We present Self-Supervised Learning Integrated Multi-modal Environment (SSLIME), a toolkit based on PyTorch that aims to accelerate research cycle in self-supervised learning: from designing a new self-supervised task to evaluating the learned representations. The toolkit treats multiple data modalities (images, videos, audio, text) as first class citizens. The toolkit aims to provide reference implementations of several self-supervised pretext tasks and also provides an extensive benchmark suite for evaluating self-supervised representations. The toolkit is designed to be easily reusable, extensible and enable reproducible research. The toolkit also aims to support efficient distributed training across multiple nodes to facilitate research on Facebook scale data.
Currently, the toolkit supports the Rotation [1] Pretext task and evaluation of features from different layers. Support for Jigsaw, Colorization and DeepCluster pretext tasks will be added in the coming months.
Please find installation instructions in INSTALL.md.
After installation, please see GETTING_STARTED.md for how to run various ssl tasks.
sslime is CC-NC 4.0 International licensed, as found in the LICENSE file.
If you use sslime in your research, please use the following BibTeX entry.
@article{sslime2019,
title={SSLIME: A Toolkit for Multi-Modal Self-Supervised Training and Benchmarking},
author={Bhooshan, Suvrat and Misra, Ishan and Fergus, Rob and Goyal, Priya},
year={2019}
}
- Gidaris, Spyros, Praveer Singh, and Nikos Komodakis. "Unsupervised representation learning by predicting image rotations." arXiv preprint arXiv:1803.07728 (2018).