how to train the model?

[nizhihao]

the model consist of four densenet, pushnet and grasp net. the author's paper say only need one Nvidia. its mean that pretrained model will no need to backward? we only train the pushnet and graspnet? My understand is right? And When I trained, whether the env is torch == 0.3.1 or torch == 1.0.0, the densenet always in train , and if i set the rotation to 16. i always out of memory when i train. and i have try to train with pytorch's DataParellel, it's also not work. What should i DO? Please Help ! Thank you !