Skip to content

Latest commit

 

History

History
32 lines (17 loc) · 1.7 KB

README.md

File metadata and controls

32 lines (17 loc) · 1.7 KB
Raw

defense

Research on adversarial defense

Contributors: Sid, Ashwath Shetty, Anisha Pal, Dwaraknath, Sanjay, Amit Jindal

Flags

  • hparams: Hyperparameter set to use. All hyperaparams are specified in hparams/resnet.py and are objects of the class HParams. You can add more to the hyperaparams and specify in the command

  • steps: The number of steps to train for. If this is not specified, steps will be calculated from the number of epochs that are mentioned in the hparams file

  • resume: Load from existing models

  • output_dir: The output directory in which the model has to be saved. It's recommended that you don't use this option since the model gets saved with an appropriate name in the runs/ folder (automatically created).

  • use_colab: If you want to train using Google Colab, set this to True. Further instructions ahead.

Training locally

Example: python3 train.py --hparams resnet18_default

That's it.

Training on colab

  • !git clone https://<access_token>@github.com/srk97/defense.git -> Paste your GitHub access tokens without <>. You can get one from developer options in GitHub settings.

  • !cd defense && git checkout <branch_name> -> Do this only if you need to run on a branch

  • !python3 defense/train.py --hparams resnet18_default --use_colab True -> Substitute with the appropriate hparams setting

This will use your google drive to save the model under runs directory which can then be shared with others. You will see a prompt asking for a token when you execute the training script on colab.