dense_softmax_test.go

package tensor

import (
	"fmt"
	"testing"

	"github.com/stretchr/testify/assert"
)

func TestSoftMax(t *testing.T) {
	testCases := []struct {
		fn             func(x Tensor, axis int, opts ...FuncOpt) (Tensor, error)
		x              Tensor
		axis           int
		expectedOutput interface{}
	}{
		{
			fn: LogSoftMax,
			x: New(
				Of(Float64),
				WithShape(3, 4),
				WithBacking([]float64{0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1}),
			),
			axis:           -1,
			expectedOutput: []float64{-1.5425355294551628, -1.4425355294551627, -1.3425355294551626, -1.2425355294551628, -1.5425355294551628, -1.4425355294551627, -1.3425355294551626, -1.2425355294551628, -1.5425355294551628, -1.4425355294551627, -1.3425355294551629, -1.2425355294551628},
		},
		{
			fn: LogSoftMax,
			x: New(
				Of(Float32),
				WithShape(3, 4),
				WithBacking([]float32{0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1}),
			),
			axis:           -1,
			expectedOutput: []float32{-1.5425355294551628, -1.4425355294551627, -1.3425355294551626, -1.2425355294551628, -1.5425355294551628, -1.4425355294551627, -1.3425355294551626, -1.2425355294551628, -1.5425355294551628, -1.4425355294551627, -1.3425355294551629, -1.2425355294551628},
		},
		{
			fn: LogSoftMax,
			x: New(
				Of(Float32),
				WithShape(3, 2, 2),
				WithBacking([]float32{0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1}),
			),
			axis:           -1,
			expectedOutput: []float32{-0.7443967, -0.64439666, -0.7443967, -0.64439666, -0.7443967, -0.64439666, -0.7443966, -0.64439666, -0.7443966, -0.64439666, -0.7443967, -0.64439666},
		},
		{
			fn: LogSoftMax,
			x: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float64{0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1}),
			),
			axis:           1,
			expectedOutput: []float64{-0.7981388693815918, -0.7981388693815918, -0.5981388693815918, -0.5981388693815919, -0.7981388693815918, -0.7981388693815918, -0.5981388693815919, -0.5981388693815919, -0.7981388693815918, -0.7981388693815918, -0.5981388693815919, -0.5981388693815918},
		},
		{
			fn: SoftMax,
			x: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float64{0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1}),
			),
			axis:           1,
			expectedOutput: []float64{0.4501660026875221, 0.45016600268752205, 0.549833997312478, 0.5498339973124778, 0.45016600268752205, 0.45016600268752205, 0.5498339973124778, 0.5498339973124778, 0.45016600268752205, 0.4501660026875221, 0.5498339973124778, 0.549833997312478},
		},
		{
			fn: SoftMax,
			x: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float64{0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1}),
			),
			axis:           -1,
			expectedOutput: []float64{0.47502081252106, 0.52497918747894, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.52497918747894},
		},
		{
			fn: SoftMax,
			x: New(
				Of(Float32),
				WithShape(3, 4),
				WithBacking([]float32{0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1}),
			),
			axis:           -1,
			expectedOutput: []float32{0.21383822, 0.23632777, 0.2611826, 0.2886514, 0.21383823, 0.23632778, 0.2611826, 0.2886514, 0.21383822, 0.23632777, 0.26118258, 0.2886514},
		},
		{
			fn: SoftMax,
			x: New(
				Of(Float64),
				WithShape(3, 4),
				WithBacking([]float64{0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1}),
			),
			axis:           -1,
			expectedOutput: []float64{0.21383822, 0.23632777, 0.2611826, 0.2886514, 0.21383823, 0.23632778, 0.2611826, 0.2886514, 0.21383822, 0.23632777, 0.26118258, 0.2886514},
		},
	}
	for i, tC := range testCases {
		t.Run(fmt.Sprintf("Example #%d - %v %v", i+1, tC.x.Shape(), tC.x.Dtype()), func(t *testing.T) {
			c := assert.New(t)

			output, err := tC.fn(tC.x, tC.axis)
			t.Logf("output: %#v", output.Data())

			c.NoError(err)
			c.NotNil(output)

			c.Equal(tC.x.Shape(), output.Shape())
			c.InDeltaSlice(tC.expectedOutput, output.Data(), 1e-6)
		})
	}
}

