Ubuntu 16.04
TF version; 1.2.0.rc1

Hi Team, I'm implementing seq2seq framework using BeamSearchDecoder and got an error when using with AttentionWrapper that has no predefined time-steps.

Here's my rough code

....
self.encoder_cell = build_encoder_cell()
self.encoder_inputs = tf.placeholder(dtype=tf.int32, shape=(None, None), name='encoder_inputs')
self.encoder_inputs_length = tf.placeholder(dtype=tf.int32, shape=(None,), name='encoder_inputs_length')
....
self.encoder_outputs, self.encoder_last_state = tf.nn.dynamic_rnn(
cell=self.encoder_cell, inputs=self.encoder_inputs_embedded,
sequence_length=self.encoder_inputs_length, dtype=self.dtype,
time_major=False)
....
self.decoder_cell = build_decoder_cell()
self.decoder_cell = seq2seq.AttentionWrapper(
cell=self.decoder_cell,
attention_mechanism=self.attention_mechanism,
attention_layer_size=self.hidden_units,
cell_input_fn=attn_decoder_input_fn,
initial_cell_state=encoder_last_state,
alignment_history=False,
name='Attention_Wrapper')

tiled_batch_size = self.batch_size * self.beam_width
self.decoder_initial_state = self.decoder_cell.zero_state(batch_size=tiled_batch_size)
....
inference_decoder = beam_search_decoder.BeamSearchDecoder(cell=self.decoder_cell,
embedding=embed_and_input_proj,
start_tokens=start_tokens,
end_token=end_token,
initial_state=self.decoder_initial_state,
beam_width=self.beam_width,
output_layer=output_layer,)

running the code made an error as below,

/home/pjh/ml_tutorials/tensorflow/tf-nmt/beam_search_decoder.pyc in init(self, cell, embedding, start_tokens, end_token, initial_state, beam_width, output_layer, length_penalty_weight)
173 self._initial_cell_state = nest.map_structure(
174 self._maybe_split_batch_beams,
--> 175 initial_state, self._cell.state_size)
176 self._start_tokens = array_ops.tile(
177 array_ops.expand_dims(self._start_tokens, 1), [1, self._beam_width])

/usr/local/lib/python2.7/dist-packages/tensorflow/python/util/nest.pyc in map_structure(func, *structure, **check_types_dict)
323
324 return pack_sequence_as(
--> 325 structure[0], [func(*x) for x in entries])
326
327

/home/pjh/ml_tutorials/tensorflow/tf-nmt/beam_search_decoder.pyc in _maybe_split_batch_beams(self, t, s)
363 print t
364 print s
--> 365 return self._split_batch_beams(t, s)
366 else:
367 return t

/home/pjh/ml_tutorials/tensorflow/tf-nmt/beam_search_decoder.pyc in _split_batch_beams(self, t, s)
314 print s
315 if isinstance(s, ops.Tensor):
--> 316 s = tensor_util.constant_value_as_shape(s)
317 else:
318 s = tensor_shape.TensorShape(s)

/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/tensor_util.pyc in constant_value_as_shape(tensor)
732 A TensorShape based on the constant value of the given tensor.
733 """
--> 734 shape = tensor.get_shape().with_rank(1)
735 if tensor.get_shape() == [0]:
736 return tensor_shape.scalar()

/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/tensor_shape.pyc in with_rank(self, rank)
630 return self.merge_with(unknown_shape(ndims=rank))
631 except ValueError:
--> 632 raise ValueError("Shape %s must have rank %d" % (self, rank))
633
634 def with_rank_at_least(self, rank):

ValueError: Shape () must have rank 1

I added a log to track the tensor argument (t) and the shape argument (s) of the _split_batch_beams function
In the log, 1024 is the size of hidden units. batch_size is 80 and beam_width is 12. For each cell, I used tf.contrib.rnn.LSTMCell()

Tensor("decoder/AttentionWrapperZeroState/checked_cell_state:0", shape=(?, 1024), dtype=float32) (t)
1024 (s)
1024 (s)
Tensor("decoder/AttentionWrapperZeroState/checked_cell_state_1:0", shape=(?, 1024), dtype=float32)
1024
1024
Tensor("decoder/AttentionWrapperZeroState/zeros_1:0", shape=(960, 1024), dtype=float32)
1024
1024
Tensor("decoder/AttentionWrapperZeroState/zeros_2:0", shape=(960, ?), dtype=float32)
Tensor("decoder/LuongAttention/strided_slice_2:0", shape=(), dtype=int32)
Tensor("decoder/LuongAttention/strided_slice_2:0", shape=(), dtype=int32)

It turned out that the reason of the error was 'None' argument in the second dimension of self.encoder_inputs placeholder (which denotes max-time-steps).
I implemented self.encoder_inputs placeholder to have (None, None) shape to support dynamic batch_size and time-steps of input feeds.

When i hardcoded max-time-steps to be 80, the code worked well.
(self.encoder_inputs = tf.placeholder(tf.int32, shape=(None, ""80"")),

The input argument of _split_batch_beams was like this

Tensor("decoder/AttentionWrapperZeroState/checked_cell_state:0", shape=(?, 1024), dtype=float32)
1024
1024
Tensor("decoder/AttentionWrapperZeroState/checked_cell_state_1:0", shape=(?, 1024), dtype=float32)
1024
1024
Tensor("decoder/AttentionWrapperZeroState/zeros_1:0", shape=(960, 1024), dtype=float32)
1024
1024
Tensor("decoder/AttentionWrapperZeroState/zeros_2:0", shape=(960, ""80""), dtype=float32)
""80""
""80""

Also the error did not occur if i used normal decoder cell (not with AttentionWrapper)
I'm wondering if this is a bug or a strict design rule for BeamSearchDecoder to support only AttentionWrapper with static_time_steps