ODC¶
Online Deep Clustering for Unsupervised Representation Learning¶
Joint clustering and feature learning methods have shown remarkable performance in unsupervised representation learning. However, the training schedule alternating between feature clustering and network parameters update leads to unstable learning of visual representations. To overcome this challenge, we propose Online Deep Clustering (ODC) that performs clustering and network update simultaneously rather than alternatingly. Our key insight is that the cluster centroids should evolve steadily in keeping the classifier stably updated. Specifically, we design and maintain two dynamic memory modules, i.e., samples memory to store samples’ labels and features, and centroids memory for centroids evolution. We break down the abrupt global clustering into steady memory update and batch-wise label re-assignment. The process is integrated into network update iterations. In this way, labels and the network evolve shoulder-to-shoulder rather than alternatingly. Extensive experiments demonstrate that ODC stabilizes the training process and boosts the performance effectively.
Citation¶
@inproceedings{zhan2020online,
title={Online deep clustering for unsupervised representation learning},
author={Zhan, Xiaohang and Xie, Jiahao and Liu, Ziwei and Ong, Yew-Soon and Loy, Chen Change},
booktitle={CVPR},
year={2020}
}
Models and Benchmarks¶
Back to model_zoo.md
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.
VOC SVM / Low-shot SVM¶
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.
| Model | Config | Best Layer | SVM | k=1 | k=2 | k=4 | k=8 | k=16 | k=32 | k=64 | k=96 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| model | resnet50_8xb64-steplr-440e |
ImageNet Linear Evaluation¶
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 file name for details of config.
| Model | Config | Feature1 | Feature2 | Feature3 | Feature4 | Feature5 | AvgPool |
|---|---|---|---|---|---|---|---|
| model | resnet50_8xb64-steplr-440e |
iNaturalist2018 Linear Evaluation¶
Please refer to resnet50_mhead_8xb32-steplr-84e_inat18.py and file name for details of config.
| Model | Config | Feature1 | Feature2 | Feature3 | Feature4 | Feature5 | AvgPool |
|---|---|---|---|---|---|---|---|
| model | resnet50_8xb64-steplr-440e |
Places205 Linear Evaluation¶
Please refer to resnet50_mhead_8xb32-steplr-28e_places205.py and file name for details of config.
| Model | Config | Feature1 | Feature2 | Feature3 | Feature4 | Feature5 | AvgPool |
|---|---|---|---|---|---|---|---|
| model | resnet50_8xb64-steplr-440e |
Semi-Supervised Classification¶
In this benchmark, the necks or heads are removed and only the backbone CNN is evaluated by appending a linear classification head. All parameters are fine-tuned.
When training with 1% ImageNet, we find hyper-parameters especially the learning rate greatly influence the performance. Hence, we prepare a list of settings with the base learning rate from
{0.001, 0.01, 0.1}and the learning rate multiplier for the head from{1, 10, 100}. We choose the best performing setting for each method. The setting of parameters are indicated in the file name. The learning rate is indicated like1e-1,1e-2,1e-3and the learning rate multiplier is indicated likehead1,head10,head100.Please use –deterministic in this benchmark.
Please refer to the directories configs/benchmarks/classification/imagenet/imagenet_1percent/ of 1% data and configs/benchmarks/classification/imagenet/imagenet_10percent/ 10% data for details.
| Model | Pretrain Config | Fine-tuned Config | Top-1 (%) | Top-5 (%) |
|---|---|---|---|---|
| model | resnet50_8xb64-steplr-440e |
Detection¶
The detection benchmarks includes 2 downstream task datasets, Pascal VOC 2007 + 2012 and COCO2017. This benchmark follows the evluation protocols set up by MoCo.
Pascal VOC 2007 + 2012¶
Please refer to faster_rcnn_r50_c4_mstrain_24k.py for details of config.
| Model | Config | mAP | AP50 |
|---|---|---|---|
| model | resnet50_8xb64-steplr-440e |
COCO2017¶
Please refer to mask_rcnn_r50_fpn_mstrain_1x.py for details of config.
| Model | Config | mAP(Box) | AP50(Box) | AP75(Box) | mAP(Mask) | AP50(Mask) | AP75(Mask) |
|---|---|---|---|---|---|---|---|
| model | resnet50_8xb64-steplr-440e |
Segmentation¶
The segmentation benchmarks includes 2 downstream task datasets, Cityscapes and Pascal VOC 2012 + Aug. It follows the evluation protocols set up by MMSegmentation.
Pascal VOC 2012 + Aug¶
Please refer to file for details of config.
| Model | Config | mIOU |
|---|---|---|
| model | resnet50_8xb64-steplr-440e |
Cityscapes¶
Please refer to file for details of config.
| Model | Config | mIOU |
|---|---|---|
| model | resnet50_8xb64-steplr-440e |