具有多种条件的优雅 2D numpy 索引

Question

I have an array of numbers that I wish to turn into dummy variables (ie arrays with 1 if condition is met, 0 if otherwise). However, the conditions can be numerous and I was wondering if there was a more elegant solution than what I'm using.

arr = np.random.randint(0, 50, size=(100, 100))

# What I'm doing

dummy = np.zeros(arr.shape)
dummy[np.where(np.logical_or.reduce((arr== 10, arr== 15, arr==16, arr==17)))] = 1

In the example, every value that is 10, 15, 16, or 17 becomes a one else a zero. For some dummy variables I have 10+ conditions and the expression can get lengthy, so I'm looking for something cleaner. I tried something like this but got a ValueError.

dummy= [1 if x in [10, 15, 16, 17] else 0 for x in arr]

Answer 1

您可以使用np.select

arr = np.select([arr==10, arr==15,arr==16,arr==17],[1,1,1,1],0)

Answer 2

Readability is a bit in the eye of the beholder, but two different ways to use the 1 in ... else 0 construct is to use a double for loop over the elements of the rows of the matrix as

dummy = [[1 if x in [10, 15, 16, 17] else 0 for x in row] for row in arr]

The output of this is not an np.array (matrix) though but rather a list of list s. Another way to do it, which "hides" the double for loop is to use np.vectorize as

dummy_func = np.vectorize(lambda x: 1 if x in [10, 15, 16, 17] else 0)
dummy = dummy_func(arr)

or as a one-liner as

dummy = np.vectorize(lambda x: 1 if x in [10, 15, 16, 17] else 0)(arr)

Of these I would probably go for the vectorized approach to keep the data type as an np.array as this is most often a more reasonable choice. And, even if I showed it as possible one-liner, I would still think it would be better to first define the function and then apply it on two different lines.

It should be noted though, that vectorize is basically just a double for loop, so the execution is rather slow compared to other numpy functions. I wouldn't be surprised if there are other ways as well, that might be able to use the built-in parallel computational behavior of numpy but then it is again a trade off between readability (intent) and speed.