redstone-ml 0.0.0

High-performance Machine Learning, Auto-Differentiation and Tensor Algebra crate for Rust
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

use crate::gradient_function::{GradientFuncTrait, GradientFunction};
use crate::{call_next_backward, FloatDataType, NdArray, Reshape, Tensor};
use std::cell::RefCell;
use std::rc::Rc;


pub(crate) struct BMMBackwards<T: FloatDataType> {
    pub(super) next_functions: [GradientFunction<T>; 2],

    pub(super) lhs: Rc<NdArray<'static, T>>,
    pub(super) rhs: Rc<NdArray<'static, T>>,
}


impl<T: FloatDataType> GradientFuncTrait<T> for BMMBackwards<T> {
    fn backward(&mut self, grad: &NdArray<T>) {
        call_next_backward!(grad.bmm(self.rhs.as_ref().transpose(1, 2)), 
                            self.next_functions[0]);
        
        call_next_backward!(self.lhs.as_ref().transpose(1, 2).bmm(grad), 
                            self.next_functions[1]);
    }
}

impl<T: FloatDataType> BMMBackwards<T> {
    pub(crate) fn new(lhs: &Tensor<T>, rhs: &Tensor<T>) -> GradientFunction<T> {
        Rc::new(RefCell::new(Self {
            next_functions: [lhs.grad_fn(), rhs.grad_fn()],
            lhs: lhs.get_ndarray(),
            rhs: rhs.get_ndarray()
        }))
    }
}