Struct GradientsParams 

pub struct GradientsParams { /* private fields */ }

Data type that contains gradients for parameters.

Implementations

impl GradientsParams

pub fn validate_for<B: AutodiffBackend, M: AutodiffModule<B>>( &self, module: &M, ) -> Result<(), GradientTransformError>

Check module membership, actual dimensions and device without reading values. Tied parameter IDs are counted once; absent gradients stay absent.

pub fn cast_for<B: AutodiffBackend, M: AutodiffModule<B>>( &self, module: &M, dtype: FloatDType, ) -> Result<Self, GradientTransformError>

Copy handles and optionally cast all present module gradients to F32/F64. This does not modify parameter storage, IDs or the source container.

pub fn scaled_for<B: AutodiffBackend, M: AutodiffModule<B>>( &self, module: &M, multiplier: f64, dtype: FloatDType, ) -> Result<Self, GradientTransformError>

Multiply each present gradient in an explicit work dtype, once per tied ID. Negative and zero multipliers are intentional caller-selected transforms.

pub fn unscaled_for<B: AutodiffBackend, M: AutodiffModule<B>>( &self, module: &M, divisor: f64, dtype: FloatDType, ) -> Result<Self, GradientTransformError>

Divide by a positive finite scale in F32/F64 without clearing the source. No clipping, nonfinite-step policy, optimizer update or dynamic scaling.

impl GradientsParams

pub fn try_to_record<B: Backend>( &self, ) -> Result<GradientsParamsRecord, RecorderError>

Snapshot all gradients without clearing them, including accumulated gradients.

B must be the backend used to register the gradients (normally the autodiff backend’s inner backend). Device reads complete before returning.

pub async fn to_record_async<B: Backend>( &self, ) -> Result<GradientsParamsRecord, RecorderError>

Asynchronously snapshot all gradients without clearing the container.

B must match the backend used to register the gradients.

pub fn from_record<B: Backend>( record: GradientsParamsRecord, device: &B::Device, ) -> Result<Self, RecorderError>

Restore recorded gradients to a device without summing or clearing entries.

Restore the model’s parameter IDs from the same checkpoint before using these gradients. B is the gradient backend, normally the inner backend.

impl GradientsParams

pub fn new() -> Self

Creates a new GradientsParams.

pub fn from_grads<B: AutodiffBackend, M: AutodiffModule<B>>( grads: B::Gradients, module: &M, ) -> Self

Extract each tensor gradients for the given module.

Note: This consumes the gradients. See [‘from_module’] to extract gradients only for a specific module.

pub fn from_module<B: AutodiffBackend, M: AutodiffModule<B>>( grads: &mut B::Gradients, module: &M, ) -> Self

Extract each tensor gradients for the given module.

pub fn from_params<B: AutodiffBackend, M: AutodiffModule<B>>( grads: &mut B::Gradients, module: &M, params: &[ParamId], ) -> Self

Extract tensor gradients for the given module and given parameters.

pub fn get<B, const D: usize>(&self, id: ParamId) -> Option<Tensor<B, D>>

where B: Backend,

Get the gradients for the given parameter id.

Notes

You should use remove if you want to get the gradients only one time.

pub fn remove<B, const D: usize>(&mut self, id: ParamId) -> Option<Tensor<B, D>>

where B: Backend,

Remove the gradients for the given parameter id.

pub fn register<B, const D: usize>(&mut self, id: ParamId, value: Tensor<B, D>)

where B: Backend,

Register a gradients tensor for the given parameter id.

Notes

If a tensor is already registered for the given parameter id, it will be replaced.

pub fn len(&self) -> usize

The number of gradients tensors registered.

pub fn is_empty(&self) -> bool

If any tensor is contained.

pub fn to_device<B: AutodiffBackend, M: AutodiffModule<B>>( self, device: &B::Device, module: &M, ) -> Self

Change the device of each tensor gradients registered for the given module.

Trait Implementations

impl Debug for GradientsParams

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

Formats the value using the given formatter.

impl Default for GradientsParams

fn default() -> Self

Returns the “default value” for a type.

impl From<GradientsParams> for GradAdaptor

fn from(grads: GradientsParams) -> Self

Converts to this type from the input type.