Question

我正在使用TensorFlow版本0.9实现双向标记GRU网络（向前1层，向后1层）。在初始化模型时，TensorFlow初始化所有变量，创建GRU单元并正确应用所有常规转换，直到运行tf.nn.bidirectional_rnn函数时，它会抛出与错误形状的Tensor合并相关联的ValueError。操作。这是代码：

# Create the cells
with tf.variable_scope('forward'):
    self.char_gru_cell_fw = tf.nn.rnn_cell.GRUCell(char_hidden_size)
with tf.variable_scope('backward'):
    self.char_gru_cell_bw = tf.nn.rnn_cell.GRUCell(char_hidden_size)

# Set initial state of the cells to be zero
self._char_initial_state_fw = \
    self.char_gru_cell_fw.zero_state(batch_size, tf.float32)
self._char_initial_state_bw = \
    self.char_gru_cell_bw.zero_state(batch_size, tf.float32)

#         Size before: batch-chrpad-chrvocabsize
#          Size after: batch-chrvocabsize
chargruinput = [tf.squeeze(input_, [1]) \
    for input_ in tf.split(1, char_num_steps, chargruinput)]

# Run the bidirectional rnn and get the corner results
_, output_state_fw, output_state_bw = \
   tf.nn.bidirectional_rnn(self.char_gru_cell_fw, 
                    self.char_gru_cell_bw, 
                    chargruinput, 
                    sequence_length=char_num_steps,
                    initial_state_fw=self._char_initial_state_fw,
                    initial_state_bw=self._char_initial_state_bw)

当我运行它时，我收到以下错误：

Traceback (most recent call last):
  File "frontbackgru.py", line 409, in <module>
    main()
  File "frontbackgru.py", line 226, in main
    config=my_config)
  File "/home/xG/Code/4-RNN/1-simple-cnn-input-classifier/gru_model.py", line 265, in __init__
    initial_state_bw=self._char_initial_state_bw)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 453, in bidirectional_rnn
    sequence_length, scope=fw_scope)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 156, in rnn
    state_size=cell.state_size)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 343, in _rnn_step
    _maybe_copy_some_through)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 1331, in cond
    _, res_f = context_f.BuildCondBranch(fn2)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 1230, in BuildCondBranch
    r = fn()
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 317, in _maybe_copy_some_through
    lambda: _copy_some_through(new_output, new_state))
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 1331, in cond
    _, res_f = context_f.BuildCondBranch(fn2)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 1230, in BuildCondBranch
    r = fn()
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 317, in <lambda>
    lambda: _copy_some_through(new_output, new_state))
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/rnn.py", line 298, in _copy_some_through
    return ([math_ops.select(copy_cond, zero_output, new_output)]
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_math_ops.py", line 1769, in select
    name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/op_def_library.py", line 704, in apply_op
    op_def=op_def)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 2262, in create_op
    set_shapes_for_outputs(ret)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1702, in set_shapes_for_outputs
    shapes = shape_func(op)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/math_ops.py", line 1578, in _SelectShape
    t_e_shape = t_e_shape.merge_with(c_shape)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/tensor_shape.py", line 570, in merge_with
    (self, other))
ValueError: Shapes (32, 50) and () are not compatible

现在，bidirectional_rnn函数的输入是：

self.char_gru_cell_fw：这是一个用整数值char_hidden_size初始化的GRUCell实例，在这种情况下为50

self.char_gru_cell_bw：这是一个用整数值char_hidden_size初始化的GRUCell实例，在这种情况下为50

chargruinput：这是一个长度为30的列表，包含形状[batch_size，charvocab]的张量，在这种情况下为[32,256]

sequence_length：一个整数，表示展开的单元格数char_num_steps，在这种情况下为30。

initial_state_fw：与GRU状态形状相同的零填充张量，在这种情况下为[32,50]

initial_state_bw：与GRU状态形状相同的零填充张量，在这种情况下为[32,50]

我尝试查看导致抛出ValueError异常的模块，但是有很多低级别的东西正在发生，这很可能正常工作，看看我上周工作的CNN是如何工作的，没有任何问题。这让我觉得在低级方法之前，我之前没有用过的rnn或rnn_cell库出了问题。

它似乎也很奇怪，因为错误与空形状有关（与标量相关而不是我认为的Tensor），但我唯一可以改变的是bidirectional_rnn函数中的标量arguments是sequence_length参数。我试图省略它并仅使用初始状态，反之亦然，但会弹出相同的错误。

有没有人有类似的问题？我的整个系统都因此而瘫痪，会喜欢一些反馈。提前致谢

Answer 1

弄清楚出了什么问题 - 参数sequence_length实际上应该是每个批次的长度为batch_size的整数列表，而不是整数本身。轻松修复，感谢您的演奏：）

TensorFlow双向GRU返回ValueError，因为<thinkly>形状不正确

1 个答案: