无法为占位符张量提供值

Question

我写了一个简单的版本双向lstm用于句子分类。但是它一直给我“您必须为占位符张量‘train_x’提供一个值”错误，这似乎来自变量初始化步骤。

data = load_data(FLAGS.data)
model = RNNClassifier(FLAGS)
init = tf.initialize_all_variables()

with tf.Session() as sess:
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)
    sess.run(init)
    print("Graph initialized..")
    print()
    np.random.seed(FLAGS.random_state)
    for epoch in range(FLAGS.max_max_epoch):

        loss = sess.run(model.cost, feed_dict={model.train_x: data.train_x, model.train_y: data.train_y, 
                                        model.embedding_placeholder: data.glove_vec})
        print("Epoch {:2d}: Loss = {:.6f} = {:.5f}".format(epoch+1, loss))
    coord.request_stop()
    coord.join(threads)

和RNNClassifier类代码(在另一个目录中)：

class RNNClassifier:

    def __init__(self, FLAGS):
        self.params = FLAGS
        with tf.device("/cpu:0"):
            self.train_x = tf.placeholder(tf.int32, [6248, 42], name='train_x')
            self.train_y = tf.placeholder(tf.int32, [6248, 3], name='train_y')
            self.embedding_placeholder = tf.placeholder(tf.float32, [1193515, 100])

        with tf.variable_scope('forward_lstm'):
            lstm_fw_cell = tf.nn.rnn_cell.LSTMCell(num_units=self.params.num_hidden, use_peepholes=False, 
                                                activation=tf.nn.relu, forget_bias=0.0, 
                                                state_is_tuple=True)
        with tf.variable_scope('backward_lstm'):
            lstm_bw_cell = tf.nn.rnn_cell.LSTMCell(num_units=self.params.num_hidden, use_peepholes=False, 
                                                activation=tf.nn.relu, forget_bias=0.0, 
                                                state_is_tuple=True)

        fw_initial_state = lstm_fw_cell.zero_state(self.params.batch_size, tf.float32)
        bw_initial_state = lstm_bw_cell.zero_state(self.params.batch_size, tf.float32)
        self._initial_state = [fw_initial_state, bw_initial_state]

        with tf.device("/cpu:0"), tf.variable_scope('softmax'):
            self.W = tf.get_variable('W', [self.params.num_hidden*2, self.params.num_classes])
            self.b = tf.get_variable('b', [self.params.num_classes], initializer=tf.constant_initializer(0.0))

        batched_inputs, batched_labels = self.batch_data()
        embed_inputs = self.use_embedding(batched_inputs)


        rnn_outputs, output_state_fw, output_state_bw  = tf.nn.bidirectional_rnn(
            cell_fw=lstm_fw_cell,
            cell_bw=lstm_bw_cell,
            inputs=embed_inputs,
            initial_state_fw=fw_initial_state,
            initial_state_bw=bw_initial_state
            )


        logits = tf.matmul(rnn_outputs[-1], self.W) + self.b

        self._cost = cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits, tf.cast(batched_labels, tf.float32)))
        optimizer = tf.train.AdamOptimizer(learning_rate=0.05).minimize(cost)


    def batch_data(self):
        # inputs = tf.convert_to_tensor(train_x, dtype=tf.int32)
        # labels = tf.convert_to_tensor(train_y, dtype=tf.int32)
        batched_inputs, batched_labels = tf.train.batch(
            tensors=[self._train_x, self._train_y],
            batch_size=self.params.batch_size,
            dynamic_pad=True,
            enqueue_many=True,
            name='batching'
    )
    return batched_inputs, batched_labels


    def use_embedding(self, batched_inputs):
        with tf.device("/cpu:0"), tf.name_scope("input_embedding"):
            embedding = tf.get_variable("embedding", shape=[1193515, 100], trainable=False)
            embedding_init = embedding.assign(self.embedding_placeholder)
            embed_inputs = tf.split(1, self.params.seq_len, tf.nn.embedding_lookup(embedding_init, batched_inputs))
            embed_inputs = [tf.squeeze(input_, [1]) for input_ in embed_inputs]
    return embed_inputs

    @property
    def cost(self):
        return self._cost

输出(包括错误)：

I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:925] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
I tensorflow/core/common_runtime/gpu/gpu_init.cc:102] Found device 0 with properties: 
name: GeForce GTX 750 Ti
major: 5 minor: 0 memoryClockRate (GHz) 1.0845
pciBusID 0000:01:00.0
Total memory: 2.00GiB
Free memory: 1.41GiB
I tensorflow/core/common_runtime/gpu/gpu_init.cc:126] DMA: 0 
I tensorflow/core/common_runtime/gpu/gpu_init.cc:136] 0:   Y 
I tensorflow/core/common_runtime/gpu/gpu_device.cc:839] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 750 Ti, pci bus id: 0000:01:00.0)
E tensorflow/core/client/tensor_c_api.cc:485] You must feed a value for placeholder tensor 'train_x' with dtype int32 and shape [6248,42]
     [[Node: train_x = Placeholder[dtype=DT_INT32, shape=[6248,42], _device="/job:localhost/replica:0/task:0/cpu:0"]()]]

Graph initialized..

W tensorflow/core/framework/op_kernel.cc:936] Out of range: PaddingFIFOQueue '_0_batching/padding_fifo_queue' is closed and has insufficient elements (requested 50, current size 0)
     [[Node: batching = QueueDequeueMany[_class=["loc:@batching/padding_fifo_queue"], component_types=[DT_INT32, DT_INT32], timeout_ms=-1, _device="/job:localhost/replica:0/task:0/cpu:0"](batching/padding_fifo_queue, batching/n)]]
W tensorflow/core/framework/op_kernel.cc:936] Out of range: PaddingFIFOQueue '_0_batching/padding_fifo_queue' is closed and has insufficient elements (requested 50, current size 0)
     [[Node: batching = QueueDequeueMany[_class=["loc:@batching/padding_fifo_queue"], component_types=[DT_INT32, DT_INT32], timeout_ms=-1, _device="/job:localhost/replica:0/task:0/cpu:0"](batching/padding_fifo_queue, batching/n)]]
E tensorflow/core/client/tensor_c_api.cc:485] PaddingFIFOQueue '_0_batching/padding_fifo_queue' is closed and has insufficient elements (requested 50, current size 0)
     [[Node: batching = QueueDequeueMany[_class=["loc:@batching/padding_fifo_queue"], component_types=[DT_INT32, DT_INT32], timeout_ms=-1, _device="/job:localhost/replica:0/task:0/cpu:0"](batching/padding_fifo_queue, batching/n)]]
     [[Node: batching/_9 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/gpu:0", send_device="/job:localhost/replica:0/task:0/cpu:0", send_device_incarnation=1, tensor_name="edge_1191_batching", tensor_type=DT_INT32, _device="/job:localhost/replica:0/task:0/gpu:0"]()]]
Traceback (most recent call last):
  File "train_lstm.py", line 66, in <module>
    model.embedding_placeholder: data.glove_vec})
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 382, in run
    run_metadata_ptr)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 655, in _run
    feed_dict_string, options, run_metadata)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 723, in _do_run
    target_list, options, run_metadata)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 743, in _do_call
    raise type(e)(node_def, op, message)
tensorflow.python.framework.errors.OutOfRangeError: PaddingFIFOQueue '_0_batching/padding_fifo_queue' is closed and has insufficient elements (requested 50, current size 0)
     [[Node: batching = QueueDequeueMany[_class=["loc:@batching/padding_fifo_queue"], component_types=[DT_INT32, DT_INT32], timeout_ms=-1, _device="/job:localhost/replica:0/task:0/cpu:0"](batching/padding_fifo_queue, batching/n)]]
     [[Node: batching/_9 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/gpu:0", send_device="/job:localhost/replica:0/task:0/cpu:0", send_device_incarnation=1, tensor_name="edge_1191_batching", tensor_type=DT_INT32, _device="/job:localhost/replica:0/task:0/gpu:0"]()]]
Caused by op u'batching', defined at:
  File "train_lstm.py", line 49, in <module>
    model = RNNClassifier(FLAGS)
  File "/home/ccrmad/Code/TDLSTM/models/rnn_classifier.py", line 34, in __init__
    batched_inputs, batched_labels = self.batch_data()
  File "/home/ccrmad/Code/TDLSTM/models/rnn_classifier.py", line 74, in batch_data
    name='batching'
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/input.py", line 595, in batch
    dequeued = queue.dequeue_many(batch_size, name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/data_flow_ops.py", line 435, in dequeue_many
    self._queue_ref, n=n, component_types=self._dtypes, name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_data_flow_ops.py", line 867, in _queue_dequeue_many
    timeout_ms=timeout_ms, name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/op_def_library.py", line 703, in apply_op
    op_def=op_def)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 2310, in create_op
    original_op=self._default_original_op, op_def=op_def)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1232, in __init__
    self._traceback = _extract_stack()

我尝试将train_x和train_y占位符初始化移到init = tf.initialize_all_variables()之前，并将它们作为两个arg提供给RNNClassifier()，但是仍然会出现相同的错误。为什么？

