custom gluonts framework with triplet loss

[sy2657]

Hi, I was interested in implementing a loss function called a triplet loss function. I see that F is a parameter in the function of hybrid_forward in the definition of the custom TrainNetwork and PredNetwork. Where do I define F and how can i operate a loss on 3 inputs ?

Here is more detail about triplet loss function (https://arxiv.org/pdf/1901.10738.pdf) :

[StatMixedML]

Very relevant topic. Hope it is being picked!

[AaronSpieler]

Hello,

have you come up with any solution?

F is just the data/tensor type, i.e. mx.NDarray or mx.Symbol. It's used to call the appropriate functions.

Fundamental difficulties that I see:

• Our data loader essentially is not flexible enough right now to allow for the appropriate sampling of sequences