Train a custom CLIP with DeepSpeed CPU offload, 16 bit precision#388

Draft

afiaka87 wants to merge 1 commit intolucidrains:mainfrom afiaka87:custom_clip_train

+293

Commits on Nov 20, 2021

(custom_clip) create train_clip.py - image text folder loader, deepspeed stage 3 round robin/gradient accumulate/cpu offload, 16-bit precision, WarmupLRDecay init, wandb logging, argparsing
Sam Sepiol
committed