YuehLinChung / DBPCL

A Self-Supervised Learning method on Representation Learning.

Density-Based Prototypical Contrastive Learning on Visual Representations

Requirements

• PyTorch
• Pytorch-Lightning
• scikit-learn

Create Environment

• conda env create -n <env_name> environment.yml
• conda activate <env_name>

Self-supervised Training

Our code only supports single-gpu and single-machine training currently.

To perform self-supervised training of a ResNet-32 model on CIFAR-100, run:

python main.py --epochs 200 --dataset cifar100 -a resnet32 --mlp-dim 256 --out-dim 128 -b 64 -t 0.1 --con 0.1 --eps 0.3 0.5 --exp-dir logs --exp-name DBPCL_cifar100_res32_bs64_eps35_round1 --seed 1 --warmup-epoch 0 [dataset folder]

Evaluation

run python eval_cli.py --help for details.

Linear Evaluation

python eval_cli.py --method lr --bs 32 --cuda 0 --j 0 --model-path [model checkpoint file] --data-folder [dataset folder] --dataset [dataset]

Finetune

python eval_cli.py --method ann --bs 32 --cuda 0 --j 4 --model-path [model checkpoint file] --data-folder [dataset folder] --dataset [dataset] --finetune

Download Pretrained Models

pretrain dataset	epochs	backbone	link
CIFAR-10	200	ResNet-32	download
CIFAR-100	200	ResNet-32	download
STL-10	200	ResNet-32	download
Tiny-ImageNet	200	ResNet-32	download