Unique random number sampling with Numpy

Question

I need to create a 10,000 x 50 array in which each row contains an ascending series of random numbers between 1 and 365, like so:

[[  4  11  14 ..., 355 360 364]
 [  2  13  15 ..., 356 361 361]
 [  4  12  18 ..., 356 361 365]
 ..., 
 [  6   9  17 ..., 356 362 364]
 [  1  10  19 ..., 352 357 360]
 [  1   9  17 ..., 356 358 364]]

The only way I've figured out to do this is by way of an iterator:

sample_dates = np.array([np.sort(np.random.choice(365, 50, replace=False)) for _ in range(10000)])

which works, but is pretty slow (~0.33 seconds to run) and I'm going to be doing this thousands of times). Is there a faster way to accomplish this?

EDIT: From what I can tell, the most expensive part of this solution is the iteration and 10k individual calls to np.random.choice, not the sorting

Answer 1

The following solution does not use sort:

l = np.array([True]*50 + [False]*315)
total = np.arange(1,366)
sample_dates = np.array([total[np.random.permutation(l)] for _ in range(10000)])

Hence it seems to be faster than the other suggested solutions (takes 0.44 seconds on my computer versus 0.77 for "Nils Werner"'s solution. The OP's solution took 0.81 seconds).

Answer 2

Considering the shapes of the arrays, I thought iterating on columns might provide some improvement. So my idea was to generate 10k numbers - with replacement. Then, on a loop, generate another 10k numbers and check for row-wise duplicates. If there are any, eliminate those and generate that many random numbers. This is also called hit and miss algorithm, if I remember correctly.

Here's the working code:

arr = np.random.choice(365, 10000)
for i in range(49):
    arr2 = np.random.choice(365, 10000)
    comp = (arr2 == arr)
    while comp.any():
        duplicate = comp if i==0 else comp.any(axis=0)
        arr2[duplicate] = np.random.choice(365, duplicate.sum())
        comp = (arr2 == arr)
    arr = np.vstack([arr, arr2])
arr = arr.T
arr.sort(axis=1)

This takes 93.4ms to complete. Your attempt takes 590ms on my computer so it provides ~6x improvement.