### Python 中 BPNN 实现并自定义隐藏层层数 在 Python 中构建具有可变数量隐藏层的反向传播神经网络 (BPNN)，可以利用 `numpy` 库来处理矩阵运算。下面是一个简单的实现方法，该方法允许指定任意数量的隐藏层及其对应的节点数。 首先，创建一个名为 `NeuralNetwork` 的类，在初始化函数中接受输入层大小、输出层大小以及一系列表示各隐藏层大小的整数值列表作为参数[^1]： ```python import numpy as np class NeuralNetwork: def __init__(self, input_size, output_size, hidden_layers): self.input_size = input_size self.output_size = output_size layer_sizes = [input_size] + list(hidden_layers) + [output_size] # 初始化权重和偏置项 self.weights = [] self.biases = [] for i in range(len(layer_sizes)-1): weight_matrix = np.random.randn(layer_sizes[i], layer_sizes[i+1]) bias_vector = np.zeros((1, layer_sizes[i+1])) self.weights.append(weight_matrix) self.biases.append(bias_vector) @staticmethod def sigmoid(x): return 1 / (1 + np.exp(-x)) @staticmethod def derivative_sigmoid(x): sigm = NeuralNetwork.sigmoid(x) return sigm * (1 - sigm) def forward(self, X): activations = [X] for w, b in zip(self.weights[:-1], self.biases[:-1]): z = np.dot(activations[-1], w) + b a = NeuralNetwork.sigmoid(z) activations.append(a) last_z = np.dot(activations[-1], self.weights[-1]) + self.biases[-1] final_output = last_z # For regression tasks or softmax activation can be used here. activations.append(final_output) return activations def backward(self, y_true, activations, learning_rate=0.01): deltas = [(activations[-1] - y_true)] dW = [] dB = [] reversed_weights = list(reversed(self.weights)) for l in range(1, len(activations)): delta_next_layer = deltas[-1].dot(reversed_weights[l-1].T) * \ NeuralNetwork.derivative_sigmoid(activations[-l-1]) deltas.append(delta_next_layer) for idx, act in enumerate(list(zip(activations[:-1], activations[1:]))[::-1]): grad_w = act[0].T.dot(deltas[idx]) grad_b = np.sum(deltas[idx], axis=0, keepdims=True) dW.insert(0, grad_w) dB.insert(0, grad_b) for i in range(len(self.weights)): self.weights[i] -= learning_rate * dW[i] self.biases[i] -= learning_rate * dB[i] ``` 此代码片段展示了如何通过传递给构造器的不同配置来自由调整模型架构中的隐藏层数量和每层内的单元数目[^2]。为了训练这个网络，还需要编写额外的方法来进行前馈计算 (`forward`) 和基于误差梯度更新权值的操作(`backward`)。