I am trying to deploy a 3d conv onnx model, however, a RuntimeError occurs, " Check failed: Only 2d kernel supported." It seems that tvm currently doesnt support Conv3D? or I just dont know how to do it? Could anyone help me? THX!
No, we don’t have support for 3D models. For that we need to update all relevant operators (conv, pooling, upsampling etc) to 3D.
This seems to be a good feature that the community can work on @masahi can you create a call for help thread on the issue tracker?
Created https://github.com/dmlc/tvm/issues/4009
I’m wondering if we can expect @merrymercy’s autoscheduler to come in near future … can somebody in the know comment?
Hi @tqchen, @masahi , I am new to tvm community. I was trying to implement maxpool3d operator based on reference implementation from tensorflow(r=1.14) into tvm codebase. For testing the same, I have written relay and topi test cases. I am also writing test cases in test_forward.py(for tensorflow) which compares tvm output with tensorflow output. I want to know is there any specific reason why you have not added any changes in test_forward.py(comparing tvm with tf) for unit testing?? Should i ignore this file altogether?