search

rosinality / stylegan2-pytorch

Implementation of Analyzing and Improving the Image Quality of StyleGAN (StyleGAN 2) in PyTorch
MIT License
2.73k stars 620 forks source link

Minimum GPU RAM requirement #67

Open matigekunstintelligentie opened 4 years ago

matigekunstintelligentie commented 4 years ago

I get a CUDA_ERROR_OUT_OF_MEMORY error when trying to convert a pkl file. I only have 8GB of memory (1080 GTX). What is the minimum requirement for converting a 1024x1024 model?

The NVIDIA repository mentions 16GB. 16 GB GPUs aren't in my budget, unfortunately. Considering buying a 2080TI with 11GB of memory. I would also like to know what the maximum batch size is for GPUs with more memory. I've been able to rain with a batch size of 2.

rosinality commented 4 years ago

You could convert weights without using gpu memory. But convert_weight.py script generates some samples for sanity check, so it could cause OOM. But converted .pt checkpoints will be generated anyway.

I was able to use 8 batch sizes during training with 24 GB memory.