salesforce/PCL

Prototypical Contrastive Learning of Unsupervised Representations (Salesforce Research)

This is a PyTorch implementation of the PCL paper:

@inproceedings{PCL,
	title={Prototypical Contrastive Learning of Unsupervised Representations},
	author={Junnan Li and Pan Zhou and Caiming Xiong and Steven C.H. Hoi},
	booktitle={ICLR},
	year={2021}
}

Requirements:

• ImageNet dataset
• Python ≥ 3.6
• PyTorch ≥ 1.4
• faiss-gpu: pip install faiss-gpu
• pip install tqdm

Unsupervised Training:

This implementation only supports multi-gpu, DistributedDataParallel training, which is faster and simpler; single-gpu or DataParallel training is not supported.

To perform unsupervised training of a ResNet-50 model on ImageNet using a 4-gpu or 8-gpu machine, run:

python main_pcl.py \ 
  -a resnet50 \ 
  --lr 0.03 \
  --batch-size 256 \
  --temperature 0.2 \
  --mlp --aug-plus --cos (only activated for PCL v2) \	
  --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 \
  --exp-dir experiment_pcl
  [Imagenet dataset folder]

Download Pre-trained Models

Linear SVM Evaluation on VOC

To train a linear SVM classifier on VOC dataset, using frozen representations from a pre-trained model, run:

python eval_svm_voc.py --pretrained [your pretrained model] \
  -a resnet50 \ 
  --low-shot (only for low-shot evaluation, otherwise the entire dataset is used) \
  [VOC2007 dataset folder]

Linear SVM classification result on VOC, using ResNet-50 pretrained with PCL for 200 epochs:

k is the number of training samples per class.

Linear Classifier Evaluation on ImageNet

Requirement: pip install tensorboard_logger
To train a logistic regression classifier on ImageNet, using frozen representations from a pre-trained model, run:

python eval_cls_imagenet.py --pretrained [your pretrained model] \
  -a resnet50 \ 
  --lr 5 \
  --batch-size 256 \
  --id ImageNet_linear \ 
  --dist-url 'tcp://localhost:10001' --multiprocessing-distributed --world-size 1 --rank 0 \
  [Imagenet dataset folder]

Linear classification result on ImageNet, using ResNet-50 pretrained with PCL for 200 epochs: