search

openai / glow

Code for reproducing results in "Glow: Generative Flow with Invertible 1x1 Convolutions"
https://arxiv.org/abs/1807.03039
MIT License
3.11k stars 515 forks source link

Any plan to make a pytorch version in the future? #3

Closed engharat closed 6 years ago

engharat commented 6 years ago

Hi, I've just read your paper and I'm impressed by the good results that you got, especially for a non adversarial model. Have you planned to make a pytorch porting in the next weeks? I imagine several things that you implemented are not still available on pytorch, so doing the porting myself would not be an easy task.

prafullasd commented 6 years ago

Thanks @engharat. We have no current plans of making a pytorch port, though would be great to see one from the community!

Amir-Arsalan commented 6 years ago

@engharat What things are not available in PyTorch? Isn't it the main missing thing 1x1 invertible convolutions?

pclucas14 commented 6 years ago

Everything seems to be available in pytorch. The torch.svd might run slower than tf.svd, however pretty much all other ops have a pytorch equivalent.

pclucas14 commented 6 years ago

@Amir-Arsalan I started to work on a pytorch implementation. In the repo you'll find a working 1x1 convolution, along with ActNorm. Hopefully I'll be able to match the paper's likelihood once I'm done. You can find the repo here https://github.com/pclucas14/pytorch-glow

asdfzxh8 commented 5 years ago

Hi, I've just read your paper and I'm impressed by the good results that you got, especially for a non adversarial model. Have you planned to make a pytorch porting in the next weeks? I imagine several things that you implemented are not still available on pytorch, so doing the porting myself would not be an easy task.

no such thing as impresx or everything or not. ts a tool. nonerx