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

// Dropout is an operator to perform elements dropout with a probability.
type Dropout[O mat.Tensor] struct {
    x       O
    prob    float64
    q       float64 // 1 - p
    randGen *rand.LockedRand
    mask    mat.Matrix // filled during the forward
}

// NewDropout returns a new Dropout Function.
func NewDropout[O mat.Tensor](x O, p float64, randGen *rand.LockedRand) *Dropout[O] {
    return &Dropout[O]{
        x:       x,
        prob:    p,
        q:       1.0 - float64(p),
        randGen: randGen,
        mask:    nil,
    }
}

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

// Forward computes the output of the function.
func (r *Dropout[O]) Forward() (mat.Tensor, error) {
    xv := r.x.Value()
    if r.q > 0.0 {
        // FIXME: avoid casting to specific type
        r.mask = bernulli.Distribution[float64](xv.Shape()[0], xv.Shape()[1], r.prob, r.randGen)
        r.mask.ProdScalarInPlace(1.0 / r.q)
    } else {
        r.mask = xv.(mat.Matrix).ZerosLike()
    }
    return xv.(mat.Matrix).Prod(r.mask), nil
}

// Backward computes the backward pass.
func (r *Dropout[O]) Backward(gy mat.Tensor) error {
    if !mat.SameDims(r.x.Value().(mat.Matrix), gy.(mat.Matrix)) {
        return fmt.Errorf("fn: matrices have incompatible dimensions")
    }
    if r.x.RequiresGrad() {
        gx := gy.(mat.Matrix).Prod(r.mask)
        r.x.AccGrad(gx)
    }
    return nil
}