rotta_rs 0.1.0

use std::sync::{ Arc, Mutex };

use crate::{ rotta_rs_module::{ arrayy::Arrayy }, ShareTensor };

pub struct RMSprop {
    parameters: Arc<Mutex<Vec<ShareTensor>>>,
    pub lr: Arrayy,
    g: Vec<Arrayy>,
    pub eps: f32,
    pub hyperparameter: f32,
    pub auto_zero_grad_execute: bool,
}

impl RMSprop {
    pub fn init(parameters: Arc<Mutex<Vec<ShareTensor>>>, lr: f32) -> RMSprop {
        let lr = Arrayy::from_vector(vec![1], vec![lr]);
        RMSprop {
            parameters,
            lr,
            g: vec![],
            eps: 1e-8,
            hyperparameter: 0.9,
            auto_zero_grad_execute: true,
        }
    }

    // zero
    pub fn zero_grad(&self) {
        for node_type in self.parameters.lock().unwrap().iter() {
            node_type.zero_grad();
        }
    }

    // optimazer
    pub fn optim(&mut self) {
        for (i, node_type) in self.parameters.lock().unwrap().iter().enumerate() {
            let node = node_type;
            if let None = self.g.get(i) {
                self.g.push(Arrayy::arrayy_from_element(node.shape(), 0.0));
            }

            // g_n = g_n-1 * hyperparameter + (1 - hyperparameter) * grad(w_n)^2
            // w_n + 1 = w_n - (lr/((g_n)^0.5 + e)) * grad(w_n)

            let eps = self.eps;
            let grad = &node.grad;
            let g_n =
                &self.g[i] * self.hyperparameter +
                (1.0 - self.hyperparameter) * grad.read().unwrap().powi(2);
            let new =
                &*node.value.read().unwrap() -
                (&self.lr / (g_n.powf(0.5) + eps)) * &*grad.read().unwrap();
            node_type.update_value(new);

            self.g[i] = g_n;
        }
    }

    pub fn update_hyperparameter(&mut self, parameter: f32) {
        self.hyperparameter = parameter;
    }
}