gdjmck / attentionbasedembeddingformetriclearning

Pytorch Implementation of paper Attention-based Ensemble forDeep Metric Learning

AttentionBasedEmbeddingForMetricLearning

Pytorch Implementation of paper Attention-based Ensemble for Deep Metric Learning

Major difference from the paper: attention maps are not followed by a sigmoid activation function and minmax norm are used instead.

The weighted sampling module code is copied from suruoxi/DistanceWeightedSampling

performance on Stanford Cars 196: 71.4% recall@1 86.9% recall@4 (8 attentions and size of each embedding is 64)

TODO:

transform attention map: att_maps = sign(att_maps) * sqrt(abs(att_maps)) before normalizing. (Motivated by tau-yihouxiang/WSDAN)

Will update here if I got better validation performance

attentionbasedembeddingformetriclearning's Issues

RuntimeError in L_metric function

Traceback (most recent call last):                                                                       
  File "train.py", line 228, in <module>                                                                 
    l_div, l_homo, l_heter = criterion.criterion(anchors, positives, negatives)                          
  File "/home/zgp/Project/ABE/criterion.py", line 48, in criterion                                       
    loss_heter = L_metric(anchors, negatives, False)                                                     
  File "/home/zgp/Project/ABE/criterion.py", line 9, in L_metric                                         
    d = torch.sum((feat1 - feat2).pow(2).view((-1, feat1.size(-1))), 1)                                  
RuntimeError: The size of tensor a (24) must match the size of tensor b (12) at non-singleton dimension 0

Please give some tips.