I used tensorflow to do some voxel feature detection, but inference is slow on CPU, so I resort to tvm for chance to accelorate it, but find no implementation is available yet. I wonder if you guys have the plan to add con3d in tvm?
1 Like
Conv3d is interesting. We will look into it, but there is no specific timeline. Maybe we will bring an auto-scheduler for general n-dimensional convolution.
This is definitely an interesting problem, note that given the current conv2d template, it should be possible to derive a simple autotvm conv3d template, would you be interested in trying it out and contribute back?
Look forward to the update for conv3d and deconv3d. Thanks.