Are you not going to make a n-pairs loss miner?

[ghost]

In the N-pairs loss paper, there is a "Hard negative class mining". But this is not implemented in this repository. Will you make this mining strategy? I just wonder it.

[nmll]

when use the N pair loss，it is increasing.may be which reason?

[KevinMusgrave]

@nmll Hard to say without more information. Perhaps your learning rate is too high?

[nmll]

Thanks，my learning rate is decaying.I do not use margin and miner,I use cos distance between the objects and then use the N pair loss,but the loss is increasing, I do not know why.the loss is increasing to 2.0 from 1,3

[KevinMusgrave]

@nmll What are you passing into the N-pairs loss?

[nmll]

@KevinMusgrave I pass into N-pairs loss the (anchors, labels,(anchor_index,pos_index,neg_index)),is the function required.

[KevinMusgrave]

@nmll Here are some other things you can try:

1. Try passing in just "anchors" and "labels":

   loss_func(anchors, labels)

2. See if the same thing happens with another loss function:

   
   from pytorch_metric_learning.losses import ContrastiveLoss
   from pytorch_metric_learning.reducers import MeanReducer

loss_func = ContrastiveLoss(reducer=MeanReducer())

3. Start with a lower learning rate.

[nmll]

Thanks a lot!