MLP 的动量反向传播

Question

我正在尝试用动量实现反向传播，但在训练循环中，我并没有得到更小的损失，而是越来越高

我已经为反向传播编写了这样的函数。我已经多次检查代码，老实说，我没有注意到任何错误。

`

def backpropagation_mini_batch_momentum(self, inputs, targets, learning_rate, batch_size, momentum = 0.9):    

    for i in range(0, len(inputs), batch_size):
        inputs_batch = inputs[i:i+batch_size]
        targets_batch = targets[i:i+batch_size]

        # Forward propagation
        self.forward_propagation(inputs_batch)

        # Backward propagation
        error = self.activations[-1] - targets_batch
        delta = error
        self.derivatives_weights[-1] = np.dot(self.activations[-2].T, delta)
        self.derivatives_biases[-1] = np.sum(delta, axis=0)
        error = np.dot(delta, self.weights[-1].T)

        for j in range(len(self.weights) - 2, -1, -1):
            delta = error * self._sigmoid_derivative(self.activations[j + 1])
            self.derivatives_weights[j] = np.dot(self.activations[j].T, delta)
            self.derivatives_biases[j] = np.sum(delta, axis=0)
            error = np.dot(delta, self.weights[j].T)

        # Update velocities for weights and biases
        for j in range(len(self.weights)):
            self.velocities_weights[j] = momentum * self.velocities_weights[j] + (learning_rate * self.derivatives_weights[j]) / batch_size
            self.velocities_biases[j] = momentum * self.velocities_biases[j] + (learning_rate * self.derivatives_biases[j]) / batch_size

            # Update weights and biases
            self.weights[j] -= self.velocities_weights[j]
            self.biases[j] -= self.velocities_biases[j]

`