Sounds great!

@rileyjmurray Could you explain a bit more when .ravel() can be used instead of .flatten()? I'm confused about when a copy is necessary.

@pcwysoc, as a rule of thumb, copies aren't necessary. Python (or rather, numpy) makes them all the time. The main time when a copy would be important is if you're explicitly using in-place operations or if you're overwriting portions of an array.

import numpy as np

a = np.ones((3,5))
b = a.ravel()
b = b + 2
print(a) # unchanged, because the change to b1 wasn't specifically in-place.

b = a.ravel()
b += 2
print(a) # changed, since "+=" is an in-place operation

a = np.ones((3,5))
b = a.ravel()
b[::2] = 0.0  # overwrite every other element of b with 0.0.
print(a) # changed, since b shares memory with a

I've started working on this.

Here are common patterns I've seen where .flatten() can be safely replaced with .ravel():

• np.concatenate([ ... vec.flatten() ... ]), since the concatenation will allocate new memory anyway.
• f(vec.flatten()) where f the return value of f is guaranteed not to point to its argument's memory. For example, f could be np.all, or np.sum, or np.isnan.
• y = f(x).flatten() where f(x) returns a new array

I'll edit this comment as I go.

Fixed in #451