Return to Computer Vision Notebooks

Basics [NumExpr]

---

• Author: Diego Inácio
• GitHub: github.com/diegoinacio
• Notebook: basics_NumExpr.ipynb
• Spec: CPU i7-9750H 2.6 GHz GPU GeForce RTX 2060

---

Basic functions and operations using NumExpr, NumPy and Python.

import numpy as np

0. Installation

---

Installation command for anaconda and pip:

$ conda install -c anaconda numexpr

or

$ pip install numexpr

import numexpr as ne

1. Evaluate

---

The string passed to evaluate is compiled into an object representing the expression and types of the arrays used by the function numexpr.

from time import time

def timeIt_numpy(a, b, runs, loops):
    output = np.zeros(runs)
    for r in range(runs):
        to = time()
        for _ in range(loops):
            current = 2*a**3 + 3*b**2
        ti = time()
        output[r] = ti - to
    return output

def timeIt_numexpr(a, b, runs, loops):
    output = np.zeros(runs)
    for r in range(runs):
        to = time()
        for _ in range(loops):
            current = ne.evaluate('2*a**3 + 3*b**2')
        ti = time()
        output[r] = ti - to
    return output

Lets compare the performance given vectors with different input sizes $N$.

%matplotlib inline
import matplotlib.pyplot as plt

%%time
fig, ax = plt.subplots(1, 1, figsize=(16, 8))

N = 16
runs = 10
loops = 1000
time_numpy = np.zeros((runs, N))
time_numexpr = np.zeros((runs, N))

# Time
for i in range(N):
    a = np.random.rand(2**(i + 1))
    b = np.random.rand(2**(i + 1))
    time_numpy[:,i] = timeIt_numpy(a, b, runs, loops)
    time_numexpr[:,i] = timeIt_numexpr(a, b, runs, loops)

# Plot
x = np.arange(N)
mu = np.mean(time_numpy, axis=0)
# std = np.std(time_numpy, axis=0)
ax.plot(mu, label="NumPy")
# ax.fill_between(x, mu - std, mu + std, alpha=0.2)

x = np.arange(N)
mu = np.mean(time_numexpr, axis=0)
# std = np.std(time_numexpr, axis=0)
ax.plot(mu, label="NumExpr")
# ax.fill_between(x, mu - std, mu + std, alpha=0.2)

ax.set_xlabel('N')
ax.set_xticks(x)
ax.set_xticklabels([f'$2^{{{e + 1}}}$'for e in range(N)])
ax.set_ylabel('time in seconds for each run')

ax.set_title(f'runs = {runs} | loops = {loops}')
ax.legend()

plt.show()

Wall time: 50.2 s