After converting mxnet model to tvm, input shape was binded and fixed.
How can we use various input size with tvm model? It fails on ‘set_input’ method.
As performance is very dependent on the shapes of operators in the network, and changing input size will change the shapes throughout the network, we currently require recompilation when the input shapes are changed (ideally the new operators should be retuned as well).
Hi, but for object detection models with background filtering it is very common to change the input size to reduce the computation overload. So TVM is not suitable in such situation.
Can you provide some examples of such a model? It depends on how many input sizes need to be supported.