/Learning-Outlier-Aware-Representation-with-Synthetic-Boundary-Samples

Primary LanguagePython

Learning Outlier-Aware Representation with Synthetic Boundary Samples

This is the PyTorch implementation for Learning Outlier-Aware Representation with Synthetic Boundary Samples from National Yang Ming Chiao Tung University, Taiwan.

Environment Setup

  • Install the required packages
pip install -r requirements.txt

Run Training

CUDA_VISIBLE_DEVICES=0 python -u train.py --arch resnet18 --training-mode SimCLR --dataset cifar100 --num-classes 100 --results-dir path --exp-name name --warmup --normalize --virtual-outlier --lamb 1 --near-region 0.01 --default-warmup --alpha 0.1 --lock-boundary
  • arguments:
    • --arch: model architecture
    • --training-mode: SimCLR, SupCon
    • --virtual-outlier: using synthetic boundary samples during training

Run Testing

python ./draw_his.py --training-mode SimCLR --normalize --ckpt ./compare_ckp/cifar100 --dataset cifar100 --classes 100
  • --ckpt: the folder contains the checkpoint name checkpoint_500.pth.tar

Run Tsne

python -u tsne.py --training-mode SimCLR --arch resnet18 --dataset cifar10 --normalize --run-name name --ckpt path_to_checkpoint

Extend work

False Negative Masking for Contrastive Learning

Environment Setup

  • Install the required packages
pip install -r requirements.txt

Run Training

python cls_train.py --dataset cifar10 --num-classes 10 --results-dir path --exp-name name --warmup --normalize --fnm-epoch 350
  • --result_dir: path folder for storing the evaluation result
  • --fnm-epoch: how many epochs of training before starting to use the false negative masking

Linear Evaluation

The model is evaluated by training a linear classifier after fixing the learned embedding.

python ./cls_linear.py --ckpt ckpt_path --result_dir path --dataset cifar10 --classes 10
  • --ckpt: path for the model checkpoint
  • --result_dir: path folder for storing the evaluation result