You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Also, this NN is usually used for generating time-series data. As far as I've seen torchgan.trainer.Trainer specializes in images. I had to hack it a bit so that the recon parameter was optional (doesn't make much sense to generate images in my case).
Maybe a new Trainer class is in order? SequenceTrainer?
Yes seems a valid thing to do. Ideally, I would like to have a common abstraction of trainer for more variants of GAN but it is not a very simple task. For the time being, we should have a SequenceTrainer.
(I also have thoughts of deprecating the Trainer in this release for an ImageTrainer and completely removing it in a later release)
That was my next suggestion actually. GANs did focus on image generation in the beginning, but I think it's safe to say that the scenario has changed.
Last thing, what if the user wants to train with the tensors directly instead of a dataloader? This other hack I did (this one looks really really bad) solved it for me, but I think that supporting this directly in torch.trainer.BaseTrainer would make more sense.
I implemented RGAN and RCGAN from [1]. Would it make sense to add it to
torchgan.models
?[1] Esteban, Cristóbal, Stephanie L. Hyland, and Gunnar Rätsch. "Real-valued (medical) time series generation with recurrent conditional gans." arXiv preprint arXiv:1706.02633 (2017).
The text was updated successfully, but these errors were encountered: