Adding more tasks to AutoTVM and nnvm_integration.py


#1

I am working on auto-tuning an NNVM graph that was imported from TensorFlow. However, when tracing the code from nnvm_integration.py through topi_integration.py, only conv2d and dense are supported as tasks for AutoTVM.

It seems that many other topi operations have schedules that are tunable. For example, reduce, softmax, and pool all support the same default schedule as conv2d and dense. Is there a reason why only dense and conv2d are supported for AutoTVM?

Thanks!


#2

So far we only tune the most computationally expensive operators (e.g., conv2d) as they are likely to have the largest impact on total running time, but other operators are also tunable, provided they already have templates or you are willing to contribute them.