System information (irrelevant for this bug)

• 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): Linux Ubuntu 16.04/Any
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): v1.3.0-rc2-20-g0787eee 1.3.0
• Python version: Python 3.5.2 :: Continuum Analytics, Inc.
• Bazel version (if compiling from source): N/A
• CUDA/cuDNN version: irrelevant
• GPU model and memory: irrelevant
• Exact command to reproduce: irrelevant

Describe the problem

tf.contrib.seq2seq.BeamSearchDecoder incorrectly truncates some of the results because the same index was previously used for a beam member that ended at a earlier step.

The root of the problem is that the while_loop body in dynamic_decode assumes that sequences are independent and will finish only once. In the same time BeamSearchDecoder creates a tree-like structure where a beam index can be reused in a later step for a state that originates from a different parent index. This causes the decoding loop to sometimes record the wrong sequence length for a beam member. Then this wrong sequence length is passed to BeamSearchDecoder.finalize which returns a truncated sequence.

Source code / logs

I use the following code to workaround the problem. This causes the right sequence to be returned but still the length returned by dynamic_decode is wrong.

class FixedBeamSearchDecoder(seq2seq.BeamSearchDecoder):
    def finalize(self, outputs, final_state, sequence_lengths):
        # BeamSearchDecoder does not follow the correct semantics of the the finished flag
        # which results in taking wrong length here and getting wrong decoded string.
        # We substitute the sequence length recorded by dynamic_decoder (which is wrong because
        # of the wrong finished flag returned by BeamSearchDecoder.step) with the length
        # recorded in BeamSearchState which is correct.
        return super().finalize(outputs, final_state, final_state.lengths)