they also appear to be relying on the tensorboard callback which seems not to work on distributed training