Metric learning models in PyTorch, recall@1
CUB2011 | CARS196 | Stanford Online Products | |
---|---|---|---|
Margin contrastive loss, semi-hard | 0.58 @ epoch60 | 0.80 @ epoch60 | 0.7526 @ epoch90 |
Lifted structured embedding | |||
Triplet loss |
Original impl of Margin contrastive loss published at: https://github.com/apache/incubator-mxnet/tree/19ede063c4756fa49cfe741e654180aee33991c6/example/gluon/embedding_learning (temporarily removed in apache/mxnet#20602)
# evaluation results are saved in ./data/log.txt
# train margin contrastive loss on CUB2011 using ResNet-50
python train.py --dataset cub2011 --model margin --base resnet50
# download GoogLeNet weights and train using LiftedStruct loss
wget -P ./data https://github.com/vadimkantorov/metriclearningbench/releases/download/data/googlenet.h5
python train.py --dataset cub2011 --model liftedstruct --base inception_v1_googlenet
# evaluate raw final layer embeddings on CUB2011 using ResNet-50
python train.py --dataset cub2011 --model untrained --epochs 1