tensorflow: tensorflow.nn.dynamic_rnn with variable as init_state cannot work with Estimator.train

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow):
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):Windows10
• TensorFlow installed from (source or binary):binary
• TensorFlow version (use command below):1.8.0
• Python version: 3.6
• Bazel version (if compiling from source):
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version:
• GPU model and memory:
• Exact command to reproduce:

Describe the problem

I wrote my own function to generate the init state variable

def get_initial_cell_state(cell, batch_size, dtype):
  state_size = cell.state_size
  i = 0
  def get_state_shape(s):
    c = _concat(1, s, static=True)
    nonlocal i
    name = "init_state_" + str(i)
    i = i + 1
    size = tf.get_variable(name, shape=c, dtype=dtype, initializer=tf.initializers.zeros)
    size = tf.tile(size, [batch_size] + [1] * (len(c) - 1))
    return size
  return nest.map_structure(get_state_shape, state_size)

And use it as below:

  rnn_output, _ = tf.nn.dynamic_rnn(
    cell=cell,
    inputs=inputs,
    initial_state=get_initial_cell_state(cell, batch_size=batch_size, dtype=tf.float32),
    parallel_iterations=128,
    dtype=tf.float32
  )

It’s compilable and runnable. But got this error:

WARNING:tensorflow:It seems that global step (tf.train.get_global_step) has not been increased. Current value (could be stable): 0 vs previous value: 0. You could increase the global step by passing tf.train.get_global_step() to Optimizer.apply_gradients or Optimizer.minimize.

I use Estimator to train. If I don’t use this init state variable, everything goes fine.