Question

我正在实现编码器解码器lstm，我必须在编码器的每一步进行自定义计算。所以，我正在使用[empty line]。但是，我在从嵌入时访问元素时遇到问题，该元素在时间步raw_rnn处形成为Batch x Time steps x Embedding dimensionality。

这是我的设置：

time

主要部分：

import tensorflow as tf
import numpy as np

batch_size, max_time, input_embedding_size = 5, 10, 16
vocab_size, num_units = 50, 64

encoder_inputs = tf.placeholder(shape=(None, None), dtype=tf.int32, name='encoder_inputs')
encoder_inputs_length = tf.placeholder(shape=(None,), dtype=tf.int32, name='encoder_inputs_length')

embeddings = tf.Variable(tf.random_uniform([vocab_size + 2, input_embedding_size], -1.0, 1.0),
                         dtype=tf.float32, name='embeddings')
encoder_inputs_embedded = tf.nn.embedding_lookup(embeddings, encoder_inputs)

cell = tf.contrib.rnn.LSTMCell(num_units)

跑步部分：

with tf.variable_scope('ReaderNetwork'):
def loop_fn_initial():
    init_elements_finished = (0 >= encoder_inputs_length)
    init_input = cell.zero_state(batch_size, tf.float32)
    init_cell_state = None
    init_cell_output = None
    init_loop_state = None
    return (init_elements_finished, init_input,
            init_cell_state, init_cell_output, init_loop_state)


def loop_fn_transition(time, previous_output, previous_state, previous_loop_state):
    def get_next_input():
        # **TODO** read tensor of shape BATCH X EMBEDDING_DIM from encoder_inputs_embedded
        #  which has shape BATCH x TIME_STEPS x EMBEDDING_DIM

    elements_finished = (time >= encoder_inputs_length)
    finished = tf.reduce_all(elements_finished)  # boolean scalar
    input_val = tf.cond(finished,
                        true_fn=lambda: tf.zeros([batch_size, input_embedding_size]), false_fn=get_next_input)
    state = previous_state
    output = previous_output
    loop_state = None
    return elements_finished, input_val, state, output, loop_state


def loop_fn(time, previous_output, previous_state, previous_loop_state):
    if previous_state is None:  # time = 0
        assert previous_output is None and previous_state is None
        return loop_fn_initial()
    return loop_fn_transition(time, previous_output, previous_state, previous_loop_state)

问题：如何在一个时间步骤访问部分嵌入并返回一个形状张量reader_loop = loop_fn encoder_outputs_ta, encoder_final_state, _ = tf.nn.raw_rnn(cell, loop_fn=reader_loop) outputs = encoder_outputs_ta.stack() def next_batch(): return { encoder_inputs: np.random.random((batch_size, max_time)), encoder_inputs_length: [max_time] * batch_size } init = tf.global_variables_initializer() with tf.Session() as s: s.run(init) outs = s.run([outputs], feed_dict=next_batch()) print len(outs), outs[0].shape？请参阅batch x embedding dim中的函数get_next_input。

谢谢。

Answer 1

我能够解决问题。由于嵌入的形状为Batch x Time steps x Embedding dimensionality，因此我会在time维度上切片。得到的张量具有形状(?, embedding dimensionality)。还需要明确设置结果张量的形状以避免错误：

ValueError：rnn / while / Merge_2：0的形状不是循环的不变量

以下是相关部分：

def get_next_input():
    embedded_value = encoder_inputs_embedded[:, time, :]
    embedded_value.set_shape([batch_size, input_embedding_size])
    return embedded_value

任何人都可以确认这是否是解决问题的正确方法？

以下是完整的参考代码：

import tensorflow as tf
import numpy as np

batch_size, max_time, input_embedding_size = 5, 10, 16
vocab_size, num_units = 50, 64

encoder_inputs = tf.placeholder(shape=(None, None), dtype=tf.int32, name='encoder_inputs')
encoder_inputs_length = tf.placeholder(shape=(None,), dtype=tf.int32, name='encoder_inputs_length')

embeddings = tf.Variable(tf.random_uniform([vocab_size + 2, input_embedding_size], -1.0, 1.0),
                         dtype=tf.float32, name='embeddings')
encoder_inputs_embedded = tf.nn.embedding_lookup(embeddings, encoder_inputs)

cell = tf.contrib.rnn.LSTMCell(num_units)
W = tf.Variable(tf.random_uniform([num_units, vocab_size], -1, 1), dtype=tf.float32, name='W_reader')
b = tf.Variable(tf.zeros([vocab_size]), dtype=tf.float32, name='b_reader')
go_time_slice = tf.ones([batch_size], dtype=tf.int32, name='GO') * 1
go_step_embedded = tf.nn.embedding_lookup(embeddings, go_time_slice)


with tf.variable_scope('ReaderNetwork'):
    def loop_fn_initial():
        init_elements_finished = (0 >= encoder_inputs_length)
        init_input = go_step_embedded
        init_cell_state = cell.zero_state(batch_size, tf.float32)
        init_cell_output = None
        init_loop_state = None
        return (init_elements_finished, init_input,
                init_cell_state, init_cell_output, init_loop_state)

    def loop_fn_transition(time, previous_output, previous_state, previous_loop_state):
        def get_next_input():
            embedded_value = encoder_inputs_embedded[:, time, :]
            embedded_value.set_shape([batch_size, input_embedding_size])
            return embedded_value

        elements_finished = (time >= encoder_inputs_length)
        finished = tf.reduce_all(elements_finished)  # boolean scalar
        next_input = tf.cond(finished,
                             true_fn=lambda: tf.zeros([batch_size, input_embedding_size], dtype=tf.float32),
                             false_fn=get_next_input)
        state = previous_state
        output = previous_output
        loop_state = None
        return elements_finished, next_input, state, output, loop_state


    def loop_fn(time, previous_output, previous_state, previous_loop_state):
        if previous_state is None:  # time = 0
            return loop_fn_initial()
        return loop_fn_transition(time, previous_output, previous_state, previous_loop_state)

reader_loop = loop_fn
encoder_outputs_ta, encoder_final_state, _ = tf.nn.raw_rnn(cell, loop_fn=reader_loop)
outputs = encoder_outputs_ta.stack()


def next_batch():
    return {
        encoder_inputs: np.random.randint(0, vocab_size, (batch_size, max_time)),
        encoder_inputs_length: [max_time] * batch_size
    }


init = tf.global_variables_initializer()
with tf.Session() as s:
    s.run(init)
    outs = s.run([outputs], feed_dict=next_batch())
    print len(outs), outs[0].shape

Tensorflow raw_rnn从嵌入矩阵中检索形状BATCH x DIM的张量

1 个答案: