search

zsyzzsoft / co-mod-gan

[ICLR 2021, Spotlight] Large Scale Image Completion via Co-Modulated Generative Adversarial Networks
Other
444 stars 67 forks source link

Why did you use TensorFlow v1 for the implementation ? #37

Open usr-ein opened 3 years ago

usr-ein commented 3 years ago

Hello,

Great paper and great results !

I would really like to try out your model, but just out of curiosity, I would like to know why did you choose to implement you model using TensorFlow v1. As mentionned in other issues, it makes running your models and loading the pre-trained weights a real pain.

I understand that older models from 2017, 2018 and 2019 relied on TF v1 because it was still the main version back then, but TF v1's development has now stopped and it doesn't support newer versions of CUDA, newer GPUs, and many libraries are now reliant on a tensorflow v2.

Sorry for posting this as an issue. I just hope you can provide a public answer for all future researchers trying to replicate your architecture.

zsyzzsoft commented 3 years ago

It's because we followed the implementation of NVIDIA's stylegan2, which was using TF v1