About the graphics card usage problem

[LiJuanapple]

In the generative adversarial network, I added 2-3 layers of deformable convolution, resulting in a large graphics card footprint. We did network training on the A800, which was not possible due to insufficient graphics cards. I would like to ask how three-dimensional deformable convolution can be used and trained on a normal graphics card?
Thank you for sharing your experience so that 3D deformable convolution can be used in the construction of networks.