Took me 2 hours to find out why the final output of a neural network was a bunch of NaN. This is always very annoying but I can’t really complain, it make sense. Just sucks.
That could be a nice way. Sadly it was in a C++ code base (using tensorflow). Therefore no such nice things (would be slow too). I skill-issued myself thinking a struct would be 0 -initialized but MyStruct input; would not while MyStruct input {}; will (that was the fix). Long story.
If you use the GNU libc the feenableexcept function, which you can use to enable certain floating point exceptions, could be useful to catch unexpected/unwanted NaNs
Took me 2 hours to find out why the final output of a neural network was a bunch of NaN. This is always very annoying but I can’t really complain, it make sense. Just sucks.
I hope it was garlic NaN at least.
I guess you can always just add an
assert not data.isna().any()in strategic locationsThat could be a nice way. Sadly it was in a C++ code base (using tensorflow). Therefore no such nice things (would be slow too). I skill-issued myself thinking a struct would be 0 -initialized but
MyStruct input;would not whileMyStruct input {};will (that was the fix). Long story.Oof. C++ really is a harsh mistress.
I too have forgotten to memset my structs in c++ tensorflow after prototyping in python.
Oof. This makes me appreciate the abstractions in Go. It’s a small thing but initializing structs with zero values by default is nice.
If you use the GNU libc the
feenableexceptfunction, which you can use to enable certain floating point exceptions, could be useful to catch unexpected/unwanted NaNsIf (var.nan){var = 0} my beloved.
It also depends on the context