Hello, I am tried to do the Auto Tuning on tvm, also I am using the tutorial example but when i try to use the other tuners. for example gridsearch or ga tuners. it gives the following errors : KeyError: ‘tile_k’
Could you provide more detail? I see it at least two posts. Seems a serious bug and we should fix it ASAP so that we could enter into v0.6
HI,
I am using same script from Tutorials of auto tvm which is for X86 CPU https://docs.tvm.ai/tutorials/autotvm/tune_relay_x86.html also i have changed the tuner to XGBTuner or GATuner. So when we try to run script it does tuning
2019-11-25 11:52:35,599 INFO Start to benchmark layout transformation… 2019-11-25 11:57:44,184 INFO Benchmarking layout transformation successful. 2019-11-25 11:57:44,192 INFO Start to run dynamic programming algorithm… 2019-11-25 11:57:44,193 INFO Start forward pass… 2019-11-25 11:57:44,503 INFO Finished forward pass. 2019-11-25 11:57:44,504 INFO Start backward pass… 2019-11-25 11:57:44,505 INFO Finished backward pass… 2019-11-25 11:57:44,505 INFO Finished DPExecutor run. 2019-11-25 11:57:44,507 INFO Writing optimal schedules to resnet-18_graph_opt.log successfully. Compile…
after this it gives
KeyError: ‘tile_k’
Maybe you can try to turn off the graph optimization part and see if the error persists or not. Similar to this tutorial that uses ApplyHistoryBest.
Have tried with graph optimization off, it still gives the same error