解决Python图像预处理中数组形状不均的ValueError问题

错误现象 #

执行prep_and_load_data()函数时触发如下错误：

ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 2 dimensions. The detected shape was (100, 2) + inhomogeneous part.

相关代码 #

data_prep.py #

import math
import os 
from random import shuffle
import constants as CONST 
import cv2
import pickle
import glob
import numpy as np

def get_size_statistics():
    heights = []
    widths = []
    img_count = 0
    DIR = CONST.TRAIN_DIR2
    for img in os.listdir(CONST.TRAIN_DIR2):
        path = os.path.join(DIR, img)
        data = cv2.imread(path)
        heights.append(data.shape[0])
        widths.append(data.shape[1])
        img_count += 1
    avg_height = sum(heights) / len(heights)
    avg_width = sum(widths) / len(widths)
    print("Average Height: " + str(avg_height))
    print("Max Height: " + str(max(heights)))
    print("Min Height: " + str(min(heights)))
    print('\n')
    print("Average Width: " + str(avg_width))
    print("Max Width: " + str(max(widths)))
    print("Min Width: " + str(min(widths)))


def label_img(name):
    word_label = name.split('.')[0]
    label = CONST.LABEL_MAP[word_label]
    label_arr = np.zeros(2)
    label_arr[label] = 1
    return label_arr

def prep_and_load_data():
    DIR2 = CONST.TRAIN_DIR2
    data = []
    image_paths = os.listdir(DIR2)
    shuffle(image_paths)
    count = 0
    max_images = 100 
    for img_path in image_paths:
        if not img_path.endswith(('.jpg', '.jpeg', '.png')):
            print(f"Skipping non-image file or directory: {img_path}")
            continue
        
        label = label_img(img_path)
        path = os.path.join(DIR2, img_path)
        image = cv2.imread(path)
        image = cv2.resize(image, (CONST.IMG_SIZE, CONST.IMG_SIZE))
        image = image.astype('float') / 255.0 
        data.append([image, label])
        count += 1
        if count == max_images:
            print(f"Reached the limit of {max_images} images.")
            break
    shuffle(data)
    print(len(data))
    print('done')
    print(np.array(data))
    return data

if __name__ == "__main__":
    prep_and_load_data()

constants.py #

TRAIN_DIR2 = 'Code\\dogs-vs-cats\\train\\train'
TEST_DIR2 = 'Code\\dogs-vs-cats\\test1\\test1'
CAT_LBL = 0
DOG_LBL = 1
CAT = 'cat'
DOG = 'dog'
LABEL_MAP = {
    CAT: CAT_LBL,
    DOG: DOG_LBL
}
DATA_SIZE = 18_000
IMG_SIZE = 110
SPLIT_RATIO = 0.8

错误原因 #

直接将包含[图像数组, 标签数组]的列表转换为numpy数组时，图像是三维数组（(110,110,3)），标签是一维数组（(2,)），numpy无法将这种嵌套的异构结构转换为规则的多维数组，从而触发形状不兼容的错误。

修复方法 #

1. 拆分特征与标签为独立数组 #

修改prep_and_load_data()函数的末尾逻辑，将图像和标签分别提取并转换为规则的numpy数组，替代直接转换整个列表的操作：

def prep_and_load_data():
    DIR2 = CONST.TRAIN_DIR2
    data = []
    image_paths = os.listdir(DIR2)
    shuffle(image_paths)
    count = 0
    max_images = 100 
    for img_path in image_paths:
        if not img_path.endswith(('.jpg', '.jpeg', '.png')):
            print(f"Skipping non-image file or directory: {img_path}")
            continue
        
        label = label_img(img_path)
        path = os.path.join(DIR2, img_path)
        image = cv2.imread(path)
        # 新增：检查图像是否读取成功
        if image is None:
            print(f"无法读取图像文件: {path}")
            continue
        image = cv2.resize(image, (CONST.IMG_SIZE, CONST.IMG_SIZE))
        image = image.astype('float') / 255.0 
        data.append([image, label])
        count += 1
        if count == max_images:
            print(f"Reached the limit of {max_images} images.")
            break
    shuffle(data)
    print(f"有效图像数量: {len(data)}")
    # 拆分图像和标签为独立数组
    X = np.array([item[0] for item in data])
    y = np.array([item[1] for item in data])
    print(f"图像数据形状: {X.shape}")
    print(f"标签数据形状: {y.shape}")
    return X, y

2. 调用函数时的适配 #

如果后续代码依赖原函数返回的列表结构，需要调整调用逻辑，接收拆分后的两个数组：

if __name__ == "__main__":
    images, labels = prep_and_load_data()

额外说明 #

新增的图像读取检查可以避免因文件损坏、路径错误导致的cv2.imread返回None，进而引发后续resize操作的崩溃问题，提升代码健壮性。

内容的提问来源于stack exchange，提问作者Luis Cruz