Problem with training data format

[Tubakabura]

My initial training data was in 'mixed' format, while the data generated from each 'run' iteration was converted to the regular format of deepmd.
Is there any way to control the format of the data generated in each iteration? Otherwise I had to run a python process to convert the data before each iteration.

[njzjz]

Alternately, you can set batch_size to mixed:N (e.g. mixed:16) which will mix systems automatically during training.

If it can not meet your requirements, PR is welcome.