DPImageBench/scripts/rq3.sh at main · 2019ChenGong/DPImageBench · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
# In RQ3, we investigate the conditional and unconditional pretraining of the studied algorithms using ImageNet as the pretraining dataset.

CUDA_VISIBLE_DEVICES=0,1 python run.py public_data.name=imagenet setup.n_gpus_per_node=2 --epsilon=10.0 pretrain.cond=true --method DP-MERF --data_name fmnist_28 eval.mode=val --exp_description val_condition_imagenet  &

CUDA_VISIBLE_DEVICES=0,1 python run.py public_data.name=imagenet setup.n_gpus_per_node=2 --epsilon=10.0 pretrain.cond=true --method DP-MERF --data_name cifar10_32 eval.mode=val --exp_description val_condition_imagenet  &

CUDA_VISIBLE_DEVICES=0,1 python run.py setup.n_gpus_per_node=2 public_data.name=imagenet --epsilon=10.0 pretrain.cond=true --method DPGAN --data_name fmnist_28 eval.mode=val --exp_description val_condition_imagenet  &

CUDA_VISIBLE_DEVICES=0,1 python run.py setup.n_gpus_per_node=2 public_data.name=imagenet --epsilon=10.0 pretrain.cond=true --method DPGAN --data_name cifar10_32 eval.mode=val --exp_description val_condition_imagenet  &

CUDA_VISIBLE_DEVICES=0 python run.py public_data.name=imagenet --epsilon=10.0 pretrain.cond=true --method DP-Kernel --data_name fmnist_28 eval.mode=val --exp_description val_condition_imagenet

CUDA_VISIBLE_DEVICES=0 python run.py public_data.name=imagenet --epsilon=10.0 pretrain.cond=true --method DP-Kernel --data_name cifar10_32 eval.mode=val --exp_description val_condition_imagenet

CUDA_VISIBLE_DEVICES=0 python run.py public_data.name=imagenet --epsilon=10.0 pretrain.cond=false --method DP-LDM --data_name cifar10_32 eval.mode=val --exp_description uncondition_imagenet

CUDA_VISIBLE_DEVICES=1 python run.py setup.n_gpus_per_node=1 public_data.name=null train.pretrain_model=./exp/dp-ldm/fmnist_28_eps10.0uncondition_imagenet-2025-01-25-21-52-53/pretrain/unet/checkpoints/last.ckpt eval.mode=val -m DP-LORA -dn fmnist_28 -e 10.0 -ed pretraining_uncondi

# Pretraining using a checkpoint

CUDA_VISIBLE_DEVICES=0,1,2 python run.py setup.n_gpus_per_node=3 model.ckpt=./exp/pdp-diffusion/<the-name-of-scripts>/pretrain/checkpoints/snapshot_checkpoint.pth pretrain.n_epochs=1200 public_data.name=imagenet --epsilon=10.0 pretrain.cond=false --method PrivImage --data_name cifar10_32 eval.mode=val --exp_description uncondition_imagenet_1200 &

# In RQ3, we investigate the pretraining dataset.

CUDA_VISIBLE_DEVICES=0 python run.py public_data.name=places365 --epsilon=10.0 pretrain.cond=true --method DP-MERF --data_name fmnist_28 eval.mode=val public_data.n_classes=365 public_data.train_path=dataset/places365 --exp_description val_condition_places365