如何稳定展平NumPy数组？为何matrix的flatten方法失效？

Let's break down what's happening here and how to fix it.

First, let's recap your test code to make sure we're aligned:

from numpy import array, eye, matrix

x = array([1, 0])
# Using numpy's matrix class
B = matrix([[1, 0], [0, 1]])
print(B.dot(x))  # Output: [[1 0]] (a 1×2 matrix object)
print(B.dot(x).flatten())  # Output still: [[1 0]] — no shift to 1D

Why flatten() isn't working #

The root issue is that numpy's matrix class is strictly a 2-dimensional structure. Every method on a matrix (including flatten()) returns another matrix instance, not a 1D ndarray. When you call flatten() on a 1×2 matrix, it just returns the same 1×2 matrix (since it's already "flat" within the 2D context) — it never drops down to a 1D array.

How to convert it to a 1D array #

Here are a few straightforward solutions:

1. Convert the matrix to an ndarray first

You can either use np.array() explicitly, or use the .A shortcut attribute that all matrix objects have (it returns the equivalent ndarray):

result_matrix = B.dot(x)
# Option 1a: Explicit conversion with np.array()
result_1d = np.array(result_matrix).flatten()
print(result_1d)  # Output: [1 0]

# Option 1b: Use the .A attribute
result_1d = result_matrix.A.flatten()
print(result_1d)  # Output: [1 0]

2. Use numpy's top-level ravel() function

Instead of calling the matrix's own flatten() method, use np.ravel() directly. This function works on both matrices and arrays, and will return a 1D ndarray:

print(np.ravel(B.dot(x)))  # Output: [1 0]

3. Stop using matrix entirely (recommended)

Numpy's matrix class is legacy — the official docs now recommend using regular ndarray for all array/matrix operations. ndarray doesn't have the strict 2D constraint, so your original workflow will work as expected:

from numpy import array, eye

x = array([1, 0])
B = array([[1, 0], [0, 1]])  # Use array instead of matrix
print(B.dot(x))  # Output: [1 0] (already a 1D array)
print(B.dot(x).flatten())  # Still outputs [1 0] if you need explicit flattening

内容的提问来源于stack exchange，提问作者Sebastian Oberhoff