In [1]:
cd ../chapter5/models-master/tutorials/rnn/ptb
In [2]:
# -*- coding:utf-8 -*-
import time
import numpy as np
import tensorflow as tf
import reader
In [3]:
class PTBInput(object):
def __init__(self,config,data,name=None):
self.batch_size = batch_size = config.batch_size #从config中读取参数存到本地变量
self.num_steps = num_steps = config.num_steps
self.epoch_size = (len(data) // batch_size - 1) // num_steps
self.input_data, self.targets = reader.ptb_producer(data,batch_size,num_steps,name=name)
In [4]:
class PTBModel(object):
def __init__(self,is_training,config,input_):
self._input = input_
batch_size = input_.batch_size #从input_中读取参数存到本地变量
num_steps = input_.num_steps
size = config.hidden_size #从config中读取参数存到本地变量，隐含节点个数
vocab_size = config.vocab_size
def lstm_cell(): #使用tf.contrib.rnn.BasicLSTMCell函数设置默认的LSTM单元
return tf.contrib.rnn.BasicLSTMCell(size,forget_bias=0.0,state_is_tuple=True) #size是隐含节点个数，forgets_bias是forget gate的bias，
attn_cell = lstm_cell
if is_training and config.keep_prob < 1: #如果是在训练状态且keepprob<1则在前面的lstm_cell之后接一个Dropout层
def attn_cell(): #使用tf.contrib.rnn.DropoutWrapper函数设置一个dropout层
return tf.contrib.rnn.DropoutWrapper(lstm_cell(),output_keep_prob=config.keep_prob)
cell = tf.contrib.rnn.MultiRNNCell([attn_cell() for _ in range(config.num_layers)],state_is_tuple=True)#用tf.contrib.rnn.MultiRNNCell函数堆叠前面构造的lstm_cell
self._initial_state = cell.zero_state(batch_size, tf.float32) #设置LSTM单元的初始化状态为0
with tf.device("/cpu:0"):
embedding = tf.get_variable( #embedding是一个向量， 将one-hot编码格式的单词转化为向量表达形式
"embedding", [vocab_size,size],dtype=tf.float32) #vocab_size是词汇表数，每个单词向量表达所需的维数为size 分别构成embedding的行和列
inputs = tf.nn.embedding_lookup(embedding,input_.input_data) #查询单词对应的向量表达获得inputs
if is_training and config.keep_prob <1: #如果是训练状态，还要在后面加上一层dropout层
inputs = tf.nn.dropout(inputs,config.keep_prob)
outputs = []
state = self._initial_state
with tf.variable_scope("RNN"): #将下面的操作设为RNN
for time_step in range(num_steps): #设置步数，用来限制梯度在反向传播过程步数
if time_step > 0: tf.get_variable_scope().reuse_variables() #第二次循环开始设置复用变量
(cell_output,state) = cell(inputs[:,time_step,:],state) #inputs的三个维度，第1个代表batch中的第几个样本，第2个代表样本中的第几个单词，第三个代表单词的
outputs.append(cell_output)
output = tf.reshape(tf.concat(outputs,1),[-1,size])
softmax_w = tf.get_variable(
"softmax_w",[size,vocab_size],dtype=tf.float32)
softmax_b = tf.get_variable("softmax_b",[vocab_size],dtype=tf.float32)
logits = tf.matmul(output,softmax_w) + softmax_b
loss = tf.contrib.legacy_seq2seq.sequence_loss_by_example( #用这个函数来计算targets和logits的偏差
[logits],
[tf.reshape(input_.targets,[-1])],
[tf.ones([batch_size * num_steps],dtype=tf.float32)])
self._cost = cost = tf.reduce_sum(loss) / batch_size #汇总batch的误差，在计算每个样本的误差
self._final_state = state #保留最终的状态
if not is_training:
return
self._lr = tf.Variable(0.0,trainable=False) #定义学习率。且设置为不可训练
tvars = tf.trainable_variables()
grads,_ = tf.clip_by_global_norm(tf.gradients(cost,tvars), #计算tvars的梯度，设置梯度的最大范数,
config.max_grad_norm) #这个Gradient Cliping方法，控制梯度的最大范数，某种程度上有正则化的效果，防止梯度爆炸问题
optimizer = tf.train.GradientDescentOptimizer(self._lr)
self._train_op = optimizer.apply_gradients(zip(grads,tvars), #定义一个训练操作，将clip过的梯度应用到所有了训练的参数上
global_step=tf.contrib.framework.get_or_create_global_step()) #生成全局统一的训练步数
self._new_lr = tf.placeholder(
tf.float32,shape=[],name="new_learning_rate") #控制学习速率
self._lr_update = tf.assign(self._lr,self._new_lr) #将新的学习速率赋值给当前的学习速率_lr
def assign_lr(self,session,lr_value):
session.run(self._lr_update,feed_dict={self._new_lr:lr_value})
@property
def input(self):
return self._input
@property
def initial_state(self):
return self._initial_state
@property
def cost(self):
return self._cost
@property
def final_state(self):
return self._final_state
@property
def lr(self):
return self._lr
@property
def train_op(self):
return self._train_op
In [5]:
class SmallConfig(object):
init_scale = 0.1
learning_rate = 1.0
max_grad_norm = 5
num_layers = 2
num_steps = 20
hidden_size = 200
max_epoch = 4
max_max_epoch = 13
keep_prob = 1.0
lr_decay = 0.5
batch_size = 20
vocab_size = 10000
In [6]:
class MediumConfig(object):
init_scale = 0.05
learning_rate = 1.0
max_grad_norm = 5
num_layers = 2
num_steps = 35
hidden_size = 650
max_epoch = 6
max_max_epoch = 39
keep_prob = 0.5
lr_decay = 0.8
batch_size = 20
vocab_size = 10000
In [7]:
class LargeConfig(object):
init_scale = 0.04
learning_rate = 1.0
max_grad_norm = 10
num_layers = 2
num_steps = 35
hidden_size = 1500
max_epoch = 14
max_max_epoch = 55
keep_prob = 0.35
lr_decay = 1 / 1.15
batch_size = 20
vocab_size = 10000
In [8]:
class TestConfig(object):
init_scale = 0.04
learning_rate = 1.0
max_grad_norm = 1
num_layers = 1
num_steps = 2
hidden_size = 2
max_epoch = 1
max_max_epoch = 1
keep_prob = 1.0
lr_decay = 0.5
batch_size = 20
vocab_size = 10000
In [9]:
def run_epoch(session,model,eval_op=None,verbose=False):
start_time = time.time()
costs = 0.0
iters = 0
state = session.run(model.initial_state) #初始化获得初始状态
fetches = {
"cost":model.cost,
"final_state":model.final_state,
}
if eval_op is not None:
fetches["eval_op"] = eval_op #创建结果的字典表
for step in range(model.input.epoch_size): #训练epoch_size
feed_dict = {}
for i,(c,h) in enumerate(model.initial_state): #每次把state装入feed_dict
feed_dict[c] = state[i].c
feed_dict[h] = state[i].h
vals = session.run(fetches,feed_dict) #跑起
cost = vals["cost"] #得到cost
state = vals["final_state"] #得到state
costs += costs #累加cost
iters += model.input.num_steps #累加迭代次数，
if verbose and step % (model.input.epoch_size // 10) == 10: #每隔10次做一次展示
print("%.3f perplexity:%.3f speed: %.0f wps" %
(step * 1.0 / model.input.epoch_size,np.exp(costs/iters),
iters * model.input.batch_size / (time.time()-start_time)))
return np.exp(costs / iters)
In [10]:
data_path='/home/wjj/TFbook/chapter7/simple-examples/data/'
raw_data = reader.ptb_raw_data(data_path) #直接读取解压后的数据
train_data,valid_data,test_data,_ = raw_data #将解压后的数据分别存为训练数据和验证数据以及测试数据
config = SmallConfig() #使用SmallConfig的配置
eval_config = SmallConfig() #测试配置eval_config需和训练配置一致
eval_config.batch_size = 1
eval_config.num_steps = 1
In [12]:
Out[12]:
In [ ]:
with tf.Graph().as_default():
initializer = tf.random_uniform_initializer(-config.init_scale,config.init_scale) #
with tf.name_scope("Train"):
train_input = PTBInput(config=config,data=train_data,name="TrainInput")
with tf.variable_scope("Model",reuse = None, initializer=initializer):
m = PTBModel(is_training=True,config=config,input_=train_input)
with tf.name_scope("Valid"):
valid_input = PTBInput(config=config,data=valid_data,name="ValidInput")
with tf.variable_scope("Model",reuse = True, initializer=initializer):
mvalid = PTBModel(is_training=False,config=config,input_=valid_input)
with tf.name_scope("Test"):
test_input = PTBInput(config=config,data=test_data,name="TestInput")
with tf.variable_scope("Model",reuse = True, initializer=initializer):
mtest = PTBModel(is_training=False,config=config,input_=test_input)
sv = tf.train.Supervisor()
with sv.managed_session() as session:
for i in range(config.max_max_epoch):
lr_decay = config.lr_decay ** max(i+1-config.max_epoch,0.0)
m.assign_lr(session,config.learning_rate * lr_decay)
print("Epoch: %d Learning rate: %.3f" % (i+1,session.run(m.lr)))
train_perplexity = run_epoch(session,m,eval_op=m.train_op,verbose=True)
print("Epoch: %d Train Perplexity: %.3f" % (i+1,train_perplexity))
valid_perplexity = run_epoch(session,mvalid)
print("Epoch: %d Valid Perplexity: %.3f" % (i+1,valid_perplexity))
test_perplexity = run_epoch(session,mtest)
print("Test Perplexity: %.3f" % test_perplexity)
In [ ]:
In [ ]:
In [ ]:
In [ ]:
In [ ]: