I'm trying to read form an CSV where the first four columns are the indexes for a multi-dimensional array. I get the error:
KeyError: 0
from:
sp = []
csvFile = open("sp.csv", "rb")
csvReader = csv.reader(csvFile)
for row in csvReader:
print row
sp[int(row[0])][int(row[1])][int(row[2])][int(row[3])] = float(row[4])
You need to initialize a dictionary at every dimension eg sp[int(row[0])] needs to be assigned to first before you can access it with [int(row[1])]
Edit. Depending on your use case, you may get away with
sp = {}
sp[(int(row[0]), int(row[1]), ..] = float(row[4])
Yet another edit. I was thinking you might use numpy and ended up at this question: Python multi-dimensional array initialization without a loop which actually reflects your problem. It contains a non-numpy solution as the accepted answer. You'd need to know the dimensions for this, though.
defaultdict and avoid having to do all that initialization -- it's called autovivification. See my answer.Instead of an array, you could use a dictionary of dictionaries like this to avoid having to preallocate the entire structure beforehand:
from collections import defaultdict
tree = lambda: defaultdict(tree)
sp = tree()
print 3 in sp[1][2] # -> False
sp[1][2][3] = 4.1
print 3 in sp[1][2] # -> True
print sp[1][2][3] # -> 4.1
sp[9][7][9] = 5.62
sp[4][2][0] = 6.29
sp[1][2][3] returns defaultdict(<function <lambda> at 0x10cf05230>, {})sp[1][2][3] before referencing its contents, so an empty defaultdict (aka a "branch" node) got created automatically by default. This is instead of a KeyError: 3 being raised because the defaultdict in sp[1][2] -- also automatically created -- doesn't have a value for that key.How about using Numpy? sp.csv might look like this:
0,0,0,4.1
1,1,2,5.2
0,1,1,3.2
Then, using Numpy, reading from file become a one-liner:
import numpy as np
sp = np.loadtxt('sp.csv', delimiter=',')
This yields a 2D record array:
array([[ 0. , 0. , 0. , 4.1],
[ 1. , 1. , 2. , 5.2],
[ 0. , 1. , 1. , 3.2]])
Converting this sparse matrix to a full ndarray works like this, assuming 0-based indexing. I'm not happy with the idx= line (there must be a more direct way), but it works:
max_indices = sp.max(0)[:-1]
fl = np.zeros(max_indices + 1)
for row in sp:
idx = tuple(row[:-1].astype(int))
fl[idx] = row[-1]
Resulting in the following ndarray fl:
array([[[ 4.1, 0. , 0. ],
[ 0. , 3.2, 0. ]],
[[ 0. , 0. , 0. ],
[ 0. , 0. , 5.2]]])
