A PyTorch implementation for Understanding and Exploring the Network with Stochastic Architectures, Zhijie Deng, Yinpeng Dong, Shifeng Zhang, and Jun Zhu (NeurIPS 2020)
- python 3
- torch 1.4.0
- torchvision
Vanilla NSA:
python main_s.py --arch wrn_r --epochs 300 --schedule 90 180 240 --droppath_rate 0. --num_ensemble 100 --decay 5e-4 --arch_seed_start 1 --arch_seed_end 5 --arch_p 0.7 --aux --aux_weight 0.1 --job-id snet-1_5-r-p0.7-aux0.1-long-bn-noaggr-ba --use_bn --batch_archNSA-i:
python main_s.py --arch wrn_r --epochs 300 --schedule 90 180 240 --droppath_rate 0. --num_ensemble 100 --decay 5e-4 --arch_seed_start 1 --arch_seed_end 5 --arch_p 0.7 --aux --aux_weight 0.1 --job-id snet-1_5-r-p0.7-aux0.1-long-bn-noaggr --use_bnNSA-id (only aggr)
python main_s.py --arch wrn_r --epochs 300 --schedule 90 180 240 --droppath_rate 0. --num_ensemble 100 --decay 5e-4 --learn_aggr --arch_seed_start 1 --arch_seed_end 5 --arch_p 0.7 --aux --aux_weight 0.1 --job-id snet-1_5-r-p0.7-aux0.1-long-bn --use_bnNSA-id:
python main_s.py --arch wrn_r --epochs 300 --schedule 90 180 240 --droppath_rate 0. --num_ensemble 100 --decay 5e-4 --learn_aggr --arch_seed_start 1 --arch_seed_end 5 --arch_p 0.7 --aux --aux_weight 0.1 --job-id snet-1_5-r-p0.7-aux0.1-longWide resnet 28-10 in our implementation:
python main_s.py --arch wrn_r --epochs 300 --schedule 90 180 240 --dropout_rate 0. --num_ensemble 1 --decay 5e-4 --learn_aggr --arch_type residual --aux --aux_weight 0.1 --job-id resnet-r-aux0.1-longWide resnet 28-10 with MC dropout:
python main_s.py --arch wrn_r --epochs 300 --schedule 90 180 240 --dropout_rate 0.2 --num_ensemble 100 --decay 5e-4 --learn_aggr --arch_type residual --aux --aux_weight 0.1 --job-id resnet-r-aux0.1-dp0.2-longIndividual training:
python main_s.py --arch wrn_r --epochs 300 --schedule 90 180 240 --droppath_rate 0. --num_ensemble 100 --decay 5e-4 --learn_aggr --arch_seed_start 1 --arch_seed_end 1 --arch_p 0.7 --aux --aux_weight 0.1 --job-id snet-1_1-r-p0.7-aux0.1-longIf you have any problem about this library or want to contribute to it, please send us an Email at:
Please cite our paper if you use this code in your own work:
PlaceHolder
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent