so ML is basically just floating point maths right could we make rustc do rnn training at compile time with const fn?
-
No, the problem is that we can't quite even do that. Because different targets have different default NaN values and different precisions. Which means comp time and run time can have differences, which is "a problem yet to be solved".
Absolutely - I don't mean "we can solve the problem of being able to do floating point maths at compile time by doing it in software", I mean "we can solve the problem of needing to do floating point maths for ML by doing that maths with our own const fn float implementation"
-
-
I'm not sure I follow. Maybe I don't know about ML enough. Wouldn't there still be float math to do at runtime as well?
-
Yes! But then we would let the processor do that bit. The point is that the training of the set would be done in compile time. For the record, this would not help anyone with anything. It's just something I was thinking about as a curiosity.
Ende der Unterhaltung
Neue Unterhaltung -
Das Laden scheint etwas zu dauern.
Twitter ist möglicherweise überlastet oder hat einen vorübergehenden Schluckauf. Probiere es erneut oder besuche Twitter Status für weitere Informationen.