DL之DNN优化技术：自定义MultiLayerNetExtend算法(BN层使用/不使用+权重初始值不同)对Mnist数据集训练评估学习过程

输出结果

设计思路

核心代码

(x_train, t_train), (x_test, t_test) = load_mnist(normalize=True)

x_train = x_train[:1000]

t_train = t_train[:1000]

max_epochs = 20

train_size = x_train.shape[0]

batch_size = 100

learning_rate = 0.01

   bn_network = MultiLayerNetExtend(input_size=784, hidden_size_list=[100, 100, 100, 100, 100], output_size=10,

                                   weight_init_std=weight_init_std, use_batchnorm=True)

   network = MultiLayerNetExtend(input_size=784, hidden_size_list=[100, 100, 100, 100, 100], output_size=10,

                               weight_init_std=weight_init_std)

   optimizer = SGD(lr=learning_rate)  

   train_acc_list = []                              

   bn_train_acc_list = []

   iter_per_epoch = max(train_size / batch_size, 1)

   for i in range(1000000000):

       #定义x_batch、t_batch

       batch_mask = np.random.choice(train_size, batch_size)

       x_batch = x_train[batch_mask]

       t_batch = t_train[batch_mask]

       for _network in (bn_network, network):

           grads = _network.gradient(x_batch, t_batch)

           optimizer.update(_network.params, grads)    

       if i % iter_per_epoch == 0:

           train_acc = network.accuracy(x_train, t_train)      

           bn_train_acc = bn_network.accuracy(x_train, t_train)

           train_acc_list.append(train_acc)

           bn_train_acc_list.append(bn_train_acc)

           print("epoch:" + str(epoch_cnt) + " | " + str(train_acc) + " - " + str(bn_train_acc))

           epoch_cnt += 1

           if epoch_cnt >= max_epochs:

               break

   return train_acc_list, bn_train_acc_list