We can already extract workloads (shapes, strides, padding, etc.) from NNVM graphs to facilitate autotuning with AutoTVM. However, I don’t know the best method to dump an NNVM graph (e.g., to JSON) that preserves the workload information. This is useful for profiling purposes, so we can know how many times an operator with a given workload occurs in the graph.
Is it possible to extend the existing save JSON pass to include this information? It seems that it is not part of the attributes out of the box.