[question] how to train on GPU with the Hausdorff Losses?

[neuronflow]

Is it possible to train on GPU with HausdorffDTLoss or HausdorffERLoss?

How did you train for https://arxiv.org/pdf/1904.10030.pdf ?

[PatRyg99]

Loss itself is always on cpu in my implementation since I use numpy. But training a net on gpu is no problem, since you only put output onto cpu once it is being calculated. I guess if you really want to calculate loss on gpu you could check up cupy library, though I don't know what it has to offer.

I haven't run any experiments in paper, I've only done an implementation based on provided description of methods. I'm not an author of a paper so it's not an official implementation - afaik there's no official implementation anywhere.

[umutdundar99]

i can recomend you some tricks to work on gpu on either tf and torch @neuronflow if you are still interested

[haibinswe]

i can recomend you some tricks to work on gpu on either tf and torch @neuronflow if you are still interested

Hello, I'm very interested in implementing HausdorffDTLoss for training 3d medical image segmentation, could you please provide some tricks for me about it?

[neuronflow]

@umutdundar99, a GPU implementation of Hausdorff loss would definitely be interesting! Apparently not only for me :)