MiCE: Mixture of Contrastive Experts for Unsupervised Image Clustering

This repo includes the PyTorch implementation of the MiCE paper, which is a unified probabilistic clustering framework that simultaneously exploits the discriminative representations learned by contrastive learning and the semantic structures captured by a latent mixture model.

Requirements and Installation

We recommended the following dependencies.

• Python 3.7
• PyTorch (>=1.3.1)
• sciki-learn
• tensorboard-logger

Unsupervised Clustering

To train MiCE on CIFAR-10 with a single GPU, please use the following command line:

CUDA_VISIBLE_DEVICES=0 python train_MiCE.py \
  --learning_rate 1.0 --lr_decay_epochs 480,640,800 --lr_decay_rate 0.1 \  
  --model resnet34_cifar --epoch 1000  --dataset cifar10 \
  --tau 1.0 --batch_size 256  

For STL-10, the command line looks like:

CUDA_VISIBLE_DEVICES=0 python train_MiCE.py \
  --learning_rate 1.0 --lr_decay_epochs 1500,2000,2500 --lr_decay_rate 0.1 \  
  --model resnet34 --epoch 3000  --dataset stl10 \
  --tau 1.0 --batch_size 256  

Evaluation

With the trained model, we can evaluate the clustering performance of the method on the same dataset using:

CUDA_VISIBLE_DEVICES=0 python eval_MiCE.py \ 
    --model resnet34_cifar  --dataset cifar10 --nu 16384 \ 
    --test_path [path to the trained model]

Acknowledgments

Part of this code is inspired by CMC and Max-Mahalanobis Training