Struct dfdx::tensor::Gradients

pub struct Gradients<E, D: Storage<E>> { /* private fields */ }

A generic container for keeping gradients of tensors keyed by the tensor’s UniqueId.

You can:

1. Insert array values into it
2. Remove entries
3. Access references to arrays
4. Access mutable references to arrays

Implementations

impl<E, D: Storage<E>> Gradients<E, D>

pub fn leaky() -> Self

Creates a Gradients object without any leaf tensor ids. This will never drop gradients for temporary tensors.

This is why this method is called leaky, because it will keep gradients from previous passes if it is used consecutively.

You should use crate::nn::ZeroGrads::alloc_grads, which will ensure non-leaf gradients are freed after backwards.

impl<E, D: Storage<E>> Gradients<E, D>

pub fn get_or_alloc_mut<S: Shape>( &mut self, t: &impl Tensorlike<S, E, D> ) -> Result<&mut D::Vec, D::Err>

Retrieves mutable gradient for t, allocating one if it isn’t present.

pub fn try_alloc_for<S: Shape>( &mut self, t: &impl Tensorlike<S, E, D> ) -> Result<(), D::Err>

Inserts a gradient for t

pub fn retain_leafs(&mut self, ids: &[UniqueId])

Drops all gradients except for the ids specified in the parameter.

pub fn drop_non_leafs(&mut self)

Keeps all gradients marked previously by Gradients::retain_leafs, and drops all others.

pub fn get<S: Shape>(&self, t: &impl Tensorlike<S, E, D>) -> Tensor<S, E, D>

Clones the gradient and transforms it into a tensor.

Panics

If no data is associated with t yet, this will panic due to an unwrap() on a .get() to the underlying hashmap.

Trait Implementations

impl<E: Clone, D: Clone + Storage<E>> Clone for Gradients<E, D>where D::Vec: Clone,

fn clone(&self) -> Gradients<E, D>

Returns a copy of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl<E: Debug, D: Debug + Storage<E>> Debug for Gradients<E, D>where D::Vec: Debug,

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.