Question

我有这样一个占位符：

self._sentence_lengths = tf.placeholder('int32', shape=[None], name='sen_len')

我还有一个embeds张量，其形状为(?, 300)。

我想根据embeds执行sentence_lengths的拆分：sentences = tf.split(embeds, self._sentence_lengths)

但是，我收到以下错误：

ValueError: Cannot infer num from shape Tensor("joint_architecture_1/encoder_1/sen_len:0", shape=(?,), dtype=int32)

最初，我用这种方式创建self._sentence_lengths（一切正常）：

self._sentence_lengths = tf.placeholder('int32', shape=[self.batch_size], name='sen_len')

我想将其更改为动态方法的原因是我不希望受到批量大小的限制。实际上，培训时可能会使用批量大小，例如128.但在进行推理时，我需要小批量。

到目前为止我一直这样做的方式是我在恢复时更改self.batch_size，但这看起来并不优雅。

有没有办法克服这个问题？

Answer 1

我设法使用一个小技巧找到了解决这个问题的方法。我创建了一个常量Tensor，它将保存我的Variable的内容。 Tensor的大小将限制sen_len的大小，但如果我们选择它足够大，那应该不是问题。

以下是我的解决方案＆gt;

的玩具示例

    embeds_raw = tf.constant(np.array([
    [1, 1],
    [1, 1],
    [2, 2],
    [3, 3],
    [3, 3],
    [3, 3],
    [4, 4],
    [4, 4],
    [4, 4],
    [4, 4],
], dtype='float32'))
# These play the role of embeddings.
embeds = tf.Variable(initial_value=embeds_raw)
# This variable plays the role of a container. We chose zeros because they are neutral to addition.
container_variable = tf.zeros([512], dtype=tf.int32, name='container_variable')
# Our placeholder for sentence lengths.
sen_len = tf.placeholder('int32', shape=[None], name='sen_len')
# Getting the length of the longest sentence.
max_l = tf.reduce_max(sen_len)
# Number of sentences.
nbr_sentences = tf.shape(sen_len)[0]
# We pad the sentence length var to match that of the container variable.
padded_sen_len = tf.pad(sen_len, [[0, 512 - nbr_sentences]], 'CONSTANT')
# We add the sentence lengths to our container variable.
added_container_variable = tf.add(container_variable, padded_sen_len)
# Create a TensorArray that will contain the split.
u1 = tf.TensorArray(dtype=tf.float32, size=512, clear_after_read=False)
# Split the embeddings by the sentence lengths.
u1 = u1.split(embeds, added_container_variable)

# Loop variables. An index and a variable containing our concatenated arrays.
i = tf.constant(0, shape=(), dtype='int32', name='i')
x = tf.constant(0, shape=[1, 2], dtype=tf.float32)

def condition(_i, _):
    """Checking whether _i is less than the number of sentences."""
    return tf.less(_i, nbr_sentences)

def body(_i, _x):
    """Padding and concatenating with _x."""
    temp = tf.pad(u1.read(_i), [[0, max_l - sen_len[_i]], [0, 0]], 'CONSTANT')
    return _i + 1, tf.concat([_x, temp], 0)

# Looping.
idx, padded_concatenated_sentences = tf.while_loop(
    condition,
    body,
    [i, x],
    shape_invariants=[tf.TensorShape([]), tf.TensorShape([None, 2])]
)

# Getting rid of the first row since it contains 0s.
padded_concatenated_sentences = padded_concatenated_sentences[1:]

# Reshaping to obtain the desired results. In our case 2 would be the word embedding dimensionality.
reshaped_elements = tf.reshape(padded_concatenated_sentences, [nbr_sentences, max_l, 2])

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    sents = sess.run(reshaped_elements, feed_dict={sen_len: [2, 1, 3, 4]})
    print(sents)

如何在Tensorflow中处理动态输入大小

1 个答案: