mat/gradfn/max.go
// Copyright 2020 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"
)
// Max is an operator to perform element-wise max.
// y = max(x1, x2)
type Max[O mat.Tensor] struct {
x1 O
x2 O
}
// NewMax returns a new Max Function.
func NewMax[O mat.Tensor](x1 O, x2 O) *Max[O] {
return &Max[O]{
x1: x1,
x2: x2,
}
}
// Operands returns the list of operands.
func (r *Max[O]) Operands() []mat.Tensor {
return []mat.Tensor{r.x1, r.x2}
}
// Forward computes the output of the function.
func (r *Max[O]) Forward() (mat.Tensor, error) {
return r.x1.Value().(mat.Matrix).Maximum(r.x2.Value().(mat.Matrix)), nil
}
// Backward computes the backward pass.
func (r *Max[O]) Backward(gy mat.Tensor) error {
x1v := r.x1.Value().(mat.Matrix)
x2v := r.x2.Value().(mat.Matrix)
if !mat.SameDims(x1v, gy) || !mat.SameDims(x2v, gy) {
return fmt.Errorf("fn: matrices have incompatible dimensions")
}
n := gy.Size()
// FIXME: avoid casting to specific type
gyData := mat.Data[float64](gy.(mat.Matrix))
x1vData := mat.Data[float64](x1v)
x2vData := mat.Data[float64](x2v)
if r.x1.RequiresGrad() {
gxData := make([]float64, n)
for i := 0; i < n; i++ {
if x1vData[i] > x2vData[i] {
gxData[i] = gyData[i]
}
}
gx := x1v.NewMatrix(mat.WithBacking(gxData))
r.x1.AccGrad(gx)
}
if r.x2.RequiresGrad() {
gxData := make([]float64, n)
for i := 0; i < n; i++ {
if x2vData[i] > x1vData[i] {
gxData[i] = gyData[i]
}
}
gx := x1v.NewMatrix(mat.WithBacking(gxData))
r.x2.AccGrad(gx)
}
return nil
}