CODEWITHSUNDEEP
pythonnumpy
rfj001
rfj001

Reputation: 8498

Vectorizing a for loop with Numpy

I have a for loop that I would like to vectorize with numpy. In the below snippet, R, A, and done are numpy arrays of length num_rows, while Q and Q1 are matrices of size (num_rows, num_cols). Also worth noting, all elements of A are between 0 and num_cols - 1, and all elements of done are either 0 or 1. I basically want to do the same thing as the below for-loop, but taking advantage of numpy vectorization.

Important Info:

Here is the loop:

    y = np.zeros((num_rows, num_cols))

    for i in range(num_rows):
        r = R[i]
        a = A[i]
        q = Q[i]
        adjustment = r
        if not done[i]:
            adjustment += (gamma*max(Q1[i]))
        q[a] = adjustment
        y[i, :] = q

I think that I have gotten my "adjustments" in a vectorized way with the following lines, I just need to do the assignment to the Q matrix and output the correct y matrix.

These are the lines that I am using to vectorize the first part:

    q_max_adjustments = np.multiply(gamma * Q1.max(1), done) # This would be numpy array of length num_rows
    true_adjustments = R + q_max_adjustments # Same dimension numpy array

An example input and output would be 

gamma = 0.99
R = numpy.array([1,2,0,3,2])
A = numpy.array([0,2,0,1,1])
done = numpy.array([0,1,0,0,1])
Q = numpy.array([[1,2,3],
                 [4,5,6],
                 [7,8,9],
                 [10,11,12],
                 [13,14,15]])
Q1 = numpy.array([[1,2,3],
                 [4,5,6],
                 [7,8,9],
                 [10,11,12],
                 [13,14,15]])

output y should be array([[ 3.97,  2.  ,  3.  ],
   [ 4.  ,  5.  ,  2.  ],
   [ 8.91,  8.  ,  9.  ],
   [10.  , 14.88, 12.  ],
   [13.  ,  2.  , 15.  ]])

EDIT

So I think that I hacked something together that works, using sparse matrices as masks and such... But it seems like this probably isn't particularly performant, given the number of steps required. Is there a more efficient way to achieve the same goal? Code is below

    q_max_adjustments = np.multiply(gamma * Q1.max(1), 1-done)
    true_adjustments = R + q_max_adjustments
    mask = np.full((num_rows, num_cols), False)
    mask[np.arange(num_rows), A] = True
    value_mask = np.multiply(np.vstack(true_adjustments), mask)
    np.copyto(Q, value_mask, where=mask)

Upvotes: 2

Views: 227

Answers (1)

Paul Panzer
Paul Panzer

Reputation: 53029

Your vectorized solution has all the right elements, but contains a couple of unnecessary complications. A streamlined version using advanced indexing would be:

>>> y = Q.astype(float)
>>> D, = np.where(1-done)
>>> y[np.arange(A.size), A] = R
>>> y[D, A[D]] += gamma * Q1[D].max(axis=1)
>>> y
array([[ 3.97,  2.  ,  3.  ],
       [ 4.  ,  5.  ,  2.  ],
       [ 8.91,  8.  ,  9.  ],
       [10.  , 14.88, 12.  ],
       [13.  ,  2.  , 15.  ]]

Upvotes: 2

Related Questions