search

stelzner / osrt

Independent PyTorch Implementation of Object Scene Representation Transformer
MIT License
46 stars 8 forks source link

Model and the number of GPUs used to acquire the trained model for teaser image. #1

Closed mch5048 closed 1 year ago

mch5048 commented 1 year ago

Hi, thanks for reproducing this interesting work. I am curious about the training setup of the reproduced model, as the original OSRT model is computation-hungry that requires 64 TPUs. This model is implemented in Pytorch, but the config file seems to set to 256 batch size which is on part with the original model (though the number of sampling rays are 1/4).

Thanks in advance!

stelzner commented 1 year ago

Hi, we ran the model as defined in the config file using a DGX machine with 8 A100 GPUs with 80GB VRAM each. I've added that information to the README.

mch5048 commented 1 year ago

I see. Thanks for the quick response.