search

znxlwm / UGATIT-pytorch

Official PyTorch implementation of U-GAT-IT: Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization for Image-to-Image Translation
MIT License
2.51k stars 476 forks source link

Runtime error CUDA out of memory #75

Open IUUI11 opened 1 year ago

IUUI11 commented 1 year ago

image I was trying to run this model on 2080ti，but it always said I do not have enough GPU memory. There is no other process using GPU image

kirill-ionkin commented 1 year ago

Hello, are you sure, you check nvidia-smi, than the procces still working? I'm checking and rewriting this repo few time and this repo unfortunately has a little bit small mistackes, like CUDA out of memory when evaluating model, because there isn't with torch.no_grad().

Also, be sure, you use light version of model, only batch_size=1 and other paramaters, which are affect to the model architecture - try to decrease them until your model can stored on GPU device.

vasyza commented 1 year ago

Same problem

gogobd commented 9 months ago

Same problem, even with "--light=True" memory is running out after 1k steps... NVIDIA GeForce RTX 3060, 12GB GPU Memory...

kirill-ionkin commented 9 months ago

Same problem, even with "--light=True" memory is running out after 1k steps... NVIDIA GeForce RTX 3060, 12GB GPU Memory...

On my opinion, there is better solution to unpaired transfer style, like VSAIT: Unpaired Image Translation via Vector Symbolic Architectures for example.

gogobd commented 9 months ago

Same problem, even with "--light=True" memory is running out after 1k steps... NVIDIA GeForce RTX 3060, 12GB GPU Memory...

On my opinion, there is better solution to unpaired transfer style, like VSAIT: Unpaired Image Translation via Vector Symbolic Architectures for example.

Thanks @kirill-ionkin for pointing it out, I was trying VSAIT recently; for reference I found it here: https://github.com/facebookresearch/vsait