使用 tensorlayer 组建 残差网络 resnet 实现 mnist 手写识别例子

最近学习残差网络，非常给力，即使是深层网络也能很快收敛
这里的代码构建了一个17层的网络，5 epoch就能达到96%以上准确率

lost-损失，acc-准确率

不过发现几个问题
1.使用训练过程中，lost值会先减小，然后会一直增大，而acc值却在一直上升
2.使用prelu神经元，lost值增大更快，acc值训练时间长会降低
3.使用elu神经元与prelu神经元相比，elu神经元lost增大更缓慢，acc值能持续增大，prelu神经元的acc值后期会下降

不使用relu主要是因为担心relu坏死问题

完整代码

import tensorflow as tf
import tensorlayer as tl

sess = tf.InteractiveSession()

# 准备数据
X_train, y_train, X_val, y_val, X_test, y_test = tl.files.load_mnist_dataset(shape=(-1,784))

# 定义 placeholder
x = tf.placeholder(tf.float32, shape=[None, 784], name='x')
y_ = tf.placeholder(tf.int64, shape=[None, ], name='y_')

# 定义模型
network = tl.layers.InputLayer(x, name='input_layer')
res_a = network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu1')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu2')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu3')
res_a = network = tl.layers.ElementwiseLayer([network, res_a], combine_fn=tf.add, name='res_add1')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu4')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu5')
res_a = network = tl.layers.ElementwiseLayer([network, res_a], combine_fn=tf.add, name='res_add2')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu6')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu7')
res_a = network = tl.layers.ElementwiseLayer([network, res_a], combine_fn=tf.add, name='res_add3')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu8')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu9')
res_a = network = tl.layers.ElementwiseLayer([network, res_a], combine_fn=tf.add, name='res_add4')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu10')
network = tl.layers.DenseLayer(network, n_units=200, act = tf.nn.elu, name='relu11')
res_a = network = tl.layers.ElementwiseLayer([network, res_a], combine_fn=tf.add, name='res_add5')
network = tl.layers.DenseLayer(network, n_units=10, act = tf.identity, name='output_layer')
# 定义损失函数和衡量指标
# tl.cost.cross_entropy 在内部使用 tf.nn.sparse_softmax_cross_entropy_with_logits() 实现 softmax
y = network.outputs
cost = tl.cost.cross_entropy(y, y_, name = 'cost')
correct_prediction = tf.equal(tf.argmax(y, 1), y_)
acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
y_op = tf.argmax(tf.nn.softmax(y), 1)

# 定义 optimizer
train_params = network.all_params
train_op = tf.train.AdamOptimizer(learning_rate=0.003, beta1=0.9, beta2=0.999,
                            epsilon=1e-08, use_locking=False).minimize(cost, var_list=train_params)

# 初始化 session 中的所有参数
tl.layers.initialize_global_variables(sess)

# 列出模型信息
network.print_params()
network.print_layers()

# 训练模型
tl.utils.fit(sess, network, train_op, cost, X_train, y_train, x, y_,
            acc=acc, batch_size=500, n_epoch=500, print_freq=5,
            X_val=X_val, y_val=y_val, eval_train=False)

# 评估模型
tl.utils.test(sess, network, acc, X_test, y_test, x, y_, batch_size=None, cost=cost)

# 把模型保存成 .npz 文件
tl.files.save_npz(network.all_params , name='model.npz')
sess.close()