tensorflow RNN LSTM语言模型

参考博客：

http://blog.csdn.net/u014595019/article/details/52605693   讲解LSTM的原理

https://www.cnblogs.com/wuzhitj/p/6297992.html    LSTM代码讲解

http://blog.csdn.net/qiqiaiairen/article/details/53239506  LSTM函数详解

首先从如下地方下载PTB数据： http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz

以上链接讲解详细，主要说明以下：

（1）数据的总量为 batch_size * num_steps * epoch_size

batch_size:一批数据的样本数

num_steps:LSTM单元的展开步数，即横向LSTM序列上有几个单元

epoch_size:训练的次数，训练一次需要batch_size * num_steps个样本，因为每个单元需要输入batch_size个样本

（2）带权重的交叉熵，第3个参数表示batch_size * num_steps个样本的权重都为1，表示在计算交叉熵时各个样本占比都相同

loss = tf.contrib.legacy_seq2seq.sequence_loss_by_example(
            [logits],
            [tf.reshape(input_.targets, [-1])],
            [tf.ones([batch_size * num_steps], dtype=tf.float32)]

        )

（3）此段代码的功能为将上次的final_state更新到本次训练的初始state，循环次数为num_layers,即LSTM单元叠加的层数

feed_dict = {}
for i, (c, h) in enumerate(model.initial_state):
     feed_dict[c] = state[i].c
     feed_dict[h] = state[i].h

一次循环的结果：

Tensor("train/Model/MultiRNNCellZeroState/BasicLSTMCellZeroState/zeros:0", shape=(20, 200), dtype=float32)
Tensor("train/Model/MultiRNNCellZeroState/BasicLSTMCellZeroState/zeros_1:0", shape=(20, 200), dtype=float32)

状态state是（20,200）的张量，20是batch_size,200是hidden_size,即向量的维度

（4）这里使用reuse=True，可以共享train_model里的变量（使用get_variable定义的），使用已经训练好的权重softmax_w和softmax_b等

关于共享变量的参考链接：http://blog.csdn.net/winycg/article/details/78650045

with tf.variable_scope("Model", reuse=False, initializer=initializer):
            train_model = PTBModel(is_training=True, config=config, input_=train_input))

with tf.variable_scope("Model", reuse=True, initializer=initializer):
            test_model = PTBModel(is_training=False, config=eval_config, input_=test_data)

举例子：

class A(object):
    def __init__(self):
        self.a = tf.get_variable('a', [1], dtype=tf.float32)
        self.b = tf.assign(self.a, [10])


with tf.variable_scope('AA', reuse=False):
    x = A()
with tf.variable_scope('AA', reuse=True):
    y = A()

sess = tf.Session()
print(sess.run(x.b))  # [10.]
print(sess.run(y.a))  # [10.]

全部代码：

import reader
import time
import numpy as np
import tensorflow as tf


class PTBInput(object):
    def __init__(self, data, config, name=None):
        self.batch_size = batch_size = config.batch_size
        self.num_steps = num_steps = config.num_steps
        self.epoch_size = (len(data) // batch_size) // num_steps
        self.input_data, self.targets = reader.ptb_producer(data, batch_size, num_steps, name=name)


class PTBModel(object):
    def __init__(self, is_training, config, input_):
        self._input = input_
        batch_size = input_.batch_size
        num_steps = input_.num_steps
        size = config.hidden_size
        vocab_size = config.vocab_size

        def lstm_cell():
            return tf.contrib.rnn.BasicLSTMCell(size, forget_bias=0.0, state_is_tuple=True)
        attn_cell = lstm_cell
        if is_training and config.keep_prob < 1:
            def attn_cell():
                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)

        self._initial_state = cell.zero_state(batch_size, tf.float32)

        with tf.device('/cpu:0'):
            embedding = tf.get_variable("embeddings", [vocab_size, size], tf.float32)
            inputs = tf.nn.embedding_lookup(embedding, input_.input_data)

        if is_training and config.keep_prob < 1:
            inputs = tf.nn.dropout(input, config.keep_prob)

