There have been some interests in running 3D networks with TVM. Relevant discussions include: https://discuss.tvm.ai/t/conv3d-is-supported-by-tvm/4118 https://discuss.tvm.ai/t/is-conv3d-op-in-tvm-on-schedule/974
This is a call for help on bringing 3D support to TVM. This will enable new application domains such as medical and other volmetric data, video, mesh & point cloud etc. Operators that need to be adapted for 3D include convolution, pad, pooling, upsampling. The only complexity is convolution, other operators can be trivially updated by adding "depth" dimension. For convolution, we may adapt existing autotvm conv2d template to 3D, or wait for @merrymercy's autoscheduler. For cuda backend, we can start by enabling conv 3d in cuDNN. Please comment below if you are interested in this topic. -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/dmlc/tvm/issues/4009
