Efficiently counting number of unique elements - NumPy / Python

Question

When running np.unique(), it first flattens the array, sorts the array, then finds the unique values. When I have arrays with shape (10, 3000, 3000), it takes about a second

Answer 1

We could leverage the fact that the elements are restricted to uint8 range by binned-counting with np.bincount and then simply count the number of non-zeros in it. Since, np.bincount expects a 1D array, we would flatten the input with np.ravel() and then feed it to bincount.

Hence, the implementation would be -

(np.bincount(a.ravel())!=0).sum()

Runtime test

Helper function to create input array with various number of unique numbers -

def create_input(n_unique):
    unq_nums = np.random.choice(np.arange(256), n_unique,replace=0)
    return np.random.choice(unq_nums, (10,3000,3000)).astype(np.uint8)

Other approach(es) :

# @Warren Weckesser's soln
def assign_method(a):
    q = np.zeros(256, dtype=int)
    q[a.ravel()] = 1
    return len(np.nonzero(q)[0])

Verification of proposed method -

In [141]: a = create_input(n_unique=120)

In [142]: len(np.unique(a))
Out[142]: 120

In [143]: (np.bincount(a.ravel())!=0).sum()
Out[143]: 120

Timings -

In [124]: a = create_input(n_unique=128)

In [125]: %timeit len(np.unique(a)) # Original soln
     ...: %timeit assign_method(a)  # @Warren Weckesser's soln
     ...: %timeit (np.bincount(a.ravel())!=0).sum()
     ...: 
1 loop, best of 3: 3.09 s per loop
1 loop, best of 3: 394 ms per loop
1 loop, best of 3: 209 ms per loop

In [126]: a = create_input(n_unique=256)

In [127]: %timeit len(np.unique(a)) # Original soln
     ...: %timeit assign_method(a)  # @Warren Weckesser's soln
     ...: %timeit (np.bincount(a.ravel())!=0).sum()
     ...: 
1 loop, best of 3: 3.46 s per loop
1 loop, best of 3: 378 ms per loop
1 loop, best of 3: 212 ms per loop