I am looking for an efficient way of turning
a=np.array([[0,1,1,0], [0,1,1,1]])
into
b=np.array([[0b0110], [0b0111]])
or into
b=np.array([[6], [7]])
The only thing I found so far is np.packbits but this only works with 8bit numbers. However my arrays have shape of around (20e6,20)
I could of course do it by hand:
a=np.array([[0,1,1,0],[0,1,1,1]]) c=2**np.arange(np.shape(a)[-1])[::-1] b=a.dot(c) b Out: array([6, 7])
But I assume that there is a faster way. Especially if the conversion directly to a binary array is possible.
Thanks for help!
Advertisement
Answer
np.packbits
does that. However, the bits of resulting array are stored in block of np.uint8
items. This means that a.shape[1] <= 8
must be true. Moreover, it packs bits on the right (in bigendian mode) so you need to shift the values. Here is the resulting code:
np.packbits(a, bitorder='big', axis=1) >> (8-a.shape[1])
If you need the code to work with a.shape[1] > 8
, then you can write a Numba code based on this previous post.