TensorFlow和單詞嵌入-TypeError：不可哈希類型：'numpy.ndarray'

Question

我希望修改http://www.brightideasinanalytics.com/rnn-pretrained-word-vectors/上的代碼，該代碼用於預測下一個單詞，從而具有預測問題答案的代碼。

這是我遇到問題的代碼的摘錄：

import tensorflow.contrib as ct

def NHIDDEN():
    return 1

g = tf.Graph()
tf.reset_default_graph()

with g.as_default():
    # lines 97-104 of original code
    # RNN output node weights and biases
    weights = { 'out': tf.Variable(tf.random_normal([NHIDDEN(), embedding_dim])) }
    biases = { 'out': tf.Variable(tf.random_normal([embedding_dim])) }

    with tf.name_scope("embedding"):
        W = tf.Variable(tf.constant(0.0, shape=[vocab_size, embedding_dim]),
                    trainable=False, name="W")
        embedding_placeholder = tf.placeholder(tf.float32, [vocab_size, embedding_dim])
        embedding_init = W.assign(embedding_placeholder)
        preimage = tf.nn.embedding_lookup(W, x2)

    # lines 107-119 of original
    # reshape input data
    x_unstack = tf.unstack(preimage)

    # create RNN cells
    rnn_cell = ct.rnn.MultiRNNCell([ct.rnn.BasicLSTMCell(NHIDDEN()), ct.rnn.BasicLSTMCell(NHIDDEN())])
    outputs, states = ct.rnn.static_rnn(rnn_cell, x_unstack, dtype=tf.float32)

    # capture only the last output
    pred = tf.matmul(outputs[-1], weights['out']) + biases['out'] 

    # Create loss function and optimizer
    cost = tf.reduce_mean(tf.nn.l2_loss(pred-y))
    optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)

    # lines 130, 134 and 135 of original
    step = 0
    acc_total = 0
    loss_total = 0

    with tf.Session(graph = g) as sess:
        # lines 138, 160, 162, 175, 178 and 182 of original
        while step < 1: # training_iters:
            _,loss, pred_ = sess.run([optimizer, cost, pred], feed_dict =
                                 {x: tf.nn.embedding_lookup(W, x2), y: tf.nn.embedding_lookup(W, y)})
            loss_total += loss
            print("loss = " + "{:.6f}".format(loss_total))
            step += 1
        print ("Finished Optimization")

我得到的錯誤是：

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-7-7a72d8d4f100> in <module>()
     42         while step < 1: # training_iters:
     43             _,loss, pred_ = sess.run([optimizer, cost, pred], feed_dict =
---> 44                                      {x: tf.nn.embedding_lookup(W, x2), y: tf.nn.embedding_lookup(W, y)})
     45             loss_total += loss
     46             print("loss = " + "{:.6f}".format(loss_total))

TypeError: unhashable type: 'numpy.ndarray'

如何修復代碼？ 是因為unstack嗎？

其他上下文： x2和y分配了np.array(list(vocab_processor.transform([s])))的返回值，其中s是一個字符串（通過傳遞不同的字符串）。 embedding_dim ， vocab_size和W是使用https://ireneli.eu/2017/01/17/tensorflow-07-word-embeddings-2-loading-pre-trained-vectors/上的代碼計算的。

Answer 1

問題發生在這里： y: tf.nn.embedding_lookup(W, y) 。 feed_dict鍵應為TensorFlow圖中的占位符。 假設y是一個包含目標值的numpy.ndarray ，則可以定義一個tf.placeholder y_來將目標值饋送到網絡中，將feed_dict的相應條目feed_dict為y_: tf.nn.embedding_lookup(W, y)和相應地修改其他張量（即使用張量y_計算損耗）。