dylanmuir / fmin_adam Goto Github PK

Hi,
I get an error of less input arguments when I run this code.
if(isequal(fun,'defaults'))
in this line.

Fractional Adam optimizer is available in the existing literature.

Hi,

I started to try out your package “Adam stochastic gradient descent optimization” on matlab file exchange.
https://se.mathworks.com/matlabcentral/fileexchange/61616-adam-stochastic-gradient-descent-optimization

Quite an impressive piece of work!
You might want to know that I have no extensions to Matlab.
I tried to apply it on a simple function handle with two inputs, but it failed:

function [value, gradients] = testFunk(x0)

x = x0(1);
y = x0(2);

value = (y-3)^2 + (x-4)^2;
gradients = [2x-8, 2y-6];

(and its application…)

fun = @TestFunk;
x0 = [0,0];
[x, fval, exitflag, output] = fmin_adam(fun, x0);

As I tried to debug fmin_adam, I realized that the parameter “options” is not properly set with all the required fields:

Reference to non-existent field 'GradObj'.
…
Reference to non-existent field 'DerivativeCheck'.
...
Reference to non-existent field 'UseParallel'.
…

So when I apply the following line of code from line

options = optimset(@fmin_adam);

results in the following struct:

options =

struct with fields:

    Display: 'final'
MaxFunEvals: 10000
    MaxIter: 1000000
     TolFun: 1.0000e-06
       TolX: 1.0000e-05
FunValCheck: []
  OutputFcn: []
   PlotFcns: []

If you wish more feedback on how your code performs, you might wish to add the support for the function to be called with few inputs.

Hi Dylan, thanks for sharing. One question regarding your implementation; why are you rejecting bad steps or what's your motivation for that? I guess that the original Adam paper is not suggesting it.