nlpodyssey/spago

// Copyright 2019 spaGO Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

package gradfn

import (
    "fmt"

    "github.com/nlpodyssey/spago/mat"
)

// SoftPlus function: f(x) = 1 / β ∗ log(1 + exp(β ∗ x))
type SoftPlus[O mat.Tensor] struct {
    x         O
    beta      O
    threshold O
}

// NewSoftPlus returns a new SoftPlus Function.
func NewSoftPlus[O mat.Tensor](x O, beta, threshold O) *SoftPlus[O] {
    return &SoftPlus[O]{
        x:         x,
        beta:      beta,
        threshold: threshold,
    }
}

// Operands returns the list of operands.
func (r *SoftPlus[O]) Operands() []mat.Tensor {
    return []mat.Tensor{r.x, r.beta, r.threshold}
}

// Forward computes the output of the function.
func (r *SoftPlus[O]) Forward() (mat.Tensor, error) {
    return r.x.Value().(mat.Matrix).ApplyWithAlpha(
        softPlus,
        r.beta.Value().Item().F64(),
        r.threshold.Value().Item().F64(),
    ), nil
}

// Backward computes the backward pass.
func (r *SoftPlus[O]) Backward(gy mat.Tensor) error {
    if !mat.SameDims(r.x.Value(), gy) {
        return fmt.Errorf("fn: matrices have incompatible dimensions")
    }
    if r.x.RequiresGrad() {
        gx := r.x.Value().(mat.Matrix).ApplyWithAlpha(
            softPlusDeriv,
            r.beta.Value().Item().F64(),
            r.threshold.Value().Item().F64(),
        )
        gx.ProdInPlace(gy.(mat.Matrix))
        r.x.AccGrad(gx)
    }
    return nil
}