TensorFlow LSTM Shape变换 for ptb model

TensorFlow的图对应的shape可以先用print出来debug，拿ptb model里面的build_rnn_graph_lstm代码为例:

import numpy as np
import tensorflow as tf


class SmallConfig(object):
  """Small config."""
  init_scale = 0.1
  learning_rate = 1.0
  max_grad_norm = 5
  num_layers = 2
  num_steps = 2
  hidden_size = 200
  max_epoch = 4
  max_max_epoch = 13
  keep_prob = 1.0
  lr_decay = 0.5
  batch_size = 3
  vocab_size = 10000

def get_lstm_cell(config, is_training):
    return tf.contrib.rnn.BasicLSTMCell(config.hidden_size, forget_bias=0.0, state_is_tuple=True, reuse=not is_training)

def build_rnn_graph_lstm(inputs, config, is_training):
    """Build the inference graph using canonical LSTM cells."""

    # Slightly better results can be obtained with forget gate biases
    # initialized to 1 but the hyperparameters of the model would need to be
    # different than reported in the paper.
    def make_cell():
        cell = get_lstm_cell(config, is_training)
        if is_training and config.keep_prob < 1:
            cell = tf.contrib.rnn.DropoutWrapper(
                cell, output_keep_prob=config.keep_prob)
        return cell

    cell = tf.contrib.rnn.MultiRNNCell(
        [make_cell() for _ in range(config.num_layers)], state_is_tuple=True)

    initial_state = cell.zero_state(config.batch_size, tf.float32)
    state = initial_state
    # Simplified version of tf.nn.static_rnn().
    # This builds an unrolled LSTM for tutorial purposes only.
    # In general, use tf.nn.static_rnn() or tf.nn.static_state_saving_rnn().
    #
    # The alternative version of the code below is:
    #
    # inputs = tf.unstack(inputs, num=config.num_steps, axis=1)
    # outputs, state = tf.nn.static_rnn(cell, inputs,
    #                                   initial_state=initial_state)
    outputs = []
    with tf.variable_scope("RNN"):
        for time_step in range(config.num_steps):
            if time_step > 0: tf.get_variable_scope().reuse_variables()
            (cell_output, state) = cell(inputs[:, time_step, :], state)
            print("cell_output = ", cell_output)
            print("cell_state = ", state)
            outputs.append(cell_output)
    outputs = tf.reshape(tf.concat(outputs, 1), [-1, config.hidden_size])
    return outputs, state

config = SmallConfig()
#inputs = tf.get_variable("inputs", initializer=[[[1.],[2.]], [[3.],[4.]], [[5.],[6.]]])
inputs = tf.placeholder(dtype = tf.float32, shape = [3, 2, 1])
print(inputs)
output, state = build_rnn_graph_lstm(inputs = inputs, config=config, is_training = True)
print("output =", output)
print("state =", state)

对应的输出为：

Tensor("Placeholder:0", shape=(3, 2, 1), dtype=float32)
cell_output =  Tensor("RNN/RNN/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/mul_2:0", shape=(3, 200), dtype=float32)
cell_state =  (LSTMStateTuple(c=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_0/cell_0/basic_lstm_cell/add_1:0' shape=(3, 200) dtype=float32>, h=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_0/cell_0/basic_lstm_cell/mul_2:0' shape=(3, 200) dtype=float32>), LSTMStateTuple(c=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/add_1:0' shape=(3, 200) dtype=float32>, h=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/mul_2:0' shape=(3, 200) dtype=float32>))
cell_output =  Tensor("RNN/RNN/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/mul_5:0", shape=(3, 200), dtype=float32)
cell_state =  (LSTMStateTuple(c=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_0/cell_0/basic_lstm_cell/add_3:0' shape=(3, 200) dtype=float32>, h=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_0/cell_0/basic_lstm_cell/mul_5:0' shape=(3, 200) dtype=float32>), LSTMStateTuple(c=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/add_3:0' shape=(3, 200) dtype=float32>, h=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/mul_5:0' shape=(3, 200) dtype=float32>))
output = Tensor("Reshape:0", shape=(6, 200), dtype=float32)
state = (LSTMStateTuple(c=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_0/cell_0/basic_lstm_cell/add_3:0' shape=(3, 200) dtype=float32>, h=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_0/cell_0/basic_lstm_cell/mul_5:0' shape=(3, 200) dtype=float32>), LSTMStateTuple(c=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/add_3:0' shape=(3, 200) dtype=float32>, h=<tf.Tensor 'RNN/RNN/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/mul_5:0' shape=(3, 200) dtype=float32>))

