如何将形状动态输入到占位符

Question

我有一个函数可以检索batch of sentences，从句子中的单词创建bag of words，从单词包中的单词创建character id matrix。

char_id_batch, word_id_batch, pos_id_batch = temp.retrieve_batch_sent(start, batch_size_counter)

我有三个占位符：

1. char_id：每个[length of bag of words, maximum length of word in bag of words]的char_id形状为batch of sentences。
2. word_id：每个[batch of sentences, maximum length of sentence in batch of sentences]的word_id形状为batch of sentences。
3. 输出：y的形状与word_id相同。

在调用retrieve函数之前，我不知道形状。我可以使用retrieve函数返回形状，但我不知道如何将它们作为char_id，word_id和y的输入。我希望从feed_dict推断形状，但它们不是。

部分代码：

char_id = tf.placeholder(dtype=tf.int32, shape=[None, None])
word_id = tf.placeholder(dtype=tf.int32, shape=[batch_size, None])
y = tf.placeholder(dtype=tf.int32, shape=[batch_size, None])

with tf.name_scope("CharacterLayer"):
    with tf.variable_scope("CharacterLayer"):
        char_embeddings = tf.Variable(tf.truncated_normal(shape=[char_codes, char_embed_size]))
        char_lookup = tf.nn.embedding_lookup(char_embeddings, char_id)
        char_train = tf.unstack(value=char_lookup, axis=1)
        char_lstm_cell = rnn.BasicLSTMCell(word_embed_size, forget_bias=1)
        output_words, _ = rnn.static_rnn(cell=char_lstm_cell, inputs=char_train, dtype=tf.float32)

for i in range(1, itr):
        char_id_batch, word_id_batch, pos_id_batch = temp.retrieve_batch_sent(start, batch_size_counter)
        start = batch_size_counter
        batch_size_counter = batch_size + batch_size_counter
        sess.run(train_op, feed_dict={char_id: char_id_batch, word_id: word_id_batch, y: pos_id_batch})

我知道如果参数未指定且不可推断，则tf.unstack不起作用。这就是我收到错误的原因：ValueError: Cannot infer num from shape (?, ?, 10)。

是否有方法可以提供shapes，还是必须手动输入shapes中的placeholders？