        outputs = []
        state = self._initial_state
        with tf.variable_scope("RNN"):
            for time_step in range(num_steps):
                if time_step > 0:
                    tf.get_variable_scope().reuse_variables()
                (cell_out, state) = cell(inputs[:, time_step, :], state)
                outputs.append(cell_out)

        output = tf.reshape(tf.concat(outputs, 1), (-1, size))
        softmax_w = tf.get_variable("softmax_w", [size, vocab_size], tf.float32)
        softmax_b = tf.get_variable("softmax_b", [vocab_size], tf.float32)
        logits = tf.matmul(output, softmax_w) + softmax_b
        loss = tf.contrib.legacy_seq2seq.sequence_loss_by_example(
            [logits],
            [tf.reshape(input_.targets, [-1])],
            [tf.ones([batch_size * num_steps], dtype=tf.float32)]
        )
        self._cost = cost = tf.reduce_sum(loss) / batch_size
        self._final_state = state

        if not is_training:
            return
        self._lr = tf.Variable(0.0, trainable=False)
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars), config.max_grad_norm)
        optimizer = tf.train.GradientDescentOptimizer(self._lr)
        # zip将梯度和变量结合，利用优化器将梯度应用到可训练的参数
        self._train_op = optimizer.apply_gradients(zip(grads, tvars),
                                                   global_step=tf.contrib.framework.get_or_create_global_step())
        self._new_lr = tf.placeholder(tf.float32, shape=[], name="new_learning_rate")
        self._lr_update = tf.assign(self._lr, self._new_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


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


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)

    for step in range(model.input.epoch_size):
        feed_dict = {}
        # 更新状态，将上次的final_state更新到本次训练的初始state
        for i, (c, h) in enumerate(model.initial_state):
            feed_dict[c] = state[i].c
            feed_dict[h] = state[i].h

        fetches = {
            "cost": model.cost,
            "final_state": model.final_state
        }
        if eval_op is not None:
            fetches["eval_op"] = eval_op

        vals = session.run(fetches, feed_dict)
        cost = vals["cost"]
        state = vals["final_state"]
        costs += cost
        iters += model.input.num_steps

        if verbose and step % (model.input.epoch_size // 10) == 10:
            print("%.3f perplexity: %.3f speed: %.0f wps" %
                  (step / model.input.epoch_size, np.exp(costs / iters),
                   iters * model.input.batch_size / (time.time() - start_time)))
    return np.exp(costs / iters)


train_data, valid_data, test_data, _ = reader.ptb_raw_data('simple-examples/data/')
config = SmallConfig()
eval_config = SmallConfig()
eval_config.batch_size = 1
eval_config.num_steps = 1

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="train_input")
        with tf.variable_scope("Model", reuse=False, initializer=initializer):
            train_model = PTBModel(is_training=True, config=config, input_=train_input)
    
    with tf.name_scope('valid'):
        valid_input = PTBInput(config=config, data=valid_data, name="valid_input")
        # 复用类里的参数
        with tf.variable_scope("Model", reuse=True, initializer=initializer):
            valid_model = PTBModel(is_training=False, config=config, input_=valid_input)

    with tf.name_scope('test'):
        test_data = PTBInput(config=eval_config, data=test_data,name="test_input")
        with tf.variable_scope("Model", reuse=True, initializer=initializer):
            test_model = PTBModel(is_training=False, config=eval_config, input_=test_data)

    sv = tf.train.Supervisor()
    with sv.managed_session() as session:
        for i in range(config.max_max_epoch):
            lr_decay = config.lr_decay ** max(config.max_epoch, 0)
            train_model.assign_lr(session, config.learning_rate * lr_decay)

            print("Epoch: %d Learning rate: %.3f" % (i + 1, session.run(train_model.lr)))
            train_perplexity = run_epoch(session, train_model, eval_op=train_model.train_op, verbose=True)
            print("Epoch: %d Train Perplexity: %.3f" % (i + 1, train_perplexity))
            valid_perplexity = run_epoch(session, valid_model)
            print("Epoch: %d Valid Perplexity: %.3f" % (i + 1, valid_perplexity))
            test_perplexity = run_epoch(session, test_model)
            print("Epoch: %d Test Perplexity: %.3f" % (i + 1, test_perplexity))