为何np.mean计算神经网络准确率比Python循环慢且结果不一致？

我在构建一个二分类神经网络时，发现运行速度极慢，排查后定位到用np.mean计算准确率的步骤平均耗时0.55秒，而自己用循环实现的计算仅需0.01秒以内，且两者输出的准确率结果还不一致。

相关代码 #

原准确率计算代码

predictions = self.neural_network.propagate(X_train)
rounded_predictions = (predictions > 0.5).astype(int)
accuracy = np.mean(rounded_predictions == y_train)

其中X_train形状为(16000,16)，y_train形状为(16000,)。

自定义循环实现的准确率计算

def compute_accuracy_score(y_train, predictions):
    """
        Calculates the accuracy score of the predictions made by the network.
        If a prediction matches its true label, it increments the count of correct predictions.
        The accuracy score is the correct predictions divided by the total predictions.
    """
    num_samples = len(y_train)
    correct_predictions = 0

    # If the prediction is correct, increment the count of correct predictions
    for true_label, predicted_label in zip(y_train, predictions):
        if true_label == predicted_label:
            correct_predictions += 1

    # Compute accuracy as the ratio of correct predictions to total number of samples
    accuracy = correct_predictions / num_samples
    return accuracy

测试运行时间的代码

def calculate_fitness(self):
    predictions = self.neural_network.propagate(X_train)
    rounded_predictions = (predictions > 0.5).astype(int)
    start_time = time.time()
    accuracy1 = compute_accuracy_score(y_train, rounded_predictions)
    end_time = time.time()
    print('Time to calculate fitness using compute accuracy:', end_time - start_time, 'seconds')
    start_time = time.time()
    accuracy = np.mean(rounded_predictions == y_train)
    end_time = time.time()
    print('Time to calculate fitness using np.mean:', end_time - start_time, 'seconds')
    print('Accuracy:', accuracy, 'Accuracy1:', accuracy1)
    self.fitness = round(float(accuracy), 4)

测试结果

Time to calculate fitness using compute accuracy: 0.022114276885986328 seconds 
Time to calculate fitness using np.mean: 0.6563560962677002 seconds 
Accuracy: 0.559151625 Accuracy1: 0.6015625 
Time to calculate fitness using compute accuracy: 0.021242141723632812 seconds 
Time to calculate fitness using np.mean: 0.6553714275360107 seconds 
Accuracy: 0.441705 Accuracy1: 0.398875 

差异原因分析 #

1. 运行速度差异：数组形状不匹配引发的超大临时数组开销

最核心的问题是rounded_predictions与y_train形状不一致，比如rounded_predictions是(16000,1)而y_train是(16000,)。此时执行rounded_predictions == y_train时，numpy会触发广播机制，将两个数组强制扩展为(16000,16000)的超大布尔数组（包含2.56亿个元素），创建这个数组以及计算全局均值的内存和计算开销极大，直接导致耗时飙升。

而自定义循环通过zip逐个遍历元素，完全避免了超大临时数组的创建和内存拷贝，因此速度快得多。即使形状看似匹配，若数组内存布局不连续（比如是切片结果）或数据类型不一致，numpy的向量操作也会产生额外转换开销，而循环遍历不受这些因素影响。

2. 结果不一致：广播后的均值计算逻辑完全错误

当rounded_predictions与y_train形状不匹配时，rounded_predictions == y_train生成的是广播后的超大数组，此时np.mean计算的是所有元素的全局均值，而非每个样本预测与真实标签的匹配率，完全偏离了准确率的计算逻辑。

而自定义循环中，zip(y_train, predictions)会按顺序逐个匹配样本的真实标签和预测值（即使预测值是长度为1的数组，Python在if判断中会自动将单元素布尔数组转为布尔值），因此计算的是正确的样本匹配率。

快速解决方案 #

先统一数组形状，确保rounded_predictions与y_train维度一致，再计算准确率：

# 方案1：展平预测数组后计算均值
accuracy = np.mean(rounded_predictions.flatten() == y_train)

# 方案2：用求和替代均值，效率更高
accuracy = np.sum(rounded_predictions.flatten() == y_train) / len(y_train)

内容的提问来源于stack exchange，提问作者Yair