I was curious how Numpy stores booleans, so I decided to explore it a bit. Right at the top of the Numpy docs it says that the boolean type is stored as a byte. That’s 8 bits instead of 1, but it probably makes computation more efficient.

Here’s the experiment:

In [2]: tf = np.array([True, False])

In [3]: tf
Out[3]: array([ True, False], dtype=bool)

In [4]: a = np.random.choice(tf, int(1e7))

tf is a Numpy array containing True and False. np.random.choice samples 10 million times in this case. The system monitor verified that this line of code resulted in a data structure occupying 10 MB in memory.

Interesting Zeros

Consider this little line of code:

In [10]: b = np.zeros(int(1e7))

In [11]: b.dtype
Out[11]: dtype('float64')

One might expect it to create 10 million floating point numbers, resulting in an additional memory use of 8 bytes * 10 million ~ 80 MB of memory. This is what happens for np.ones. But np.zeros uses almost no memory. This means that something very clever is happening, and it’s using a sparse data structure.

It’s the subtleties that make these things interesting.