[SciPy-User] Numpy/MATLAB difference in array indexing

[federico vaggi]

Hi everyone,

I was trying to port some code from MATLAB to Scipy, and I noticed a slight bug in the functionality of numpy.tile vs repmat in matlab:

For example:

a = np.random.rand(10,2)

b = tile(a[:,1],(1,5))

b.shape
Out[86]: (1, 50)

While MATLAB gives:

>>  a = rand(10,2);
>> b = repmat(a(:,1),[1,5]);
>> size(b)

ans =

    10     5

This is obviously trivial to fix**, but I'm wondering what causes the difference?  If you take a vertical slice of an array in numpy that's seen as a row vector, while in MATLAB its seen as a column vector?

Is it worth making a note in here:  http://www.scipy.org/NumPy_for_Matlab_Users ?

Federico

**
The easiest way I found was:

b = tile(a[:,1],(5,1)).T

[Pauli Virtanen]

18.03.2012 23:18, federico vaggi kirjoitti:
> I was trying to port some code from MATLAB to Scipy, and I noticed a
> slight bug in the functionality of numpy.tile vs repmat in matlab:
>
> For example:
>
> a = np.random.rand(10,2)
>
> b = tile(a[:,1],(1,5))

a[:,1] is an 1-d array, and therefore considered as a (1, N) vector in
2-d context. This is not a bug --- the Numpy constructs do not always
map exactly to Matlab ones.

--
Pauli Virtanen

_______________________________________________
SciPy-User mailing list
SciPy...@scipy.org
http://mail.scipy.org/mailman/listinfo/scipy-user

[Wes McKinney]

In [3]: b = tile(a[:,1:2],(1,5))

In [4]: b.shape
Out[4]: (10, 5)

[Sturla Molden]

On 18.03.2012 23:37, Pauli Virtanen wrote:

>> b = tile(a[:,1],(1,5))
>
> a[:,1] is an 1-d array, and therefore considered as a (1, N) vector in
> 2-d context. This is not a bug --- the Numpy constructs do not always
> map exactly to Matlab ones.

Yes.

Also, the need for "repmats" (np.repeat, np.tile, np.hstack, np.vstack)
and "reshapes" (np.reshape, np.ndarray.reshape) is less prominent in
NumPy because of broadcasting. Using MATLAB idioms like reshape and
repmat instead of broadcasting is a common mistake (or bad habit) when
coming to NumPy for MATLAB.

In my experience, 99% of cases for

a .* reshape(b,m,n)
a .* repmat(b,m,n)

in MATLAB will just map to NumPy constructs like these:

a * b
a * b[:,np.newaxis]

This, in addition to view arrays, make NumPy much more memory efficient.

Not to mention that a.T is O(1) in NumPy whereas a' is O(N*M) in MATLAB.

Sturla