Network drop out and learning rate warm up requested as additional training parameters

Network dropout of about 20 percent made all the difference for my SD 1.5 models and is a powerful generalization tool. Learning rate warmup for cosine I think is essential for a finely tuned model and it would be great to adjust that.