problem for tripletloss

[amsturdy]

Is there some problem in forward of tripletloss?
Is it possible as listed as below?

[Podidiving]

Hi. I'm also have problems understanding this implementation of triplet loss
Can you explain the logic of L.reshape
dist has shape of [BatchSIze x BatchSIze], so what is the shape of ap_dist & an_dist after applying L.reshape(dist, (0, 0, 1)) L.reshape(dist, (1, 0, -1)) respectively? I'm confused with paddle and don't know what 0 and -1 mean in L.reshape

[yongqyu]

Hello. I am not an author, but I think I can share this.
You can check it out here.
0 is the same shape as before, -1 is the new axis.
And ap and an calculate distance in every case, respectively, but as you can see in the code, they are transposed.
These two matrices are later combined for all combinations of the 3 vectors.
@amsturdy @Podidiving

[amsturdy]

The author computed "ap - an" distance firstly, but selected "pan" examples later. I think "apn" examples should be selected.

In the uploaded picture, "pa - an" distance computed firstly, and "pan" examples selected later.

[hitchhiker3010]

Hi @amsturdy
I felt there was an issue with the TripletLoss that was implemented on the original repo and made changes similar to yours but my anchor is on 1st dimension instead of the 2nd dimension that's in your change, would that make sense or do you see any challenges with that? I'm computing "ap - an" distance firstly and selecting "apn" examples later.
Would also like to understand if this change helped get better results?
Thanks in advance