首先从阅读论文开始。
先后阅读了如下文章
关于《A Critical Review of Recurrent Neural Networks for Sequence Learning》的阅读理解
《Understanding LSTM Networks》——文章对 LSTM 结构为什么这样设计,做了一步步的推理解释
关于《Supervised Sequence Labelling with Recurrent Neural Networks》的阅读理解
……一些文章
然后是一些tensorflow实现RNN或LSTM的例子。
目前,把普通的神经网络改造成RNN的成果如下。对RNN用tensorflow实现的逻辑可以理顺,但是实现起来有错误,提示维度不匹配。正在检查原因。
import numpyas np
import pandasas pd
import tensorflowas tf
# 转为onehot编码
def turn_onehot(df):
for keyin df.columns:
oneHot = pd.get_dummies(df[key])
for oneHotKeyin oneHot.columns:# 防止重名
oneHot = oneHot.rename(columns={oneHotKey: key +'_' +str(oneHotKey)})
df = df.drop(key,axis=1)
df = df.join(oneHot)
return df
# 获取一批次的数据
def get_batch(x_date, y_date, batch):
global pointer
x_date_batch = x_date[pointer:pointer + batch]
y_date_batch = y_date[pointer:pointer + batch]
pointer = pointer + batch
return x_date_batch, y_date_batch
# 生成layer
def add_layer(input_num, output_num, x, layer, active=None):
# 生成权重
with tf.name_scope('layer' + layer +'/W' + layer):
W = tf.Variable(tf.random_normal([2*input_num, output_num],dtype=tf.float32),name='W' + layer)
tf.summary.histogram('layer' + layer +'/W' + layer, W)
# 加入L2正则化
if isregularization:
tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(lambda1)(W))
# 生成偏移量
with tf.name_scope('layer' + layer +'/b' + layer):
b = tf.Variable(tf.zeros([output_num]) +0.1,dtype=tf.float32,name='b' + layer)
tf.summary.histogram('layer' + layer +'/b' + layer, b)
# 激活
with tf.name_scope('layer' + layer +'/l' + layer):
l = active(tf.matmul(x, W) + b)# 使用sigmoid激活函数,备用函数还有relu
tf.summary.histogram('layer' + layer +'/l' + layer, l)
return l
hiddenDim =1000 # 隐藏层神经元数
lambda1 =0.5 # 正则化超参数
save_file ='./train_model.ckpt'
pointer =0
time_step =1
istrain =True # 启用训练模式
istensorborad =False # 启用tensorboard
isregularization =False # 启用正则化
if istrain:
samples =2000
batch =1 # 每批次的数据输入数量
else:
samples =550
batch =1 # 每批次的数据输入数量
with tf.name_scope('inputdate-x-y'):
# 导入
df = pd.DataFrame(pd.read_csv('GHMX.CSV',header=0))
# 产生 y_data 值 (n, 1)
y_date = df['number'].values
y_date = y_date.reshape((-1,1))
# 产生 x_data 值 (n, 4+12+31+24)
df = df.drop('number',axis=1)
df = turn_onehot(df)
x_data = df.values
###生成神经网络模型
# 占位符
with tf.name_scope('inputs'):
x = tf.placeholder(tf.float32,shape=[None, time_step,71],name='x_input')
y_ = tf.placeholder(tf.float32,shape=[None,1],name='y_input')
keep_prob = tf.placeholder(tf.float32,name='keep_prob')
# 生成神经网络
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=71,forget_bias=1.0,state_is_tuple=True)
lstm_cell = tf.nn.rnn_cell.DropoutWrapper(cell=lstm_cell,input_keep_prob=1.0,output_keep_prob=keep_prob)
mlstm_cell = tf.nn.rnn_cell.MultiRNNCell([lstm_cellfor _in range(3)])
init_state = mlstm_cell.zero_state(batch,dtype=tf.float32)
outputs, date = tf.nn.dynamic_rnn(mlstm_cell,inputs=x,initial_state=init_state,time_major=False)
h_date= outputs[:, -1, :]
y = add_layer(71,1, h_date,'1', tf.nn.relu)
# 计算loss
with tf.name_scope('loss'):
# loss = tf.reduce_mean(tf.reduce_sum(tf.square(y - y_), name='square'), name='loss') #损失函数,损失不下降,换用别的函数
# loss = -tf.reduce_sum(y_*tf.log(y)) #损失仍然不下降
# loss = -tf.reduce_sum(y_*tf.log(tf.clip_by_value(y,1e-10,1.0)) , name='loss')
loss = tf.losses.mean_squared_error(labels=y_,predictions=y)
#tf.add_to_collection('losses', mse_loss) # 损失集合
#loss = tf.add_n(tf.get_collection('losses'))
tf.summary.scalar('loss', loss)
# 梯度下降
with tf.name_scope('train_step'):
train_step = tf.train.GradientDescentOptimizer(0.0005).minimize(loss)# 有效的学习率0.000005
# 初始化
init = tf.global_variables_initializer()
sess = tf.Session()
if istensorborad:
merged = tf.summary.merge_all()
writer = tf.summary.FileWriter('logs/', sess.graph)
sess.run(init)
# 保存/读取模型
saver = tf.train.Saver()
if not istrain:
saver.restore(sess, save_file)
for iin range(samples):
x_date_batch, y_date_batch = get_batch(x_data, y_date, batch)
feed_dict = {x: x_date_batch, y_: y_date_batch, keep_prob:1.0}
if istrain:
_, loss_value, y_value, y__value = sess.run((train_step, loss, y, y_),feed_dict=feed_dict)
print('y=', y_value,'----ture=', y__value)
print(loss_value)
else:
loss_value, y_value, y__value = sess.run((loss, y, y_),feed_dict=feed_dict)
print('y=', y_value,'----ture=', y__value)
print(loss_value)
if istensorborad:
result = sess.run(merged,feed_dict=feed_dict)
writer.add_summary(result, i)
# 保存模型
if istrain:
saver.save(sess, save_file)