Struct juice::solver::Solver

pub struct Solver<SolverB, B>
where
    SolverB: IBackend + SolverOps<f32>,
    B: IBackend + LayerOps<f32>,
{
    pub worker: Box<dyn ISolver<SolverB, B>>,
    /* private fields */
}

Solver that optimizes a [Layer][1] with a given objective. [1]: ../layer/index.html

Fields

worker: Box<dyn ISolver<SolverB, B>>

The implementation of the Solver

Implementations

impl<SolverB, B> Solver<SolverB, B>

where SolverB: IBackend + SolverOps<f32> + 'static, B: IBackend + LayerOps<f32> + 'static,

pub fn from_config( net_backend: Rc<B>, obj_backend: Rc<SolverB>, config: &SolverConfig ) -> Solver<SolverB, B>

Create Solver from [SolverConfig][1] [1]: ./struct.SolverConfig.html

This is the preferred method to create a Solver for training a neural network.

impl<SolverB, B> Solver<SolverB, B>

where SolverB: IBackend + SolverOps<f32> + 'static, B: IBackend + LayerOps<f32> + 'static,

pub fn train_minibatch( &mut self, mb_data: ArcLock<SharedTensor<f32>>, mb_target: ArcLock<SharedTensor<f32>> ) -> ArcLock<SharedTensor<f32>>

Train the network with one minibatch

pub fn network(&self) -> &Layer<B>

Returns the network trained by the solver.

This is the recommended method to get a usable trained network.

pub fn mut_network(&mut self) -> &mut Layer<B>

Returns the network trained by the solver.

This is the recommended method to get a trained network, if you want to alter the network. Keep in mind that altering the network might render the solver unusable and continuing training the network with it will yield unexpected results.

Trait Implementations

impl<SolverB, B> Debug for Solver<SolverB, B>

where SolverB: IBackend + SolverOps<f32> + Debug, B: IBackend + LayerOps<f32> + Debug,

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

Formats the value using the given formatter.