Troubleshooting tf.contrib.seq2seq.TrainingHelper

Question

I managed to create a sequence for a sequence in tensorflow using the tf.contrib.seq2seq classes in version 1.1.

For know that I use TrainingHelper to train my model. But does this helper support previously decoded values ​​in the decoder for training or just ground rights? If this is not the case, how can I supply the previously decoded value as input to the decoder instead of ground truth values?

Answer 1

TrainingHelper delivers basic truth at every turn. If you want to use the outputs of the decoder, you can use the scheduled sampling [1]. Scheduled fetching is implemented in ScheduledEmbeddingTrainingHelper and ScheduledOutputTrainingHelper , so you can use one of two (depending on your specific application) instead of TrainingHelper . See also this topic: planned sampling in Tensorflow .

[1] https://arxiv.org/pdf/1506.03099.pdf