python - Function approximation with tensorflow, sigmoid vs relu6 -
im trying aproximate sin()
function (actually can aproximate anything) tensorflow in deep neural network, has 2 layers 10 , 5 neurons each, tried many optimizers , adam seems best 1 (also found paper recommending it)
my problem if use relu6 activation function wich recommended aproximation looks https://imgur.com/a/pcsre, has lot of edges. in other hand if use sigmoid aproximation more soft less neurons looks https://imgur.com/a/gevss , don't know why fails that.
i mapped function interval [0, 1]
before feeding train step.
any light on appreciated, in advance!
Comments
Post a Comment