Python计算梯度报错ValueError：序列赋值数组元素问题排查

Let's break down your problem step by step—this is a common gotcha when working with numerical gradient functions like the one from CS231n!

The Root Cause of the Error #

Your issue has nothing to do with the shape of your input array (row or column vector)—it's all about the custom function f you're passing in.

The eval_numerical_gradient function is designed to compute gradients for scalar-valued functions (functions that return a single number, not an array). But your f(a) just returns the input array directly. Here's what happens under the hood:

1. When you compute fx = f(x) and fxh = f(x+h), both are arrays (matching the shape of your input).
2. The difference fxh - fx is also an array, and dividing by h gives another array.
3. You then try to assign this entire array to grad[ix]—but grad[ix] is a single scalar position in the gradient array. Numpy throws the "setting an array element with a sequence" error because you can't stuff an array into a single scalar slot.

Why Both Row and Column Arrays Fail #

Whether you pass np.array([1,2,3]) (1D row) or np.array([[1],[2],[3]]) (2D column), the core problem stays the same: your function returns an array instead of a scalar. For example:

• With the row array, when ix = (0,), fxh - fx becomes array([h, 0, 0]), and dividing by h gives array([1, 0, 0]). Trying to assign this to grad[0] (a scalar) is invalid.
• With the column array, fxh - fx becomes array([[h], [0], [0]]), which also can't fit into grad[0,0] (a scalar).

Fixing the Issue: Use a Scalar-Valued Function #

To make the code work, you need to modify f to return a single number. Here are two valid examples:

# Example 1: Sum all elements of the array
def f(a):
    return np.sum(a)

# Example 2: A custom linear combination (scalar output)
def f(a):
    return 2*a[0] + 3*a[1] + a[2]

With these functions, fx and fxh are scalars, so (fxh - fx)/h is also a scalar—perfect for assigning to grad[ix].

Are Your Input Arrays Valid? #

Yes! Both row and column-shaped numpy arrays are valid inputs for this gradient function, as long as your f is scalar-valued. For example, using f(a) = np.sum(a):

• Input np.array([[1],[2],[3]]) will return a gradient array of array([[1.], [1.], [1.]])
• Input np.array([1,2,3]) will return array([1., 1., 1.])

Both results are correct, matching the gradient of the sum function (which is 1 for every element).

内容的提问来源于stack exchange，提问作者LoveMeow