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"
)

// Softmax is a single-input softmax function.
type Softmax[O mat.Tensor] struct {
    x O
    y mat.Matrix // initialized during the forward pass (required by the backward pass)
}

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

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

// Forward computes the output of this function.
func (r *Softmax[O]) Forward() (mat.Tensor, error) {
    r.y = r.x.Value().(mat.Matrix).Softmax()
    return r.y, nil
}

// Backward computes the backward pass.
func (r *Softmax[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() {
        y := r.y
        n := y.Size()
        jb := y.NewMatrix(mat.WithShape(n, n), mat.WithBacking(mat.InitializeMatrix(n, n, func(row, col int) float64 {
            vRow := y.ScalarAt(row).F64()
            if row == col {
                return vRow * (1 - vRow)
            }
            vCol := y.ScalarAt(col).F64()
            return -(vRow * vCol)
        })))
        gx := jb.Mul(gy.(mat.Matrix))
        r.x.AccGrad(gx)
    }
    return nil
}