Tensorflow：无法理解 ctc_beam_search_decoder() 输出序列

Question

我正在使用 Tensorflow 的 tf.nn.ctc_beam_search_decoder() 解码 RNN 的输出，进行一些多对多映射（即，每个网络单元的多个 softmax 输出）。

网络输出和波束搜索解码器的简化版本是：

import numpy as np
import tensorflow as tf

batch_size = 4
sequence_max_len = 5
num_classes = 3

y_pred = tf.placeholder(tf.float32, shape=(batch_size, sequence_max_len, num_classes))
y_pred_transposed = tf.transpose(y_pred,
                                 perm=[1, 0, 2])  # TF expects dimensions [max_time, batch_size, num_classes]
logits = tf.log(y_pred_transposed)
sequence_lengths = tf.to_int32(tf.fill([batch_size], sequence_max_len))
decoded, log_probabilities = tf.nn.ctc_beam_search_decoder(logits,
                                                           sequence_length=sequence_lengths,
                                                           beam_width=3,
                                                           merge_repeated=False, top_paths=1)

decoded = decoded[0]
decoded_paths = tf.sparse_tensor_to_dense(decoded)  # Shape: [batch_size, max_sequence_len]

with tf.Session() as session:
    tf.global_variables_initializer().run()

    softmax_outputs = np.array([[[0.1, 0.1, 0.8], [0.8, 0.1, 0.1], [0.8, 0.1, 0.1], [0.8, 0.1, 0.1], [0.8, 0.1, 0.1]],
                                [[0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7]],
                                [[0.1, 0.7, 0.2], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7]],
                                [[0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7]]])

    decoded_paths = session.run(decoded_paths, feed_dict = {y_pred: softmax_outputs})
    print(decoded_paths)

这种情况下的输出是：

[[0]
 [1]
 [1]
 [1]]

我的理解是输出张量的维度应该是 [batch_size, max_sequence_len]，每行包含找到的路径中相关类的索引。

在这种情况下，我希望输出类似于：

[[2, 0, 0, 0, 0],
 [2, 2, 2, 2, 2],
 [1, 2, 2, 2, 2],
 [2, 2, 2, 2, 2]]

我对 ctc_beam_search_decoder 的工作原理有什么不了解的地方？

Answer 1

如tf.nn.ctc_beam_search_decoder documentation所示，输出的形状不是[batch_size, max_sequence_len]。相反，它是

[batch_size, max_decoded_length[j]]

（在你的情况下 j=0）。

基于 this paper (which is cited in the github repository 的第 2 节开头），max_decoded_length[0] 从上方以 max_sequence_len 为界，但它们不一定相等。相关引用为：

Let S be a set of training examples drawn from a fixed distribution D_{XxZ}. The input space X = (R^m) is the set of all sequences of m dimensional real valued vectors. The target space Z = L* is the set of all sequences over the (finite) alphabet L of labels. In general, we refer to elements of L* as label sequences or labellings. Each example in S consists of a pair of sequences (x, z). The target sequence z = (z1, z2, ..., zU) is at most as long as the input sequence x = (x1, x2, ..., xT ), i.e. U<=T. Since the input and target sequences are not generally the same length, there is no a priori way of aligning them.

其实max_decoded_length[0]取决于具体的矩阵softmax_outputs。特别是，具有完全相同维度的两个这样的矩阵可能导致不同的max_decoded_length[0]。

例如，如果您替换行

softmax_outputs = np.array([[[0.1, 0.1, 0.8], [0.8, 0.1, 0.1], [0.8, 0.1, 0.1], [0.8, 0.1, 0.1], [0.8, 0.1, 0.1]],
                                [[0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7]],
                                [[0.1, 0.7, 0.2], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7]],
                                [[0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7], [0.1, 0.2, 0.7]]])

与行数

np.random.seed(7)
r=np.random.randint(0,100,size=(4,5,3))
softmax_outputs=r/np.sum(r,2).reshape(4,5,1)

你会得到输出

[[1 0 1]
 [1 0 1]
 [1 0 0]
 [1 0 0]]

（在上面的示例中，softmax_outputs 由 logits 组成，它与您提供的矩阵具有完全相同的维度）。

另一方面，将种子更改为 np.random.seed(50) 会得到输出

[[1 0]
 [1 0]
 [1 0]
 [0 1]]

P.S.

关于你问题的最后一部分：

In this case I would expect the output to be similar to:
[[2, 0, 0, 0, 0],
 [2, 2, 2, 2, 2],
 [1, 2, 2, 2, 2],
 [2, 2, 2, 2, 2]]

请注意，根据documentation，num_classes实际上代表num_labels + 1。具体来说：

The inputs Tensor's innermost dimension size, num_classes, represents num_labels + 1 classes, where num_labels is the number of true labels, and the largest value (num_classes - 1) is reserved for the blank label.

For example, for a vocabulary containing 3 labels [a, b, c], num_classes = 4 and the labels indexing is {a: 0, b: 1, c: 2, blank: 3}.

所以你的真实标签是 0 和 1，2 是为空白标签保留的。空白标签表示观察无标签的情况（第3.1节here）：

A CTC network has a softmax output layer (Bridle, 1990) with one more unit than there are labels in L. The activations of the first |L| units are interpreted as the probabilities of observing the corresponding labels at particular times. The activation of the extra unit is the probability of observing a ‘blank’, or no label. Together, these outputs define the probabilities of all possible ways of aligning all possible label sequences with the input sequence.

Tensorflow：无法理解 ctc_beam_search_decoder() 输出序列

Tensorflow: Can't understand ctc_beam_search_decoder() output sequence

python

tensorflow

beam-search