robust-synthesis

Implements image synthesis with a robust classifier based on Computer Vision with a Single (Robust) Classifier. Implements image generation, inpainting, image-to-image translation, and super-resolution on CIFAR-10 and ImageNette images. This repo is meant to be run in Google Colab.

Installation

1. Clone this repository with git clone [email protected]:ShaanGondalia/robust-synthesis.git.
2. Visit the authors' GitHub repository to download the pre-trained CIFAR-10 model. Add this model to the models folder in this repository.
   • If you want to use ImageNette, download the pre-trained ImageNet model instead.
3. Upload the root directory of the repository to your Google Drive, so that Colab can access your code.
4. Change the WORKDIR constant in main.ipynb to the path of your repository in Google Drive.

ImageNette

Download the ImageNette dataset from here. Add the downloaded data folder to the robust-synthesis/data folder. Make sure the folder is called imagenette.

Usage

Once the initial setup is complete, run main.ipynb in Google Colab. CUDA is required. See specific sections of the notebook for each synthesis technique.