Understanding the meaning of the weights that we get with model.state_dict() and the method .predict()

[TiToLiOn]

Hi Lu,

I'm trying to reproduct the working of the method .predict() with a function, but I have found some differences between what I thought and what I see. I explain my dude with my code, it's just a simple example to show my dude.

The structure of my net is 1 input layer with the 2 neurons variables of the pde, 2 hide layers with 5 neurons each one and 1 output layer with 1 neuron which is the value predicted of my net.

When I execute model.state_dict() I get what I expected in part, butI receive more information that I don't know how to use.

First of all I receive in order, weights and bias of each of the layers of my net

But apart from that, I got parameters called beta1 power, beta2_power and weights and bias of adam that I don't know how they matter in the prediction.

My function trying to replicate the method .predict is

It operates with the matrix of the weights and bias just of my net and composing with activation function, but ignores the rest of parametres, and I though that this was the normal way to predict. Could you explain me the meaning of the parametres beta1 power, beta2_power and weights and bias of adam, and how does the .predict() method works.

Thank you a lot.

[lululxvi]

Check the Adam paper for those variables.