Unsupervised Learning of Visual Features by Contrasting Cluster Assignments
Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, Armand Joulin
In this repository, we implement SwAV in VISSL. To train a model, use the configs specified here.
To use a pre-trained SwAV ResNet-50 model, simply do:
import torch
model = torch.hub.load('facebookresearch/swav', 'resnet50')We provide several baseline SwAV pre-trained models with ResNet-50 architecture in torchvision format. We also provide models pre-trained with DeepCluster-v2 and SeLa-v2 obtained by applying improvements from the self-supervised community to DeepCluster and SeLa (see details in the appendix of our paper).
| method | epochs | batch-size | multi-crop | ImageNet top-1 acc. | url |
|---|---|---|---|---|---|
| SwAV | 800 | 4096 | 2x224 + 6x96 | 75.3 | model |
| SwAV | 400 | 4096 | 2x224 + 6x96 | 74.6 | model |
| SwAV | 200 | 4096 | 2x224 + 6x96 | 73.9 | model |
| SwAV | 100 | 4096 | 2x224 + 6x96 | 72.1 | model |
| SwAV | 200 | 256 | 2x224 + 6x96 | 72.7 | model |
| SwAV | 400 | 256 | 2x224 + 6x96 | 74.3 | model |
| SwAV | 400 | 4096 | 2x224 | 70.1 | model |
| DeepCluster-v2 | 800 | 4096 | 2x224 + 6x96 | 75.2 | model |
| DeepCluster-v2 | 400 | 4096 | 2x160 + 4x96 | 74.3 | model |
| DeepCluster-v2 | 400 | 4096 | 2x224 | 70.2 | model |
| SeLa-v2 | 400 | 4096 | 2x160 + 4x96 | 71.8 | model |
| SeLa-v2 | 400 | 4096 | 2x224 | 67.2 | model |
We provide SwAV models with ResNet-50 networks where we multiply the width by a factor ×2, ×4, and ×5.
| network | parameters | epochs | ImageNet top-1 acc. | url |
|---|---|---|---|---|
| RN50-w2 | 94M | 400 | 77.3 | model |
| RN50-w4 | 375M | 400 | 77.9 | model |
| RN50-w5 | 586M | 400 | 78.5 | model |
If you use SwAV, please use the following BibTeX entry.
@article{caron2020unsupervised,
title={Unsupervised Learning of Visual Features by Contrasting Cluster Assignments},
author={Caron, Mathilde and Misra, Ishan and Mairal, Julien and Goyal, Priya and Bojanowski, Piotr and Joulin, Armand},
booktitle={Proceedings of Advances in Neural Information Processing Systems (NeurIPS)},
year={2020}
}