yeshaokai / robustness-aware-pruning-admm Goto Github PK

Inside the run.sh.example, it says

#admm compression in adversarial setting #python -u adv_main.py --config_file config.yaml --stage pretrain #python -u adv_main.py --config_file config.yaml --stage admm #python -u adv_main.py --config_file config.yaml --stage retrain

From my understanding, the pretrain stage will train a standard network, and admm stage will deploy weight pruning using ADMM and adversarial training on the standard network. So what is the stage retrain ? Do I need it to completely reproduce results in the paper?

Typo in config file leads to the example being unable to run.
admm_master/ADMM_examples/mnist/config.py
line 39: 'self.alpxha' should become 'self.alpha'

Thanks for the code implementation.
Can you please release the model checkpoints of the results in the paper?