Answer 1

看起来问题就在这个方法中，在RNNClassifier中。

def batch_data(self):
    # ...
    batched_inputs, batched_labels = tf.train.batch(
        tensors=[self._train_x, self._train_y],
        batch_size=self.params.batch_size,
        dynamic_pad=True,
        enqueue_many=True,
        name='batching')

tf.train.batch()的两个张量参数是self._train_x和self._train_y。在RNNClassifier构造函数中，您似乎将这些创建为tf.placeholder()张量：

def __init__(self, FLAGS):
    # ...
    with tf.device("/cpu:0"):
        self.train_x = tf.placeholder(tf.int32, [6248, 42], name='train_x')
        self.train_y = tf.placeholder(tf.int32, [6248, 3], name='train_y')

...though我假设self._train_x和self.train_x之间的区别是复制粘贴错误，因为self._train_x似乎没有在其他地方定义。

现在，关于tf.train.batch()的一个令人惊讶的事情是，它在一个完全独立的线程(称为“队列运行程序”)中使用它的输入，并在调用tf.train.start_queue_runners()时启动。这个线程对依赖于占位符的子图调用Session.run()，但不知道如何填充这些占位符，所以这个调用失败了，导致了您正在看到的错误。

那你该怎么解决呢？一种选择是使用“进料队列跑步者”，对其有实验支持。一个更简单的选择是使用tf.train.slice_input_producer()从输入数据中生成切片，如下所示：

def batch_data(self, train_x, train_y):
    input_slice, label_slice = tf.train.slice_input_producer([train_x, train_y])
    batched_inputs, batched_labels = tf.train.batch(
        tensors=[input_slice, label_slice],
        batch_size=self.params.batch_size,
        dynamic_pad=False,   # All rows are the same shape.
        enqueue_many=False,  # tf.train.slice_input_producer() produces one row at a time.
        name='batching')
    return batched_inputs, batched_labels

# ...
# Create batches from the entire training data, where `array_input` and
# `array_labels` are two NumPy arrays.
batched_inputs, batched_labels = model.batch_data(array_input, array_labels)

Answer 2

这就是我做的..。

我可以将初始化输入变量的方式更改为：

data = load_data(FLAGS.data)
model = RNNClassifier(FLAGS, data)
init = tf.initialize_all_variables()

with tf.Session() as sess:
    sess.run(init)
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)
    for epoch in range(FLAGS.max_max_epoch):
        sess.run(model.train_step)
        loss, acc = sess.run([model.mean_cost, model.accuracy])
        print("Epoch {:2d}: Loss = {:.6f}; Training Accuracy = {:.5f}".format(epoch+1, loss, acc))
    print()
    coord.request_stop()
    coord.join(threads)

在RNNClassifier类中，我可以替换

    self.train_x = tf.placeholder(tf.int32, [6248, 42], name='train_x')
    self.train_y = tf.placeholder(tf.int32, [6248, 3], name='train_y')
    self.embedding_placeholder = tf.placeholder(tf.float32, [1193515, 100])

(并删除use_embedding())到

def __init__(self, FLAGS, data):
    self._train_x = tf.convert_to_tensor(data.train_x, dtype=tf.int32)
    self._train_y = tf.convert_to_tensor(data.train_y, dtype=tf.int32)
    embedding = tf.get_variable("embedding", shape=self.embedding_shape, trainable=False)
    self.embedding_init = embedding.assign(data.glove_vec)

这样，在调用队列运行程序之前，一切都是用RNNClassifier(FLAGS, data)初始化的。

我仍然不知道为什么以前的方法，也就是使用占位符，没有工作。我已经检查了dtype和数据的形状，它们都匹配。这个职位似乎也有类似的问题，“答案”建议用tf.Variable代替占位符。我已经尝试过这个->输入值现在可以输入，但是obv不像占位符那样工作，只提供初始化的值。

更新：--我认为错误可能是由RNNClassifier类中的操作输入批处理引起的，它以某种方式将输入数据输入到图形中(如果您知道错误的原因，请再次纠正我，谢谢！)相反，在初始化会话之前，我已经将批处理和嵌入函数重新定位到我的主代码中，这样可以确保所有输入都定义到主管道，批处理和预取将在tf图中处理(不使用占位符)。

