Pengfight / Bayesian-Data-Selection

The implement of 'Towards Accelerated Model Training via Bayesian Data Selection'

Environment Settings

• numpy==1.20.1
• scikit_learn==1.2.1
• scipy==1.6.2
• timm==0.5.4
• torch==2.0.1
• torchvision==0.15.1
• tqdm==4.59.0
• transformers==4.18.0

How to run

After you have downloaded the repository, you can train the model using Bayesian data selection by running the example script below.

• For CIFAR-10

CUDA_VISIBLE_DEVICES=0 python clip_prioritized_train_bayes_ema.py --num_epochs 200 --dataset cifar10_clip --save_name bayes_e200_tau4_alpha.2_2e2d_ema --alpha .2 --num_effective_data 200 --prior_precision 10 --tau 4

• For CIFAR-100

CUDA_VISIBLE_DEVICES=0 python clip_prioritized_train_bayes_ema.py --num_epochs 200 --dataset cifar100_clip --save_name bayes_e200_ema --alpha .3 --tau 12 --num_effective_data 1000 --prior_precision 10

• For WebVision-100

CUDA_VISIBLE_DEVICES=0 python clip_prioritized_train_bayes_ema.py --num_epochs 200 --dataset webvision --num_classes 100 --save_name bayes_e200_ema --alpha .3 --tau 10 --num_effective_data 400 --prior_precision 10