shallow iteration with nditer

问题:

I've got such an array:

>>>y = np.random.randint(0, 255, (2,2,3)) >>>array([[[242, 14, 211],            [198,  7,   0]],            [[235,  60,  81],            [164,  64, 236]]]) 

And I have to iterate over each triplet element (unfortunately vectorization won't help me here...). So I tried:

for i, j in np.nditer(y): print y[i, j], 

hoping I'd get such an output:

[242, 14, 211], [198, 7, 0], [235, 60, 81], [164, 64, 236], but no luck!

I get the error:

Traceback (most recent call last):    File "", line 1, in      for i, j in np.nditer(y):    print y[i,j]  TypeError: iteration over a 0-d array 

I'm quite sure I'm making a very obvious error... can anyone please help me?

回答1:

It looks like you just need to flatten this down a level. You can use the chain operator from itertools.

from itertools import chain       y = np.random.randint(0, 255, (2,2,3) b = chain.from_iterable(y) # where b is a generator 

list(b) output

[array([ 51, 119,  84]),  array([ 50, 110, 193]),  array([165, 157,  52]),  array([239, 119,  83])] 

回答2:

Or reshape y

for i in y.reshape(-1,3):     print i 

A double iteration also works:

for x in y:     for z in x:         print z 

Plain nditer iterates over each element of y (nditer does not give you the indices):

for i in np.nditer(y):     print i        # wrong y[i] 

You'll need to dig more into the flags and documentation for nditer to iterate over 2 of its dimensions. While nditer gives access to the underlying iteration mechanism, you don't usually need to use it - unless you are doing something unusual, or trying to speed up code with cython.

---

Here's an example of getting 2 values from iteration on an nditer object. There is one value for each array in the op list. Both x and z are () shape arrays.

for x,z in np.nditer([y,y]):     print x,z 

There's more on the use of nditer at http://docs.scipy.org/doc/numpy/reference/arrays.nditer.html

---

This doc page has an example using external_loop that dishes out the array in subarrays, rather than individually. I can accomplish the same with the 3d y by reordering its axes:

y3=y.swapaxes(2,0).copy(order='C') for i in np.nditer(y3,order='F',flags=['external_loop']):     print i,  [242  14 211] [198   7   0] [235  60  81] [164  64 236] 

So we can use nditer to do this shallow iteration, but is it worth it?

---

In Iterating over first d axes of numpy array, I stumbled upon ndindex:

for i in np.ndindex(y.shape[:2]):     print y[i], # [242  14 211] [198   7   0] [235  60  81] [164  64 236] 

ndindex uses nditer. The trick to generating shallow iteration is to use a subarray using just the dimensions you want to iterate over.

class ndindex(object):     def __init__(self, *shape):         ...         x = as_strided(_nx.zeros(1), shape=shape, strides=_nx.zeros_like(shape))         self._it = _nx.nditer(x, flags=['multi_index', 'zerosize_ok'], order='C')     def __next__(self):         next(self._it)         return self._it.multi_index 

Or stripping out the essential parts of ndindex I get:

xx = np.zeros(y.shape[:2]) it = np.nditer(xx,flags=['multi_index'])                                while not it.finished:     print y[it.multi_index],     it.iternext() # [242  14 211] [198   7   0] [235  60  81] [164  64 236] 

文章来源: shallow iteration with nditer