全连接神经网络

Question

我需要（适当地）实现软最大和梯度损失以及一个可选的辍学（最好是反向），我该怎么做呢？这是我的代码。

    def train(self, features, targets, optimizer, decay_rate_1 = None, 
              decay_rate_2 = None, epsilon = None):
        # Feed Forward
        l0 = features
        l1 = self.activation_function(np.dot(l0, self.w0)+ self.bh)
        l2 = softmax(np.dot(l1, self.w1)+ self.bo)

        # Backpropagation
        l2_error = l2 - targets
        l2_delta = l2_error * self.activation_function(l2, deriv=True)
        l1_error = l2_delta.dot(self.w1.T)
        l1_delta = l1_error * self.activation_function(l1, deriv=True)

        if optimizer == 'sgd':
            # Update Weights
            self.w1 -= self.lr * l1.T.dot(l2_delta)
            self.w0 -= self.lr * l0.T.dot(l1_delta)