A new approach to unsupervised learning leveraging domain structure and invariance.
Paper as Roberts*, Mani*, Garg, and Lipton.
Paper as Mani*, Roberts*, Garg, and Lipton.
SlidesLive Poster Session Video
Pranav Mani*1 pmani@andrew.cmu.edu
Manley Roberts*1 manleyroberts@cmu.edu, manley@abacus.ai
Saurabh Garg1 sgarg2@andrew.cmu.edu
Zachary C. Lipton1 zlipton@cmu.edu
*: Denotes equal contribution 1: Machine Learning Department, Carnegie Mellon University
- Install a recent version of Python 3.
pip install -r requirements.txt- 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.
- In experiment_config.yml, replace "project" and "entity" with the appropriate project and entity for WandB.
Attributions are available in LICENSE_ATTRIBUTION