use crate::err::{Error, Result};
use crate::shapes::Shape;
use crate::tensor::backend::{AutogradBackend, Backend};
use crate::tensor::base::Tensor;
use crate::tensor::dtype::DType;
use crate::tensor::grad::RequiresGrad;
pub struct Gradients<B: Backend + AutogradBackend>(B::Grads);
impl<B: Backend + AutogradBackend> Gradients<B> {
pub(crate) fn from_backend(inner: B::Grads) -> Self {
Self(inner)
}
pub fn get<S, K, G>(&self, tensor: &Tensor<S, B, K, G>) -> Result<Option<Tensor<S, B, K>>>
where
S: Shape,
K: DType,
G: RequiresGrad,
{
let Some(storage) = B::get_grad(tensor.inner(), &self.0)? else {
return Ok(None);
};
Tensor::from_gradient_storage(tensor, storage).map(Some)
}
pub fn require<S, K, G>(&self, tensor: &Tensor<S, B, K, G>) -> Result<Tensor<S, B, K>>
where
S: Shape,
K: DType,
G: RequiresGrad,
{
self.get(tensor)?.ok_or_else(|| {
Error::Backward(crate::err::BackwardError::Recipe {
operation: crate::shapes::error::OperationKind::Storage,
reason: "the backward pass produced no gradient for this tensor",
})
})
}
#[must_use]
pub fn as_backend(&self) -> &B::Grads {
&self.0
}
#[must_use]
pub fn as_backend_mut(&mut self) -> &mut B::Grads {
&mut self.0
}
#[must_use]
pub fn into_backend(self) -> B::Grads {
self.0
}
}