tf.nn.static_rnn - 输入必须是一个序列

Question

我正在尝试创建 2 层 lstm（包括 dropout），但收到一条错误消息，提示“输入必须是序列”。

我使用嵌入作为输入，但不确定如何将它们更改为序列？非常感谢任何解释。

这是我的图表定义：

    with tf.name_scope('Placeholders'):
        input_x = tf.placeholder(tf.int32, [None, n_steps], name='input_x')
        input_y = tf.placeholder(tf.float32, [None, n_classes], name='input_y')
        dropout_keep_prob = tf.placeholder(tf.float32, name='dropout_keep_prob')


    with tf.name_scope('Embedding_layer'):
        embeddings_var = tf.Variable(tf.random_uniform([vocab_size, EMBEDDING_DIM], -1.0, 1.0), trainable=True)
        embedded_chars = tf.nn.embedding_lookup(embeddings_var, input_x)
        print(embedded_chars, 'embed')


    def get_a_cell(lstm_size, keep_prob):
        lstm = tf.nn.rnn_cell.BasicLSTMCell(lstm_size)
        drop = tf.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=dropout_keep_prob)
        return drop


    with tf.name_scope('lstm'):
        cell = tf.nn.rnn_cell.MultiRNNCell(
            [get_a_cell(num_hidden, dropout_keep_prob) for _ in range(num_layers)]
        )

    lstm_outputs, state = tf.nn.static_rnn(cell=cell,inputs=embedded_chars, dtype=tf.float32)

    with tf.name_scope('Fully_connected'):
        W = tf.Variable(tf.truncated_normal([num_hidden, n_classes], stddev=0.1))
        b = tf.Variable(tf.constant(0.1, shape=n_classes))
        output = tf.nn.xw_plus_b(lstm_outputs,W,b)
        predictions = tf.argmax(output, 1, name='predictions')

    with tf.name_scope('Loss'):
        # Cross-entropy loss and optimizer initialization
        loss1 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=output, labels=input_y))
        global_step = tf.Variable(0, name="global_step", trainable=False)
        optimizer = tf.train.AdamOptimizer(learning_rate=1e-3).minimize(loss1, global_step=global_step)

    with tf.name_scope('Accuracy'):
        # Accuracy metrics
        accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.round(tf.nn.softmax(output)), input_y), tf.float32))

    with tf.name_scope('num_correct'):
        correct_predictions = tf.equal(predictions, tf.argmax(input_y, 1))
        num_correct = tf.reduce_sum(tf.cast(correct_predictions, 'float'), name='num_correct')

编辑： 将 static_rnn 更改为 dynamic_rnn 时，错误消息更改为以下内容，在偏差 (b) 变量上失败：

TypeError: 'int' object is not iterable

在我将偏差项改成这样之后：

b = tf.Variable(tf.random_normal([n_classes]))

并得到一个新的错误信息：

ValueError: Shape must be rank 2 but is rank 3 for 'Fully_connected/xw_plus_b/MatMul' (op: 'MatMul') with input shapes: [?,27,128], [128,6].

Answer 1

如果我们假设您使用tf.dynamic_rnn（对于tf.static_rnn 的情况，第一个问题是因为您没有以正确的格式提供输入，tf.static_rnn 除了张量序列，例如张量列表[batch_size x seq_len] 而不是一个形状为 [batch_size x seq_len x dim] 的张量，而 tf.dynamic_rnn 处理这样的张量作为输入）

我邀请您阅读tf.nn_dynamic_rnn 的文档，以了解对于您的分类问题，您可能不想使用lstm_outputs，但state 基本上包含RNN 的最后一个输出，因为lstm_output 包含所有输出，而在这里你只对 last_output 感兴趣（除非你想对分类做一些注意力，这里你需要所有的输出）。

要获得最后一个输出，您基本上需要这样做：

lstm_outputs, state = tf.nn.dynamic_rnn(cell=cell,inputs=embedded_chars, dtype=tf.float32)
last_output = state[-1].h

state[-1] 获取最后一个单元格的状态，然后h 包含最后一个输出并将last_output 传递给您的前馈网络。

完整代码

（工作，但计算错误的准确度见 cmets）

n_classes = 6
n_steps = 27
num_hidden=128
dropout_keep_prob =0.5
vocab_size=10000
EMBEDDING_DIM=300
num_layers = 2

with tf.name_scope('Placeholders'):
    input_x = tf.placeholder(tf.int32, [None, n_steps], name='input_x')
    input_y = tf.placeholder(tf.float32, [None, n_classes], name='input_y')
    dropout_keep_prob = tf.placeholder(tf.float32, name='dropout_keep_prob')


with tf.name_scope('Embedding_layer'):
    embeddings_var = tf.Variable(tf.random_uniform([vocab_size, EMBEDDING_DIM], -1.0, 1.0), trainable=True)
    embedded_chars = tf.nn.embedding_lookup(embeddings_var, input_x)
    print(embedded_chars, 'embed')


def get_a_cell(lstm_size, keep_prob):
    lstm = tf.nn.rnn_cell.BasicLSTMCell(lstm_size)
    drop = tf.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=dropout_keep_prob)
    return drop


with tf.name_scope('lstm'):
    cell = tf.nn.rnn_cell.MultiRNNCell(
        [get_a_cell(num_hidden, dropout_keep_prob) for _ in range(num_layers)]
    )

lstm_outputs, state = tf.nn.dynamic_rnn(cell=cell,inputs=embedded_chars, dtype=tf.float32)
last_output = state[-1].h
with tf.name_scope('Fully_connected'):
    W = tf.Variable(tf.truncated_normal([num_hidden, n_classes], stddev=0.1))
    b = tf.Variable(tf.constant(0.1, shape=[n_classes]))
    output = tf.nn.xw_plus_b(last_output,W,b)
    predictions = tf.argmax(output, 1, name='predictions')

with tf.name_scope('Loss'):
    # Cross-entropy loss and optimizer initialization
    loss1 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=output, labels=input_y))
    global_step = tf.Variable(0, name="global_step", trainable=False)
    optimizer = tf.train.AdamOptimizer(learning_rate=1e-3).minimize(loss1, global_step=global_step)

with tf.name_scope('Accuracy'):
    # Accuracy metrics
    accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.round(tf.nn.softmax(output)), input_y), tf.float32))

with tf.name_scope('num_correct'):
    correct_predictions = tf.equal(predictions, tf.argmax(input_y, 1))
    num_correct = tf.reduce_sum(tf.cast(correct_predictions, 'float'), name='num_correct')