https://docs.google.com/spreadsheets/d/1iB9iQ7p68j7MldXvhn7xEjoj2ajJLXeJtPSKtjqXa_I/edit?usp=sharing
The overall architecture of our DCT-Transformer and Watermarking System
train from scratch (not pretraining) ---> python train.py --batch_size 32 --dataset /coco --pretrain_iter 0
train from scratch (pretraining) ---> python train.py --batch_size 32 --dataset /coco --pretrain_iter 5000
load pretrain checkpoint ---> python train.py --batch_size 32 --resume_pretrain /checkpoint/pretrain500.pyt
load train checkpoint ---> python train.py --batch_size 32 --resume /checkpoint/train500.pyt --> auto skip necst pretraining
CUDA_VISIBLE_DEVICES=[rank of GPU] python -m torch.distributed.launch --nproc_per_node=[GPU size] torch_ddp_train.py --dataset /coco
COCO Dataset please refer to https://cocodataset.org/#home
https://github.com/ando-khachatryan/HiDDeN