Great question—this is a super common gotcha when you're just starting out with Python's variable scopes! Let's break this down clearly:

The Core Issue: Variable Scope + Uncaptured Return Values #

In Python, variables defined inside a function (like your weight and i in grad_d()) are local to that function. That means they only exist while the function is running, and you can't access them directly from outside the function's code block.

Your function does return weight and i at the end, but you never saved those returned values to variables in the outer scope. When you try to run print('Weight: ', weight) outside the function, Python has no idea what weight you're referring to—hence the NameError.

Fixed Code + Additional Improvements #

Here's your code with the key fixes, plus a couple of tweaks to avoid other potential errors:

import numpy as np
from scipy import optimize as opt
import time

def grad_d():
    # Use a numpy array instead of a list for smoother math operations
    weight = np.array([0, 0])
    learnrate = 0.01
    tol = 1.e-5
    itmax = 1000
    for i in range(itmax):
        deltaweight = - learnrate * opt.rosen_der(weight)
        weight = weight + deltaweight
        # Fix: deltaweight is an array—use norm to check convergence (not abs + learnrate)
        if np.linalg.norm(deltaweight) < tol:
            break
    return weight, i

# Capture the function's return values in outer-scope variables
final_weight, total_iterations = grad_d()

print('Weight: ', final_weight)
print('Iterations: ', total_iterations)

What Changed? #

• Captured Return Values: We now assign the values returned by grad_d() to final_weight and total_iterations, which are accessible outside the function.
• Numpy Array for weight: Your original weight = [0, 0] is a regular list, which can cause issues when doing math with numpy arrays (like the output of opt.rosen_der()). Switching to np.array([0, 0]) fixes this.
• Fixed Convergence Check: abs(deltaweight) won't work because deltaweight is an array—this would return an array of element-wise absolute values, which can't be compared to a scalar. Using np.linalg.norm(deltaweight) calculates the magnitude of the array, which is the right way to check if your weight update is small enough to stop.

内容的提问来源于stack exchange，提问作者D K