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"
)

// SwishB function: f(x) = x * sigmoid.
//
// Reference: "Searching for Activation Functions" by Ramachandran et al, 2017.
// (https://arxiv.org/pdf/1710.05941.pdf)
type SwishB[O mat.Tensor] struct {
    x    O
    beta O // scalar
}

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

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

// Forward computes the output of the function.
func (r *SwishB[O]) Forward() (mat.Tensor, error) {
    y := r.x.Value().(mat.Matrix).ApplyWithAlpha(swishB, r.beta.Value().Item().F64())
    return y, nil
}

// Backward computes the backward pass.
func (r *SwishB[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(swishBDeriv, r.beta.Value().Item().F64())
        gx.ProdInPlace(gy.(mat.Matrix))
        r.x.AccGrad(gx)
    }
    if r.beta.RequiresGrad() {
        gb := r.beta.Value().(mat.Matrix).ZerosLike()
        // FIXME: avoid casting to specific type
        for i, x := range r.x.Value().Data().F64() {
            deriv := swishBBetaDeriv(x, r.beta.Value().Item().F64())
            gyi := gy.(mat.Matrix).ScalarAt(i).F64()
            gb.AddScalarInPlace(deriv * gyi)
        }
        r.beta.AccGrad(gb)
    }
    return nil
}