search

blue-oil / blueoil

Bring Deep Learning to small devices
https://blueoil.org
Apache License 2.0
248 stars 86 forks source link

Tolerance in elf binaries should be variable as argument #935

Open tk26eng opened 4 years ago

tk26eng commented 4 years ago

Now the tolerance to be tested in elf binaries like lm_x86.elf is fixed to 0.00001. It should be variable because the tolerance is up to network and dataset. Some developers try to choose the bigger tolerance in blue-oil/converter/templates/include/dlk_test.h and recompile it again when a calculated value and a ground truth doesn't match. We can improve efficiency if we can change the tolerance as argument like below.

./lm_x86.elf [input.npy] [output.npy] [tolerance(optional)]
iizukak commented 4 years ago

I want to know @primenumber 's opinion.

tk26eng commented 4 years ago

@iizukak I discussed with @primenumber about this directly. This feature to change tolerance is acceptable for him. But an warning should be shown when the user changes the tolerance because we might miss a bug with a bigger tolerance setting.

iizukak commented 4 years ago

But an warning should be shown when the user changes the tolerance because we might miss a bug with a bigger tolerance setting.

Sound good.