Integrating Transformer modules

[mantianjiang]

Currently, it seems that only samples of the shape (batch_size, features) can be obtained for training; how can we obtain samples of the shape (seq_len, batch_size, features) from the experience pool for training?

[alex-petrenko]

When training with RNNs, the learner naturally gets samples with the shape (seq_len, batch_size, features)
So this should be supported out of the box.

Try parameters --use_rnn=True --recurrence=<seq_len> --rollout=<seq_len>