4

I am trying to make a simulator of incoming messages to output ports. I am given the number of input ports (N) and the number of output ports (M) and a flattened matrix of probabilities of N*M (mat). I am also given the number of incoming messages as an array (in_freq). I figured that in order to direct each message from an input port o an output one I should use numpy.random.choice, but didn't manage to make it work.

what I tried is:

 for k in range (N):  
       enqueue = np.random.choice(M, in_freq[k], p=[(mat[k*N:k*N+M-1])])

when enqueue is an array of the destinations of the messages.

But encountered the message:

ValueError: object too deep for desired array

Mark
  • 90,562
  • 7
  • 108
  • 148
Syoma Dukhan
  • 43
  • 3

1 Answers1

1

The problem

The ValueError that you're seeing is, as hpaulj pointed out in the comments, due to the fact that your p argument is wrapped in list brackets []. This extra enclosing list makes np.random.choice interpret your p as a 2D array of shape (1, x) instead of a 1D array. Since your first argument to np.random.choice is an integer, the function expects p to be 1D, and so throws an error when it sees that it is not (see the choice docs for complete details).

You can fix that ValueError by just removing the brackets around p:

enqueue = np.random.choice(M, in_freq[k], p=mat[k*N:k*N+M-1])

However, now your code will raise a new and different ValueError:

ValueError: a and p must have same size

This ValueError is caused by the fact that the way you're generating the row slices of your flattened NxM matrix mat is not correct.

The solution

From the description in your question, it seems like what you intended was to loop over the rows of the original 2D mat by slicing the 1D mat. Here's how you can fix the slicing in your code so that you iterate over the rows of mat in your for loop:

import numpy as np

N = 10
M = 5

in_freq = np.random.randint(0, 10, N)
mat = np.random.rand(N, M)
# normalize so that each row is a probability dist
mat /= mat.sum(axis=1, keepdims=True)
# flatten to match OP's question
mat = mat.flat

for k in range(N):
    print((M*k, M*(k + 1)))
    enqueue = np.random.choice(M, in_freq[k], p=mat[M*k:M*(k + 1)])

Testing

In order to demonstrate that the correct slices are now being generated, I added a print statement that outputs the slice indices in each iteration of the for loop. Here's the output from that:

(0, 5)
(5, 10)
(10, 15)
(15, 20)
(20, 25)
(25, 30)
(30, 35)
(35, 40)
(40, 45)
(45, 50)

which shows that the slices required to iterate over the rows of mat are indeed being generated as expected now.

Notes about converting a 2D index to its 1D flattened equivalent

Given an array with N rows and M columns, the general formula for converting a 2D x,y index into its flattened 1D equivalent i is:

i = x + M*y

You can find a more in-depth discussion in this old thread.

tel
  • 13,005
  • 2
  • 44
  • 62