为何np.vectorize会将一维数组中的np.uint8元素转换为int类型，却保留列表元素的原类型？

Great question! The difference in behavior comes down to how np.vectorize handles numpy array inputs vs. Python sequence (like list) inputs during its two key execution phases: type inference and actual function application. Let's break this down step by step.

1. How np.vectorize works at a high level #

np.vectorize operates in two core stages:

• Type inference: It runs your function on the first element of the input to determine the output dtype and shape (this is why you see the first set of print statements in all your examples).
• Actual execution: It iterates over all input elements, applies your function, and populates the output array.

The discrepancy you're observing happens entirely in the second stage.

2. Why numpy array inputs get converted to Python ints #

When you pass a numpy array (like np.array([np.uint8(1)])) to np.vectorize:

• During the execution phase, numpy internally handles array element access by converting numpy scalar types (like np.uint8) to their corresponding Python native types—in this case, a Python int. This is standard behavior when numpy exposes array elements to pure Python functions; most numeric dtypes get unwrapped into Python's built-in scalar types.
• So during the second (execution) pass, the element passed to f() is a Python int instead of the original np.uint8 scalar, hence the "no dtype exists on x" message.

3. Why list inputs preserve the numpy uint8 type #

When you pass a Python list (like [np.uint8(1)]) to np.vectorize:

• np.vectorize treats it as a generic Python sequence, not a numpy array. It simply iterates over the list elements directly, passing each original np.uint8 object to your function without any conversion.
• That's why both the type inference pass and execution pass see the np.uint8 scalar, so you get the x.dtype is uint8 message twice.

4. What about passing a single np.uint8 scalar? #

When you pass np.uint8(1) directly to the vectorized function:

• np.vectorize still runs the type inference pass (using the scalar itself) and then executes the function again on the same scalar. Since there's no numpy array involved to trigger type unwrapping, the np.uint8 type is preserved both times.

A quick note on np.vectorize #

Keep in mind that np.vectorize is not a true vectorization tool—it's a convenience wrapper that loops over elements under the hood. For performance-critical code, you'd want to use actual numpy vectorized operations instead. But for understanding type behavior, this quirk is a great example of how numpy interacts with pure Python functions.

内容的提问来源于stack exchange，提问作者zabop