Matthew Chau

I'm having the same error. Both "pred" and "full" contain non-midi notes if I train the model on my own. No error would happen if I use the pretrained model.

I found consecutive "xxsep"s in the predicted stream. Not sure what the reason is..

I got the same problem. I solved it by making this change to `scripts/generate_song_data_beam.py:106` `p |= 'tfrecord_list' >> beam.Create(FLAGS.input)` to `p |= 'tfrecord_list' >> beam.Create([FLAGS.input])`