Hello! Thanks for your work! I see you have several versions of the parameters here. Do you use previously obtained parameters to get the best version? Can I autotune network with previous results?
Thank you! I autotune network with previous results? I’ve tuned model once, but I believe it’s can be faster, can we load model with this parameters and tune it again?
@iamlion12 I didn’t experience that this is possible. In other words there is no way to autotune already autotuned model. You can always modify by hand the *.log file itself, but this seems tricky.
Although, if there is any alternate way to improve already autotuned model with AutoTVM it’ll be really good to know how.
I did not try it but noticed this method ( tvm.autotvm.tuner.Tuner.load_history) in one of the tutorials. https://github.com/dmlc/tvm/blob/master/tutorials/autotvm/tune_relay_cuda.py#L199
Not sure whether it will improve the performance for already tuned model, but you should take a look.