cross-view-asymmetric-cluster-contrastive's Introduction

Cross-view Asymmetric Cluster Contrastive

Installation

cd Dual-Cluster-Contrast
python setup.py develop

Prepare Datasets

cd examples && mkdir data

Download the person datasets Market-1501, and MSMT17 from (https://virutalbuy-public.oss-cn-hangzhou.aliyuncs.com/share/data.zip), which is provided by existing public code ClusterContrast(https://github.com/alibaba/cluster-contrast). Then unzip them under the directory like

examples/data/market1501/Market-1501-v15.09.15
exmpales/data/msmt17/MSMT17_V1

Training on supervised

CUDA_VISIBLE_DEVICES=0  python examples/main.py -b 128 -a resnet50 -d market1501 --momentum 0.1 --w 0.25 --num-instances 16 --logs-dir ./examples/market1501_supervised
CUDA_VISIBLE_DEVICES=0  python examples/main.py -b 128 -a resnet50 -d msmt17 --momentum 0.1 --w 0.5 --num-instances 16 --logs-dir ./examples/msmt17_supervised

Training on unsupervised

CUDA_VISIBLE_DEVICES=0,1  python examples/main_unsupervised.py -b 128 -a resnet50 -d market1501 --iters 400 --w 0.5 --momentum 0.1 --eps 0.4 --num-instances 16  --logs-dir ./example/market1501_unsupervised

In inspired by the Cluster-Contrast, the MSMT 17 is conducted with four gpus

CUDA_VISIBLE_DEVICES=0,1,2,3  python examples/main_unsupervised.py -b 256 -a resnet50 -d msmt17 --iters 400 --w 0.5 --momentum 0.1 --eps 0.6 --num-instances 16  --logs-dir ./example/msmt_unsupervised

The code is implemented based on the public code:https://github.com/alibaba/cluster-contrast

cross-view-asymmetric-cluster-contrastive's People

Contributors

Stargazers

Watchers

cross-view-asymmetric-cluster-contrastive's Issues

paper

Can you tell me which journal this article was submitted to? What is the status of the article now?

scalar & 温度超参数

1、class DCCLoss(nn.Module): def __init__(self, num_features, num_classes, scalar=20.0, momentum=0.0, weight=None, size_average=True,init_feat=[]):

scalar=20.0 这个参数作用是什么？

2、这两个损失计算的时候温度超参数去哪了？
loss_ccc = F.cross_entropy(inputs_ccc, targets, size_average=self.size_average)
loss_icc = F.cross_entropy(inputs_icc, targets, size_average=self.size_average)