Optimizer and .txt based captions.

Hei i noticed that the training code has no setting for optimizers, if we could use different optimizers it would be great. 
Also adding an option for image/txt pair based dataloading would be even better and maybe automatic bucketing to bring it closer to the current trainers.