但如果用注释中的代码：

import numpy as np
import tensorflow as tf


class SmallConfig(object):
  """Small config."""
  init_scale = 0.1
  learning_rate = 1.0
  max_grad_norm = 5
  num_layers = 2
  num_steps = 2
  hidden_size = 200
  max_epoch = 4
  max_max_epoch = 13
  keep_prob = 1.0
  lr_decay = 0.5
  batch_size = 3
  vocab_size = 10000

def get_lstm_cell(config, is_training):
    return tf.contrib.rnn.BasicLSTMCell(config.hidden_size, forget_bias=0.0, state_is_tuple=True, reuse=not is_training)

def build_rnn_graph_lstm(inputs, config, is_training):
    """Build the inference graph using canonical LSTM cells."""

    # Slightly better results can be obtained with forget gate biases
    # initialized to 1 but the hyperparameters of the model would need to be
    # different than reported in the paper.
    def make_cell():
        cell = get_lstm_cell(config, is_training)
        if is_training and config.keep_prob < 1:
            cell = tf.contrib.rnn.DropoutWrapper(
                cell, output_keep_prob=config.keep_prob)
        return cell

    cell = tf.contrib.rnn.MultiRNNCell(
        [make_cell() for _ in range(config.num_layers)], state_is_tuple=True)

    initial_state = cell.zero_state(config.batch_size, tf.float32)
    state = initial_state
    # Simplified version of tf.nn.static_rnn().
    # This builds an unrolled LSTM for tutorial purposes only.
    # In general, use tf.nn.static_rnn() or tf.nn.static_state_saving_rnn().
    #
    # The alternative version of the code below is:
    #
    inputs = tf.unstack(inputs, num=config.num_steps, axis=1)
    outputs, state = tf.nn.static_rnn(cell, inputs,
                                      initial_state=initial_state)
    # outputs = []
    # with tf.variable_scope("RNN"):
    #     for time_step in range(config.num_steps):
    #         if time_step > 0: tf.get_variable_scope().reuse_variables()
    #         (cell_output, state) = cell(inputs[:, time_step, :], state)
    #         print("cell_output = ", cell_output)
    #         print("cell_state = ", state)
    #         outputs.append(cell_output)
    # outputs = tf.reshape(tf.concat(outputs, 1), [-1, config.hidden_size])
    return outputs, state

config = SmallConfig()
#inputs = tf.get_variable("inputs", initializer=[[[1.],[2.]], [[3.],[4.]], [[5.],[6.]]])
inputs = tf.placeholder(dtype = tf.float32, shape = [3, 2, 1])
print(inputs)
output, state = build_rnn_graph_lstm(inputs = inputs, config=config, is_training = True)
print("output =", output)
print("state =", state)

对应的输出为：

Tensor("Placeholder:0", shape=(3, 2, 1), dtype=float32)
output = [<tf.Tensor 'rnn/rnn/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/mul_2:0' shape=(3, 200) dtype=float32>, <tf.Tensor 'rnn/rnn/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/mul_5:0' shape=(3, 200) dtype=float32>]
state = (LSTMStateTuple(c=<tf.Tensor 'rnn/rnn/multi_rnn_cell/cell_0/cell_0/basic_lstm_cell/add_3:0' shape=(3, 200) dtype=float32>, h=<tf.Tensor 'rnn/rnn/multi_rnn_cell/cell_0/cell_0/basic_lstm_cell/mul_5:0' shape=(3, 200) dtype=float32>), LSTMStateTuple(c=<tf.Tensor 'rnn/rnn/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/add_3:0' shape=(3, 200) dtype=float32>, h=<tf.Tensor 'rnn/rnn/multi_rnn_cell/cell_1/cell_1/basic_lstm_cell/mul_5:0' shape=(3, 200) dtype=float32>))

所以总结一下，PTB LSTM的输入每次都是一批单词，大小为batch_size, 例如以下官网上的例子，其中每一个time_steps都会投一批词进去，这就是inputs的shape为[batch_size, time_steps, embedding_size]，至于第三维参数，取决于Embedding的方法。

 t=0  t=1    t=2  t=3     t=4
[The, brown, fox, is,     quick]
[The, red,   fox, jumped, high]

words_in_dataset[0] = [The, The]
words_in_dataset[1] = [brown, red]
words_in_dataset[2] = [fox, fox]
words_in_dataset[3] = [is, jumped]
words_in_dataset[4] = [quick, high]
batch_size = 2, time_steps = 5

用一张图来展示多层LSTM状态的变化：

这正和http://blog.csdn.net/taoqick/article/details/79475350 一文中最后一图成对应的关系。