relay.frontend.from_pytorch set fixed input size, but I need input size can change at inference，is there any way to handle this?
No, the input shape needs to be fixed.
ok, thinks ,any way.
Dynamic input shape will involve dynamic shape kernel codegen, which is still WIP.
What’s the latest status? Could relay.frontend.from_pytorch support pytorch danamic input shape now?
No it is not supported. Dynamic codegen is possible but currently the compiled model would run extremely slow, so I don’t think it is worth supporting for now.
So for pytorch dynamic shape do you usually convert pytorch model to tensorflow/onnx for tvm supporting?
I’ve never had a need to support dynamic shape input model.
Any updates ? I noticed the vm can support dynamic batch/shape now. Is it possible to import dynamic shape pytorch model now ?