priyankclarifai / -online-soft-mining-and-class-aware-attention-pytorch

Implementation of Weighted Contrastive Loss (Deep Metric Learning by Online Soft Mining and Class-Aware Attention)

Online-Soft-Mining-and-Class-Aware-Attention-Pytorch

Implementation of Weighted Contrastive Loss from

Deep Metric Learning by Online Soft Mining and Class-Aware Attention (https://arxiv.org/pdf/1811.01459v2.pdf)
Xinshao Wang, Yang Hua1, Elyor Kodirov, Guosheng Hu, Neil M. Robertson

I will release the tensorflow implementation of the same soon :)

For an input vector x : n x d

dist refers to the pairwise distance between normalized feature vectors, of the shape n x n, dij = dist[i][j]

A refers to the pairwise attention score Aij = min(ai , aj)

Use (person re-id):

criterion_osm_caa = OSM_CAA_Loss()
if use_gpu:
   imgs, pids = imgs.cuda(), pids.cuda()
imgs, pids = Variable(imgs), Variable(pids)
outputs, features = model(imgs)
if use_gpu:
  loss = criterion_osm_caa(features, pids , model.module.classifier.weight.t())         
else:
  loss = criterion_osm_caa(features, pids , model.classifier.weight.t())