • struct is also good for structured array data!
      • repmat(struct('image',[],'eig',[]),1,10) can create an array of structure which is easy to
      • Matlab cell is like the python list which is a universal container for any heterogenous things.

  • String and Printing

    • Python string is so easy to use "%s-%d" % (str,int) syntax can perform string formatting anywhere; Matlab has more cumbersome syntax, with several choices
      • sprintf(%s-%d", (str,int))
      • compose(%s-%d", (str,int))
    • Python print function print is so flexible. Matlab counterpart is disp but more for display array and structure instead of string
    • Python print(str) will implicitly add \n to the end to change line. For Matlab, fprintf will not add \n you have to add yourself. But disp will change line automatically

  • Flow Control

    • for i=1:2:10 to for i in range(1,10,2):
    • Note Python add : to any control command; Matlab don’t

  • Common command

    • Print out words: In Python, print("%.f, %.d".%( x, arg2)); In matlab fprintf("x: %.1f, arg1: %d", x, arg2). Or using disp(var)
    • Formatted string control: In python, "%.f, %.d".%( x, arg2), in matlab str = sprintf("x: %.1f, arg1: %d", x, arg2)

  • Subfunction define

    • in matlab
function out=Y(in) 
.... 
end

in python

def Y(param):
    return out

Tricky points

  • Complicated control of ndarray and matrix in numpy
    • Reduce into 1D: M.flat return iterator, M.flatten() return a 1_D array
    • In matlab A(:) will give you a 1d array, but in python A[:] will return the same shape.
    • Aside from reshape there are quite a few handy functions in matlab See this.
  • Array Storage in Memory
    • A subtle difference is the convention of storing multidimensional array in matlab or python. As matlab use column major order, and numpy and torch use row major order. Major order
    • Note row major is the convention of C, Pascal etc. Column major is the convention of Fortran. (Matlab and Julia inherit this) So in python you can specify these 2 conventions by C for row major and F for column major. output=output.reshape((row*col, depth), order='F') (Really confusing….)
    • Because of this, the 2 language can have very different result when applying reshape operation. If you really want to match the result from 2 languages, carefully match your input and then:
      • See which numbers come consecutively on the first dimension.
      • And then permute those axis to left most and then do reshape.
      • For python, you should look at the right most dimension, entries will be stored in that dimension first.
    • See the following corresponding code in matlab vs python
def pixelshuffle(inputs, upscale_factor=3):
    batch_size, channels,  in_height, in_width = inputs.size()
    channels //= upscale_factor ** 2  #*3 if 3D
    #out_depth = in_depth * self.upscale_factor  if 3D
    out_height = in_height * upscale_factor
    out_width = in_width * upscale_factor
    input_view = inputs.contiguous().view(
        batch_size, channels, upscale_factor, upscale_factor, in_height, in_width)
    #shuffle_out = input_view.permute(0, 1, 5, 2, 6, 3, 7, 4).contiguous() if 3D
    shuffle_out = input_view.permute(0, 1, 4, 2, 5, 3).contiguous()
    return shuffle_out.view(batch_size, channels, out_height, out_width)

Corresponds to matlab code

function [out] = pixelshuffle(input, upscale)
    if nargin == 1
        upscale = 2;
    end
    H = size(input,1); W = size(input,2);
    Ch = size(input,3) / upscale.^2;
    B = size(input,4);
    input_fact = reshape(input,[H,W,upscale,upscale,Ch,B]);
    out = reshape(permute(input_fact, [4,1,3,2,5,6]), [H * upscale,W * upscale,Ch,B]);
end

Slide of Reshape function

Let Matlab and Python Talk to Each Other

  • For common non-binary file types, they usually have good readers like
    • csv , python import csv
    • json
    • Which are good for exchange of configurations and string information
  • .mat format could be read in python, with some drawbacks
    • Old simple .mat files could be read by scipy.io.loadmat
    • new .mat files (v7.3+) could be read by h5py.File
      • The matrix and tensor will be read directly as datasets
      • But the objects will be formed in a reference array, and you feed the reference back to the file to get the object into python. See Refs in h5py.
  • .npy format could be read in matlab

For more detailed exposition on this topic, see the other notes Data Transport between Python Matlab, Matlab Python Hybrid Programming

Matlab Python Coding
Binxu Wang
Authors
Binxu Wang
Research Fellow