CODEWITHSUNDEEP
python-3.xnumpyrandomcython
Alex
Alex

Reputation: 4264

Improve speed of numpy.random functions

I am writing a simulation of cancer cell population growth, and I am using numpy.random functions to model the cells acquiring or losing mutations. I have determined through profiling that the bottleneck in the code (around 70% of runtime) is the first few lines that contain the numpy.random functions. Here the variable num_steps is a large number, around one million:

def simulate(mu, gamma, beta, num_steps, threshold):
    mutation_num = 0 # the index of the mutation (we assume each mutation only occurs once)
    population = {() : 1} # represents population: tuple of mutations and number of cells with those mutations
    for epoch in range(num_steps):
        next_population = {}
        for mutations, size in population.items():      
            born = np.random.binomial(size, birth_rate)
            if np.random.binomial(born, gamma):
                return True 
            mut_loss = 0 # initializing in case variable is not created
            if mutations:
                mut_gain, mut_loss, mut_same = np.random.multinomial(born, [mu, beta, 1-mu-beta])
            else:
                mut_gain, mut_same = np.random.multinomial(born, [mu, 1-mu])
.....

Is there a way to make the np.random.binomial and np.random.multinomial functions run faster? I tried using Cython, but that did not help.

Upvotes: 0

Views: 1093

Answers (1)

hpaulj
hpaulj

Reputation: 231325

To illustrate my comment:

In [81]: timeit np.random.binomial(1,1,1000)
46.4 µs ± 1.53 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
In [82]: %%timeit
    ...: for _ in range(1000):
    ...:     np.random.binomial(1,1)
    ...: 
4.77 ms ± 186 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

If possible generate many random values with one call rather than one at a time.

Upvotes: 2

Related Questions