Getting the Best Performance Out of NumPy

TL;DR

- Avoid implicit copy operations

  # implicity-copy
  b = a * 2
  # in-place
  a *= 2

- Reshaping involves a copy when also transposing

  # triggers a copy due to transpose
  c = a.T.reshape((1, -1))
  # does not trigger a copy
  b = a.reshape((1, -1))

- flatten always returns a copy. Use ravel, when possible

  # flatten
  d = a.flatten()
  e = a.ravel()

- Use broadcasting instead of np.tile

- Fancy indexing yields a copy

  b1 = a[::10]                 # array view: does not yield a copy, takes 804 ns per loop
  b2 = a[np.arange(0, n, 10)]  # fancy indexing: creates a copy, takes 14.1 ms per loop

- Logical indexing can be done using np.compress, which is faster than fancy indexing

  i = np.random.random_sample(n) < .5
  b1 = a[i]                       # fancy indexing: 59.8 ms per loop
  b2 = np.compress(i, a, axis=0)  # takes 24.1 ms per loop

- Use np.take as alternative to fancy indexing, when possible

  i = np.arange(0, n, 10)
  b1 = a[i]                   # 13 ms per loop
  b2 = np.take(a, i, axis=0)  # 4.87 ms per loop