Great question! Let's unpack this clearly—while both functions can return coordinates of non-zero elements in a boolean array, they're not identical in functionality or output format, and each has its own distinct use cases.

First, let's break down the output difference using your example:

• When you run res = np.where(((aa == 7) | (aa == 8))), you get a tuple of arrays where each array corresponds to indices for a single dimension. In your case, (array([0, 0]), array([2, 3])) means the matching elements sit at row indices [0,0] and column indices [2,3].
• When you run res2 = np.argwhere(((aa == 7) | (aa == 8))), you get a 2D array where each row represents a full coordinate pair of a matching element: array([[0, 2], [0, 3]]). In fact, this is exactly equivalent to transposing the tuple from np.where: np.transpose(np.where(condition)) will give you the same result as np.argwhere(condition).

The key distinction comes in the core functionality of each function:

• numpy.where has two distinct modes:
  1. Single argument (boolean array): This is the mode you used, where it returns indices of non-zero elements.
  2. Three arguments (condition, x, y): This is a powerful vectorized tool to choose elements from x or y based on the condition. For example, np.where(aa > 7, aa, 0) would replace all elements in aa that are ≤7 with 0, keeping the rest as-is. This is a completely separate use case that np.argwhere doesn't support.
• numpy.argwhere is specialized solely for returning coordinate pairs (or tuples for higher dimensions) of non-zero elements in a boolean array. Its output format is often more convenient when you need to iterate over individual coordinates (e.g., looping through each matching element's position).

When to use which? #

• Use np.where if:
  • You need to split indices by dimension (e.g., collect all row indices of matches separately from column indices)
  • You want to perform element-wise selection between two arrays based on a condition
• Use np.argwhere if:
  • You want an easy-to-iterate list of coordinate pairs for matching elements
  • You prefer a 2D array format where each row represents a complete position

So to answer your original question: When only given a boolean array, both functions locate the same elements, but their output formats differ. And since np.where has an additional critical use case (the three-argument selection), they are not functionally identical overall.

内容的提问来源于stack exchange，提问作者f. c.