Have a problem with build multilayer perceptron about rust-autograd HOT 3 CLOSED

What is the desired value in the dataset?

I try to predict float value by its cosine and sine. More precisely, input is [(sin(y)+1.0)/2.0, (cos(y)+1.0)/2.0]

I changed (also in gist)

let z = g.matmul(xww2, w3) + b;

to

let z = g.sigmoid(g.matmul(xww2, w3) + b);

and

let mean_loss = g.reduce_mean(g.sigmoid_cross_entropy(z, &y), &[0,1], false);

to

let mean_loss = g.reduce_mean(g.square(g.sub(z, &y)), &[0,1], false);

Now error is near 0.03 which is much better (but still not perfect). Thank you for help.
(Although, it would be very nice to have example for such cases with MSE loss function and a lot of layers and etc - for those ones who migrating from fann)

from rust-autograd.

I'm sorry I can't figure out the reason why that is, but

let values = z.eval(&[x.given(x_test.view()), y.given(y_test.view())]).unwrap();

should be like below because z is not normalized to be between 0~1.

let values = g.sigmoid(z).eval(&[x.given(x_test.view()), y.given(y_test.view())]).unwrap();

stuck on error ~ 0.585, which is very large for task detecting value from 0.0 to 1.0.

In your code, acc seems to be (binary) cross entropy. What is the desired value in the dataset?

from rust-autograd.

Although, it would be very nice to have example for such cases with MSE loss function and a lot of layers and etc - for those ones who migrating from fann)

Sorry about the poor example code.
I know that the num of use cases is important for this kind of tools, but I don't have enough time to do it...

from rust-autograd.