Changing something from iterating over a numpy array to vectorization

问题

I am trying to speed up the piece of code below by vectorization:

[rows,cols] = flow_direction_np.shape
elevation_gain = np.zeros((rows,cols), np.float)

for [i, j], flow in np.ndenumerate(flow_direction_np):
    try:
        if flow == 32:
            elevation_gain[i - 1, j - 1]  = elevation_gain[i - 1, j - 1] + sediment_transport_np[i, j]
        elif flow == 64:
            elevation_gain[i - 1, j]  = elevation_gain[i - 1, j] + sediment_transport_np[i, j]
        elif flow == 128:
            elevation_gain[i - 1, j + 1]  = elevation_gain[i - 1, j + 1] + sediment_transport_np[i, j]
        elif flow == 16:
            elevation_gain[i, j - 1]  = elevation_gain[i, j - 1] + sediment_transport_np[i, j]
        elif flow == 1:
            elevation_gain[i, j + 1]  = elevation_gain[i, j + 1] + sediment_transport_np[i, j]
        elif flow == 2:
            elevation_gain[i + 1, j + 1]  = elevation_gain[i + 1, j + 1] + sediment_transport_np[i, j]
        elif flow == 4:
            elevation_gain[i + 1, j]  = elevation_gain[i + 1, j] + sediment_transport_np[i, j]
        elif flow == 8:
            elevation_gain[i + 1, j - 1]  = elevation_gain[i + 1, j - 1] + sediment_transport_np[i, j]
    except IndexError:
            elevation_gain[i, j] = 0

This is how my code looks at the moment:

elevation_gain = np.zeros_like(sediment_transport_np)
nrows, ncols = flow_direction_np.shape
lookup = {32: (-1, -1),
            16:  (0, -1), 
            8:   (+1, -1),
            4:   (+1,  0),
            64: (-1,  0),
            128:(-1,  +1),
            1:   (0,  +1),
            2:   (+1,  +1)}

# Initialize an array for the "shifted" mask
shifted = np.zeros((nrows+2, ncols+2), dtype=bool)

# Pad elevation gain with zeros
tmp = np.zeros((nrows+2, ncols+2), elevation_gain.dtype)
tmp[1:-1, 1:-1] = elevation_gain
elevation_gain = tmp

for value, (row, col) in lookup.iteritems():
    mask = flow_direction_np == value

    # Reset the "shifted" mask
    shifted.fill(False)
    shifted[1:-1, 1:-1] = mask

    # Shift the mask by the right amount for the given value
    shifted = np.roll(shifted, row, 0)
    shifted = np.roll(shifted, col, 1)

    # Set the values in elevation change to the offset value in sed_trans
    elevation_gain[shifted] = elevation_gain[shifted] + sediment_transport_np[mask]

The trouble I am having is they aren't giving me the same result at the end any suggestions where I am going wrong?

回答1:

You can significantly improve your performance using np.where to get the indices where your conditions happen:

ind = np.where( flow_direction_np==32 )

you will see that ind is a tuple with two elements, the first is the indices of the first axis and the second of the second axis of your flow_direction_np array.

You can work out with this indices to apply the shifts: i-1, j-1 and so on...

ind_32 = (ind[0]-1, ind[1]-1)

Then you use fancy indexing to update the arrays:

elevation_gain[ ind_32 ] += sediment_transport_np[ ind ]

EDIT: applying this concept to your case would give something like this:

lookup = {32: (-1, -1),
          16: ( 0, -1),
           8: (+1, -1),
           4: (+1,  0),
          64: (-1,  0),
         128: (-1, +1),
           1: ( 0, +1),
           2: (+1, +1)}

for num, shift in lookup.iteritems():
    ind = np.where( flow_direction_np==num )
    ind_num = ind[0] + shift[0], ind[1] + shift[1]
    elevation_gain[ ind_num] += sediment_transport_np[ ind ]

回答2:

The reason that you're getting different results is due to the way python handles negative indexing.

For other folks reading, this question (and answer) are a follow up from here: Iterating through a numpy array and then indexing a value in another array

First off, I apologize that the "vectorized" code is as dense as it is. There's a through explanation in my earlier answer, so I won't repeat it here.

Your original code (in the original question) is actually subtly different than the version you posted here.

Basically, before you had

for [i, j], flow in np.ndenumerate(flow_direction_np):
    try:
        if flow == 32:
            ...
        elif ...
            ...

and you were getting an index error when i+1 or j+1 was greater than the size of the grid.

Just doing:

for [i, j], flow in np.ndenumerate(flow_direction_np):
    try:
        if flow == 32:
            ...
        elif ...
            ...
    except IndexError:
        elevation_change[i, j] = 0

is actually incorrect because it defines different boundary conditions on different sides of the grid.

In the second case, when j-1 or i-1 is negative, the value from the opposite side of the grid will be returned. However, when j+1 or i+1 is greater than the size of the grid, 0 will be returned. (Thus the "different boundary conditions".)

In the vectorized version of the code, 0 is returned both when the indices are negative and when they're beyond the bounds of the grid.

As a quick example, notice what happens with the following:

In [1]: x = [1, 2, 3]

In [2]: x[0]
Out[2]: 1

In [3]: x[1]
Out[3]: 2

In [4]: x[2]
Out[4]: 3

In [5]: x[3]
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-5-ed224ad0520d> in <module>()
----> 1 x[3]

IndexError: list index out of range

In [6]: x[-1]
Out[6]: 3

In [7]: x[-2]
Out[7]: 2

In [8]: x[-3]
Out[8]: 1

In [9]: x[-4]
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-9-f9c639f21256> in <module>()
----> 1 x[-4]

IndexError: list index out of range

In [10]:

Notice that negative indices up to the size of the sequence are valid and return the "opposite end" of the sequence. So, x[3] raises an error, while x[-1] just returns the other end.

Hopefully that's a bit clearer.

来源：https://stackoverflow.com/questions/17744481/changing-something-from-iterating-over-a-numpy-array-to-vectorization

标签

python

loops

numpy

iterator

vectorization