virtual Type infereSigType (const vector<Type>& args)
{
assert (args.size() == arity());
interval i = args[0]->getInterval();
if (i.valid & i.lo>0) {
return castInterval(floatCast(args[0]), interval(log(i.lo), log(i.hi)));
} else {
return floatCast(args[0]);
}
}
virtual Type infereSigType (const vector<Type>& args)
{
assert (args.size() == arity());
return floatCast(args[0]);
}
ExponentialLearningRateFct(SdlFloat initialEta_ = floatCast(0.2),
SdlFloat alpha_ = floatCast(0.85)) // recommended by Tsuruoka et al. (ACL 2009)
: initialEta(initialEta_),
alpha(alpha_),
numUpdates(100) // make compiler happy by initializing
{}
ConstantLearningRateFct(SdlFloat eta_ = floatCast(0.1)) : eta(eta_) {}
/**
\param stepSize step size, this is what you should play with to speed up convergence
\param numIters expected number of iterations (calls)
*/
NocedalLearningRateFct(SdlFloat stepSize_ = floatCast(0.5))
: stepSize(stepSize_)
, alpha((SdlFloat)0.602)
, length((SdlFloat)100) // make compiler happy by initializing
{}
virtual Type infereSigType (const vector<Type>& args)
{
assert (args.size() == 1);
return castInterval(floatCast(args[0]), interval(-1,1));
}
