Training Techniques | DeepCluster, Weight Decay, SGD with Momentum |
---|---|
Architecture | 1x1 Convolution, Bottleneck Residual Block, Batch Normalization, Convolution, Global Average Pooling, Residual Block, Residual Connection, ReLU, Max Pooling, Softmax |
ID | rn50_in1k_deepclusterv2_400ep_2x160_4x96 |
SHOW MORE |
Training Techniques | DeepCluster, Weight Decay, SGD with Momentum |
---|---|
Architecture | 1x1 Convolution, Bottleneck Residual Block, Batch Normalization, Convolution, Global Average Pooling, Residual Block, Residual Connection, ReLU, Max Pooling, Softmax |
ID | rn50_in1k_deepclusterv2_400ep_2x224 |
SHOW MORE |
Training Techniques | DeepCluster, Weight Decay, SGD with Momentum |
---|---|
Architecture | 1x1 Convolution, Bottleneck Residual Block, Batch Normalization, Convolution, Global Average Pooling, Residual Block, Residual Connection, ReLU, Max Pooling, Softmax |
ID | rn50_in1k_deepclusterv2_800ep_2x224_6x96 |
SHOW MORE |
DeepClusterV2 is a self-supervision approach for learning image representations. DeepCluster iteratively groups the features with a standard clustering algorithm, k-means, and uses the subsequent assignments as supervision to update the weights of the network. The second version of DeepCluster is obtained by obtained by applying various training improvements introduced in other self-supervised learning papers. Among these improvements are the use of stronger data augmentation, MLP projection head, cosine learning rate schedule, temperature parameter, memory bank, and multi-clustering.
Get started with VISSL by trying one of the Colab tutorial notebooks.
@misc{caron2021unsupervised,
title={Unsupervised Learning of Visual Features by Contrasting Cluster Assignments},
author={Mathilde Caron and Ishan Misra and Julien Mairal and Priya Goyal and Piotr Bojanowski and Armand Joulin},
year={2021},
eprint={2006.09882},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
BENCHMARK | MODEL | METRIC NAME | METRIC VALUE | GLOBAL RANK |
---|---|---|---|---|
ImageNet | DeepClusterV2 ResNet-50 (800 epochs, 2x224+6x96) | Top 1 Accuracy | 75.18% | # 226 |
ImageNet | DeepClusterV2 ResNet-50 (400 epochs, 2x160+4x96) | Top 1 Accuracy | 74.32% | # 239 |
ImageNet | DeepClusterV2 ResNet-50 (400 epochs, 2x224) | Top 1 Accuracy | 70.01% | # 272 |