============================

• Contrasstive Loss

• Semi-Hard Sampling

  Sampling strategy in FaceNet

• Lifted Structure Loss

• Binomial BinDeviance Loss

• Distance Weighted Sampling

• NCA Loss

  Learning a Nonlinear Embedding by Preserving Class Neighbourhood Structure -Ruslan Salakhutdinov and Geoffrey Hinton

• VTHM sampling + RAW (weight assignment) (our method)

  In submission

• Car-196

  first 98 classes as train set and last 98 classes as test set

• CUB-200-2011

  first 100 classes as train set and last 100 classes as test set

• Stanford-Online-Products

  for the experiments, we split 59,551 images of 11,318 classes for training and 60,502 images of 11,316 classes for testing

• In-Shop-clothes-Retrieval

  For the In-Shop Clothes Retrieval dataset, 3,997 classes with 25,882 images for training. And the test set are partitioned to query set with 3,985 classes(14,218 images) and gallery set with 3,985 classes (12,612 images).

Pre-trained Inceptionn-BN(inception-v2) used in most deep metric learning papers

Download site: http://data.lip6.fr/cadene/pretrainedmodels/bn_inception-239d2248.pth

wget http://data.lip6.fr/cadene/pretrainedmodels/bn_inception-239d2248.pth

mkdir pretrained_models

cp   bn_inception-239d2248.pth    pretrained_models/

• Computer with Linux or OSX
• For training, an NVIDIA GPU is strongly recommended for speed. CPU is supported but training may be slow.

The pre-trained model inception-v2 is transferred from Caffe, it can only work normally on specific version of Pytorch. Please create an env as follows:

• PyTorch : (0.2.30) (I have tried 0.3.0 and 0.1.0, performance is lower than 0.2.03 by 10% on rank@1)

If you are not required to used inception-BN, you better use my New repository is at https://github.com/bnulihaixia/VGG_dml.

Performance is nearly the same as BN-inception, training speed is a bit faster.

which can work normally on pytorch 0.4.0 +

*** VTHM+RAW :***

sh run_train_00.sh