It is desired to submit a job to the SGE several times each running after the last one finishes. This is useful especially for training on the GPU since your job gets killed before the training finishes.

This will replace bob.learn.tensorflow!14 (merged)

ping @tiago.pereira