additional changes for lower memory usage

[Thomas-MMJ]

lower batch size with higher gradient accumulation uses less memory for same training benefit; use_8bit_adam greatly reduces memory for adam; gradient_checkpointing greatly reduces memory; mixed_precision bf16 is faster for same memory usage.

Note that I can't test these locally yet (on windows) so not positive they are all of benefit. Will probably test tomorrow on colab.

[1blackbar]

Theres a colab cell missing to download trained dreambooth bin file , also cell missing to prune it, for newbies this is kinda not useful at all if you cant downoad model, theres 3 bin files , takes ages to download one just to figure out its not the right one....
Results on t4 are pretty bad compared to tex iversion , i have colabpro with p100 that might be better (11 tflops)but training on t4 ( 6 tflops) needs to be like on cards with 100 tflops where it takes 15 minutes, with t4 it should take equivalent so 240 minutes for tfloppage they have , maybe changing that would get better results, at the moment this is not working really to get likeness of a person into SD
https://colab.research.google.com/github/ShivamShrirao/diffusers/blob/main/examples/dreambooth/DreamBooth_Stable_Diffusion.ipynb

Also theres xformers precompiled for p100 on colab pro here, can you include them ?
https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb

Please dont train on "sks"as its a gun ,better use something random like tgsdswetafa , theres no personalisation to change that
i think --tokenizer_name= is missing and its important
can You convert to ckpt ? thers no way to use bin files in webui