Siamese networks have become a common structure in various recent models for unsupervised visual representation learning. These models maximize the similarity between two augmentations of one image, subject to certain conditions for avoiding collapsing solutions. In this paper, we report surprising empirical results that simple Siamese networks can learn meaningful representations even using none of the following: (i) negative sample pairs, (ii) large batches, (iii) momentum encoders. Our experiments show that collapsing solutions do exist for the loss and structure, but a stop-gradient operation plays an essential role in preventing collapsing. We provide a hypothesis on the implication of stop-gradient, and further show proof-of-concept experiments verifying it. Our “SimSiam” method achieves competitive results on ImageNet and downstream tasks. We hope this simple baseline will motivate people to rethink the roles of Siamese architectures for unsupervised representation learning.
@inproceedings{chen2021exploring,
title={Exploring simple siamese representation learning},
author={Chen, Xinlei and He, Kaiming},
booktitle={CVPR},
year={2021}
}
Back to model_zoo.md to download models.
In this page, we provide benchmarks as much as possible to evaluate our pre-trained models. If not mentioned, all models were trained on ImageNet1k dataset.
The classification benchmarks includes 4 downstream task datasets, VOC, ImageNet, iNaturalist2018 and Places205. If not specified, the results are Top-1 (%).
The Best Layer indicates that the best results are obtained from which layers feature map. For example, if the Best Layer is feature3, its best result is obtained from the second stage of ResNet (1 for stem layer, 2-5 for 4 stage layers).
Besides, k=1 to 96 indicates the hyper-parameter of Low-shot SVM.
Self-Supervised Config | Best Layer | SVM | k=1 | k=2 | k=4 | k=8 | k=16 | k=32 | k=64 | k=96 |
---|---|---|---|---|---|---|---|---|---|---|
resnet50_8xb32-coslr-100e | feature5 | 84.21 | 39.71 | 49.65 | 62.79 | 69.97 | 74.73 | 78.30 | 81.06 | 82.44 |
resnet50_8xb32-coslr-200e | feature5 | 85.20 | 39.85 | 50.44 | 63.73 | 70.93 | 75.74 | 79.42 | 82.02 | 83.44 |
The Feature1 - Feature5 don't have the GlobalAveragePooling, the feature map is pooled to the specific dimensions and then follows a Linear layer to do the classification. Please refer to resnet50_mhead_8xb32-steplr-90e.py for details of config.
The AvgPool result is obtained from Linear Evaluation with GlobalAveragePooling. Please refer to resnet50_8xb512-coslr-90e_in1k for details of config.
Self-Supervised Config | Feature1 | Feature2 | Feature3 | Feature4 | Feature5 | AvgPool |
---|---|---|---|---|---|---|
resnet50_8xb32-coslr-100e | 15.85 | 34.02 | 46.00 | 60.90 | 67.92 | 67.88 |
resnet50_8xb32-coslr-200e | 15.57 | 37.21 | 47.28 | 62.21 | 69.85 | 69.80 |
The detection benchmarks includes 2 downstream task datasets, Pascal VOC 2007 + 2012 and COCO2017. This benchmark follows the evluation protocols set up by MoCo.
Please refer to faster_rcnn_r50_c4_mstrain_24k_voc0712.py for details of config.
Self-Supervised Config | AP50 |
---|---|
resnet50_8xb32-coslr-100e | 79.97 |
resnet50_8xb32-coslr-200e | 79.85 |
Please refer to mask_rcnn_r50_fpn_mstrain_1x_coco.py for details of config.
Self-Supervised Config | mAP(Box) | AP50(Box) | AP75(Box) | mAP(Mask) | AP50(Mask) | AP75(Mask) |
---|---|---|---|---|---|---|
resnet50_8xb32-coslr-100e | 38.3 | 57.6 | 41.7 | 34.4 | 54.8 | 36.9 |
resnet50_8xb32-coslr-200e | 38.8 | 58.0 | 42.3 | 34.9 | 55.3 | 37.6 |
The segmentation benchmarks includes 2 downstream task datasets, Cityscapes and Pascal VOC 2012 + Aug. It follows the evluation protocols set up by MMSegmentation.
Please refer to fcn_r50-d8_512x512_20k_voc12aug.py for details of config.
Self-Supervised Config | mIOU |
---|---|
resnet50_8xb32-coslr-100e | 46.11 |
resnet50_8xb32-coslr-200e | 46.27 |