GiilDe / adv-ddpo

Installation

Requires Python 3.10 or newer.

git clone [e-mail]:GiilDe/adv-ddpo.git
pip install -e .

Usage

First run the inversion

python scripts/invert.py --config config/cifar_inversion.py

Then, run RL training Run

accelerate launch scripts/train.py --config "config/cifar.py"

To run training on CIFAR. You can also run on MNIST by using config/base.py instead.

If you wish to experiment with pretraining, i.e learning to attack using diffusion loss, run:

python pretraining.py

After saving a pretrained model you can train it using RL by specifying its path in your config file.

The training uses wandb for logging. To see the logs, you need to click the link in the output of the training script.