这是我的密码：

def batch(x, y):
    with tf.device("/cpu:0"):
        x = tf.convert_to_tensor(x, dtype=tf.int32)
        y = tf.convert_to_tensor(y, dtype=tf.int32)
        batched_x, batched_y = tf.train.batch(
            tensors=[x, y],
            batch_size=50,
            dynamic_pad=True,
            enqueue_many=True,
            name='batching'
            )
    return (batched_x, batched_y)


def embed(df):
    with tf.device("/cpu:0"), tf.variable_scope("embed"):
        embedding = tf.get_variable("embedding", shape=df.embed_shape, trainable=False)
        embedding_init = embedding.assign(df.embed_vec)
    return embedding_init


data = load_data(FLAGS.data)
pretrained_embed = embed(data)
batched_train_x, batched_train_y = batch(data.train_x, data.train_y)
batched_test_x, batched_test_y = batch(data.train_x, data.train_y)

model = RNNClassifier(FLAGS, pretrained_embed)
logits = model.inference(batched_train_x)
test_pred = model.inference(batched_test_x, reuse=True)
loss = model.loss(logits, batched_train_y)
test_loss = model.loss(test_pred, batched_test_y)
train_op = model.training(loss[0])

init = tf.group(tf.initialize_all_variables(),
                tf.initialize_local_variables())


with tf.Session() as sess:
    t0 = time.time()
    sess.run(init)
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)
    np.random.seed(FLAGS.random_state)
    for epoch in range(FLAGS.max_max_epoch):
        _, output = sess.run([train_op, loss])
        loss_value, acc = output
        print("Epoch {:2d}: Loss = {:.6f}; Training Accuracy = {:.5f}".format(epoch+1, loss_value, acc))
    print()
    coord.request_stop()
    coord.join(threads)

RNNClassifier：

class RNNClassifier:

    def __init__(self, FLAGS, embedding_init):
        self.batch_size = FLAGS.batch_size
        self.num_hidden = FLAGS.num_hidden
        self.num_classes = FLAGS.num_classes
        self.seq_len = FLAGS.seq_len
        self.embedding_init = embedding_init

    def inference(self, batched_inputs, reuse=None):
        embed_inputs = tf.nn.embedding_lookup(self.embedding_init, tf.transpose(batched_inputs))

        with tf.variable_scope('hidden', reuse=reuse):
            with tf.variable_scope('forward_lstm'):
                lstm_fw_cell = tf.nn.rnn_cell.LSTMCell(num_units=self.num_hidden, use_peepholes=False, 
                                                    activation=tf.nn.relu, forget_bias=0.0, 
                                                    initializer=tf.random_uniform_initializer(-1.0, 1.0),
                                                    state_is_tuple=True)
                lstm_fw_cell = tf.nn.rnn_cell.DropoutWrapper(cell=lstm_fw_cell, input_keep_prob=0.7)

            with tf.variable_scope('backward_lstm'):
                lstm_bw_cell = tf.nn.rnn_cell.LSTMCell(num_units=self.num_hidden, use_peepholes=False, 
                                                    activation=tf.nn.relu, forget_bias=0.0, 
                                                    initializer=tf.random_uniform_initializer(-1.0, 1.0),
                                                    state_is_tuple=True)
                lstm_bw_cell = tf.nn.rnn_cell.DropoutWrapper(cell=lstm_bw_cell, input_keep_prob=0.7)

            fw_initial_state = lstm_fw_cell.zero_state(self.batch_size, tf.float32)
            bw_initial_state = lstm_bw_cell.zero_state(self.batch_size, tf.float32)

            rnn_outputs, output_state_fw, output_state_bw  = tf.nn.bidirectional_rnn(
                cell_fw=lstm_fw_cell,
                cell_bw=lstm_bw_cell,
                inputs=tf.unpack(embed_inputs),
                # sequence_length=self.seq_len,
                initial_state_fw=fw_initial_state,
                initial_state_bw=bw_initial_state
                )

        with tf.variable_scope('output', reuse=reuse):
            with tf.variable_scope('softmax'):
                W = tf.get_variable('W', [self.num_hidden*2, self.num_classes],
                                    initializer=tf.truncated_normal_initializer(stddev=0.1))
                b = tf.get_variable('b', [self.num_classes], initializer=tf.constant_initializer(0.1))
            logits = tf.matmul(rnn_outputs[-1], W) + b
        return logits


    def loss(self, logits, labels):
        cost = tf.nn.softmax_cross_entropy_with_logits(logits, tf.cast(labels, tf.float32))
        # self.total_cost = tf.reduce_sum(self.cost)
        mean_cost = tf.reduce_mean(cost)
        correct_pred = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
        return mean_cost, accuracy


    def training(self, cost):
        optimizer = tf.train.AdamOptimizer(learning_rate=0.001)
        train_op = optimizer.minimize(cost)
        return train_op

另外，我希望堆叠溢出的Tensorflow社区能更活跃。