speed of array vs matrix
Hello, I have noticed a significant speed difference between the array and the matrix implementation of the dot product, especially for not-so-big matrices. For example: In [1]: import numpy as np In [2]: b = np.random.rand(104,1) In [3]: bm = np.mat(b) In [4]: a = np.random.rand(8, 104) In [5]: am = np.mat(a) In [6]: %timeit np.dot(a, b) 1000000 loops, best of 3: 1.74 us per loop In [7]: %timeit am * bm 100000 loops, best of 3: 6.38 us per loop The results for two different PCs (PC1 with windows/EPD6.2-2 and PC2 with ubuntu/numpy-1.3.0) and two different sizes are below: array matrix 8x104 * 104x1 PC1 1.74us 6.38us PC2 1.23us 5.85us 8x10 * 10x5 PC1 2.38us 7.55us PC2 1.56us 6.01us For bigger matrices the timings seem to asymptotically approach. Is it something worth trying to fix or should I just accept this as a fact and, when working with small matrices, stick to array? Thanks, Luca
On Mon, Oct 25, 2010 at 7:48 AM, Citi, Luca <lciti@essex.ac.uk> wrote:
Hello, I have noticed a significant speed difference between the array and the matrix implementation of the dot product, especially for not-so-big matrices. For example:
In [1]: import numpy as np In [2]: b = np.random.rand(104,1) In [3]: bm = np.mat(b) In [4]: a = np.random.rand(8, 104) In [5]: am = np.mat(a) In [6]: %timeit np.dot(a, b) 1000000 loops, best of 3: 1.74 us per loop In [7]: %timeit am * bm 100000 loops, best of 3: 6.38 us per loop
The results for two different PCs (PC1 with windows/EPD6.2-2 and PC2 with ubuntu/numpy-1.3.0) and two different sizes are below:
array matrix
8x104 * 104x1 PC1 1.74us 6.38us PC2 1.23us 5.85us
8x10 * 10x5 PC1 2.38us 7.55us PC2 1.56us 6.01us
For bigger matrices the timings seem to asymptotically approach.
Is it something worth trying to fix or should I just accept this as a fact and, when working with small matrices, stick to array?
Looks like call overhead, it will become significant when the dot product itself doesn't take much time. If you want to spend time profiling it might be possible to find some spots in the matrix class that can be sped up. My guess is that it is the time taken to create the matrices. But if you don't want to get sidetracked sticking to arrays is the easy way to go. Chuck
On Mon, Oct 25, 2010 at 6:48 AM, Citi, Luca <lciti@essex.ac.uk> wrote:
Hello, I have noticed a significant speed difference between the array and the matrix implementation of the dot product, especially for not-so-big matrices. For example:
In [1]: import numpy as np In [2]: b = np.random.rand(104,1) In [3]: bm = np.mat(b) In [4]: a = np.random.rand(8, 104) In [5]: am = np.mat(a) In [6]: %timeit np.dot(a, b) 1000000 loops, best of 3: 1.74 us per loop In [7]: %timeit am * bm 100000 loops, best of 3: 6.38 us per loop
The results for two different PCs (PC1 with windows/EPD6.2-2 and PC2 with ubuntu/numpy-1.3.0) and two different sizes are below:
array matrix
8x104 * 104x1 PC1 1.74us 6.38us PC2 1.23us 5.85us
8x10 * 10x5 PC1 2.38us 7.55us PC2 1.56us 6.01us
For bigger matrices the timings seem to asymptotically approach.
Is it something worth trying to fix or should I just accept this as a fact and, when working with small matrices, stick to array?
I think the fixed overhead comes from the subclassing of arrays. The subclassing is done in Python and if an operation creates a matrix then __array_finalize__ is called. All that adds up to overhead. http://github.com/numpy/numpy/blob/master/numpy/matrixlib/defmatrix.py I wrote a mean-variance optimizer with matrices. Switching to arrays gave me a big speed up.
Thank you for your replies. I might just use arrays. Maybe in the future something like pypy might help reduce the python overhead. Best, Luca
participants (3)
-
Charles R Harris -
Citi, Luca -
Keith Goodman