SoftMarginTripletLoss_Pytorch

This repository is based on previous works from Part_ReID.

Please refer to In Defense of the Triplet Loss for Person Re-Identification and Deeply-Learned Part-Aligned Representations for Person Re-Identification for more details.

Formulation

install

cd online_triplet_loss
sh ./mask.sh

Usage

definition

from online_triplet_loss.online_triplet_loss import OnlineTripletLoss as LossFunction

triplet_loss_fun = LossFunction(all_triplets=False,
                                positive_type=LossFunction.SampleMethod_ALL,
                                negative_type=LossFunction.SampleMethod_ALL,
                                margin_type=LossFunction.MarginType_SOFTMARGIN,
                                margin=0.3, mu=1.)

Foward

loss = triplet_loss_fun(features, labels)
accuracy = triplet_loss_fun.accuracy

optional arguments

sample strategy:

• SampleMethod_ALL
• SampleMethod_HARD
• SampleMethod_MODERATE

margin type:

• hard margin
• soft margin

mu

• This parameter is adopted for keep a balance between the intra-class loss for postive pairs and the triplet distance.
• mu=1. is highly recommend, which means that we ignore the intra-class loss.