search

MoritzKappel / HF-NHMT

Official PyTorch Implementation of 'High-Fidelity Neural Human Motion Transfer from Monocular Video'
MIT License
85 stars 11 forks source link

Gpu memory cost #2

Closed hzhao1997 closed 3 years ago

hzhao1997 commented 3 years ago

Does the default batch size equal 1？ If so, is there any solutions to lower the memory cost with no more than 11GB ？: )

MoritzKappel commented 3 years ago

Hi, the default batch size is 1 and should not be changed, as images are processed sequentially. To reduce VRAM requirements, you can train each network component individually (i.e. run the training script 3 times) by deactivating the other components in the config file.

hzhao1997 commented 3 years ago

Thanks for your quick reply and I will try it.