KiyoshiKAWASAKI/crosr

Paper CROSR official code and my modifications

Official code for "Classification-Reconstruction Learning for Open-Set Recognition"

• Paper: https://arxiv.org/abs/1812.04246
• Venue: CVPR 2019

Preparation

• Dependencies:
  -- OpenCV-Python -- Chainer
  -- LibMR
  -- Scikit-Image
• Download outlier datsets from https://github.com/facebookresearch/odin
• Modify hard-coded paths to the datasets

Usage

• train_*.py trains networks for feature extraction Example:

$ python3 train_ladder.py --model_file models/vggbtlladder.py --model_name VGGBtlLadder --rloss_weight 1

• train_openmax_multilayer.py trains the open-set classifier using features from the network trained above

• tests/eval_*.py runs evaluation

• print_score.py prints the scores

Reference

• CIFAR 10 training code in chainer https://github.com/mitmul/chainer-cifar10