如何使用纯Numpy将ndarray指定列元素转换为特定类型？以及如何通过np.array()将列表转为ndarray时指定元素类型？

Great question! Since you're restricted to pure NumPy and need to convert each column of a string array to a specific type, structured arrays are exactly what you need—they’re designed for scenarios where different columns (or "fields") have distinct data types, and they play perfectly with NumPy’s core functionality.

1. Why Structured Arrays Fit Your Use Case #

Unlike regular NumPy arrays (which require a single uniform dtype), structured arrays act as 1D arrays where each element is a "record" containing multiple fields (your columns) with their own types. This is ideal for your CSV-derived string array, where you need to cast columns to int, float, and string types separately.

2. Step-by-Step Implementation for Your Example #

Let’s walk through your sample input: [['1', '1.6', 'hey'], ['2', '5', 'tr5']]

First, create your initial string array:

import numpy as np

# Your input list converted to a string ndarray
str_arr = np.array([['1', '1.6', 'hey'], ['2', '5', 'tr5']], dtype=np.str_)

Define your target dtype

Specify each column’s name and desired type using a list of tuples. For your example:

• Column 1 → int
• Column 2 → float
• Column 3 → String (we’ll use 'U10' for a Unicode string of max length 10; adjust as needed)

target_dtype = [('col1', np.int_), ('col2', np.float_), ('col3', 'U10')]

Convert to a structured array

Use np.core.records.fromarrays—this function takes each column as a separate array and packs them into a structured array. Note we transpose str_arr first because fromarrays expects columns as input:

structured_arr = np.core.records.fromarrays(str_arr.T, dtype=target_dtype)

What you get

Your structured array will look like this (print it to see):

array([(1, 1.6, 'hey'), (2, 5. , 'tr5')],
      dtype=[('col1', '<i8'), ('col2', '<f8'), ('col3', '<U10')])

Accessing columns/values

You can access entire columns by their field name:

# Get all integers from column 1
structured_arr['col1']  # Output: array([1, 2])

# Get all floats from column 2
structured_arr['col2']  # Output: array([1.6, 5. ])

3. Can you specify per-element types with np.array() directly? #

For regular 2D arrays: No—np.array() requires a uniform dtype for all elements. If you try to create a mixed-type 2D array directly, NumPy will fall back to object dtype (where each element is a Python object), which is slow, unsafe, and not recommended for large datasets.

Structured arrays are the NumPy-native way to handle mixed column types instead. If you really need a "2D-like" mixed structure (though it’s less efficient), you could convert the structured array to an object dtype array:

obj_arr = np.array(structured_arr.tolist(), dtype=object)
# Output: array([[1, 1.6, 'hey'], [2, 5.0, 'tr5']], dtype=object)

But again, structured arrays are better for performance and type safety.

4. Manual Alternative (If You Prefer) #

If you want to cast columns one by one before combining, you can do this:

# Cast each column individually
col1 = str_arr[:, 0].astype(np.int_)
col2 = str_arr[:, 1].astype(np.float_)
col3 = str_arr[:, 2]

# Zip columns into records and create structured array
structured_arr = np.array(list(zip(col1, col2, col3)), dtype=target_dtype)

This works, but fromarrays is more concise for larger datasets.

内容的提问来源于stack exchange，提问作者Sokolovskii Vladislav