A new approach to unsupervised learning leveraging domain structure and invariance.
NOTE: this is a deprecated repository. Look for all future updates at this repo on the ACMI Lab Github.
Paper as Roberts*, Mani*, Garg, and Lipton.
Paper as Mani*, Roberts*, Garg, and Lipton.
SlidesLive Poster Session Video
Pranav Mani*1 [email protected]
Manley Roberts*1 [email protected]
Saurabh Garg1 [email protected]
Zachary C. Lipton1 [email protected]
*: Denotes equal contribution 1: Machine Learning Department, Carnegie Mellon University
- Install a recent version of Python 3 (we used Python 3.10.4).
pip install -r requirements.txt- You may receive errors for Pillow installation, if so follow instructions here: https://pillow.readthedocs.io/en/latest/installation.html
- Install ImageNet by the instructions at https://www.image-net.org/download.php and replace 'root folder' in ImageNet and ImageNetSubset classes in dataset.py with the root folder of the installation (one level above the train/validation split folders). The test dataset we use is composed of the validation dataset from ImageNet, the validation dataset is split out of the train dataset of ImageNet.
- Details on downloading the FieldGuide dataset can be found here https://sites.google.com/view/fgvc6/competitions/butterflies-moths-2019. Extract images from training.rar into '~/FieldGuideAllImagesDownload/'. Then run ./data_utils/create_FieldGuide_directories.ipynb to create the FieldGuide-28 and FieldGuide-2 train, val and test directories.
- Starting on line 215 of experiment_runner.py, replace "project" and "entity" with the appropriate project and entity for WandB.
- From https://github.com/wvangansbeke/Unsupervised-Classification, download CIFAR-10 SCAN Loss, CIFAR-100 SCAN Loss, and Imagenet-50 SCAN Loss pth.tar files into ./pretrain/scan_cifar_pretrain/ and ./pretrain/scan_imagenet_pretrain/, depending on the dataset.
Attributions are available in LICENSE_ATTRIBUTION