func TestSoftMaxB(t *testing.T) {
	testCases := []struct {
		fn             func(output, grad Tensor, axis int, opts ...FuncOpt) (Tensor, error)
		output         Tensor
		grad           Tensor
		axis           int
		expectedOutput interface{}
	}{
		{
			fn: SoftMaxB,
			output: New(
				Of(Float64),
				WithShape(3, 4),
				WithBacking([]float64{0.21383822, 0.23632777, 0.2611826, 0.2886514, 0.21383823, 0.23632778, 0.2611826, 0.2886514, 0.21383822, 0.23632777, 0.26118258, 0.2886514}),
			),
			grad: New(
				Of(Float64),
				WithShape(3, 4),
				WithBacking([]float64{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           -1,
			expectedOutput: []float64{-0.003474116568224552, -0.0014762147035963322, 0.0009803563066858392, 0.00396997522759976, -0.003474116880376028, -0.001476214931490494, 0.0009803561238580223, 0.003969975025543781, -0.0034741159267098936, -0.0014762139946130218, 0.0009803570151630109, 0.003969976093553957},
		},
		{
			fn: LogSoftMaxB,
			output: New(
				Of(Float64),
				WithShape(3, 4),
				WithBacking([]float64{-1.5425355294551628, -1.4425355294551627, -1.3425355294551626, -1.2425355294551628, -1.5425355294551628, -1.4425355294551627, -1.3425355294551626, -1.2425355294551628, -1.5425355294551628, -1.4425355294551627, -1.3425355294551629, -1.2425355294551628}),
			),
			grad: New(
				Of(Float64),
				WithShape(3, 4),
				WithBacking([]float64{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           -1,
			expectedOutput: []float64{-0.011383822036598441, -0.003632778232153768, 0.0038817407844924366, 0.01113485948425977, -0.005597937295155945, -0.001445223403599799, 0.0020925260396803457, 0.004950634659075405, 0.00018794744628654992, 0.0007423314249541871, 0.00030331129486827163, -0.0012335901661089598},
		},
		{
			fn: SoftMaxB,
			output: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float64{0.47502081252106, 0.52497918747894, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.52497918747894}),
			),
			grad: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float64{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           -1,
			expectedOutput: []float64{-0.002493760401928919, 0.0024937604019289205, -0.0024937604019289183, 0.002493760401928922, -0.002493760401928915, 0.002493760401928922, -0.002493760401928912, 0.0024937604019289253, -0.0024937604019289183, 0.0024937604019289253, -0.0024937604019289183, 0.0024937604019289183},
		},
		{
			fn: SoftMaxB,
			output: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float64{0.4501660026875221, 0.45016600268752205, 0.549833997312478, 0.5498339973124778, 0.45016600268752205, 0.45016600268752205, 0.5498339973124778, 0.5498339973124778, 0.45016600268752205, 0.4501660026875221, 0.5498339973124778, 0.549833997312478}),
			),
			grad: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float64{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           1,
			expectedOutput: []float64{-0.004950331454237199, -0.004950331454237198, 0.004950331454237199, 0.0049503314542372, -0.004950331454237196, -0.004950331454237193, 0.004950331454237203, 0.0049503314542372065, -0.004950331454237193, -0.0049503314542372, 0.0049503314542372065, 0.004950331454237193},
		},
		{
			fn: LogSoftMaxB,
			output: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float64{-0.7981388693815918, -0.7981388693815918, -0.5981388693815918, -0.5981388693815919, -0.7981388693815918, -0.7981388693815918, -0.5981388693815919, -0.5981388693815919, -0.7981388693815918, -0.7981388693815918, -0.5981388693815919, -0.5981388693815918}),
			),
			grad: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float64{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           1,
			expectedOutput: []float64{-0.008006640107500884, -0.007009960161251325, 0.00800664010750088, 0.007009960161251332, -0.004019920322502654, -0.003023240376253103, 0.004019920322502661, 0.0030232403762530968, -3.32005375044292e-05, 0.0009634794087451421, 3.320053750442642e-05, -0.0009634794087451543},
		},
		{
			fn: LogSoftMaxB,
			output: New(
				Of(Float32),
				WithShape(3, 2, 2),
				WithBacking([]float32{-0.7981388693815918, -0.7981388693815918, -0.5981388693815918, -0.5981388693815919, -0.7981388693815918, -0.7981388693815918, -0.5981388693815919, -0.5981388693815919, -0.7981388693815918, -0.7981388693815918, -0.5981388693815919, -0.5981388693815918}),
			),
			grad: New(
				Of(Float32),
				WithShape(3, 2, 2),
				WithBacking([]float32{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           1,
			expectedOutput: []float64{-0.008006640107500884, -0.007009960161251325, 0.00800664010750088, 0.007009960161251332, -0.004019920322502654, -0.003023240376253103, 0.004019920322502661, 0.0030232403762530968, -3.32005375044292e-05, 0.0009634794087451421, 3.320053750442642e-05, -0.0009634794087451543},
		},
		{
			fn: SoftMaxB,
			output: New(
				Of(Float32),
				WithShape(3, 2, 2),
				WithBacking([]float32{0.4501660026875221, 0.45016600268752205, 0.549833997312478, 0.5498339973124778, 0.45016600268752205, 0.45016600268752205, 0.5498339973124778, 0.5498339973124778, 0.45016600268752205, 0.4501660026875221, 0.5498339973124778, 0.549833997312478}),
			),
			grad: New(
				Of(Float32),
				WithShape(3, 2, 2),
				WithBacking([]float32{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           1,
			expectedOutput: []float32{-0.004950331454237199, -0.004950331454237198, 0.004950331454237199, 0.0049503314542372, -0.004950331454237196, -0.004950331454237193, 0.004950331454237203, 0.0049503314542372065, -0.004950331454237193, -0.0049503314542372, 0.0049503314542372065, 0.004950331454237193},
		},
		{
			fn: SoftMaxB,
			output: New(
				Of(Float32),
				WithShape(3, 4),
				WithBacking([]float32{0.21383822, 0.23632777, 0.2611826, 0.2886514, 0.21383823, 0.23632778, 0.2611826, 0.2886514, 0.21383822, 0.23632777, 0.26118258, 0.2886514}),
			),
			grad: New(
				Of(Float64),
				WithShape(3, 4),
				WithBacking([]float32{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           -1,
			expectedOutput: []float32{-0.003474116568224552, -0.0014762147035963322, 0.0009803563066858392, 0.00396997522759976, -0.003474116880376028, -0.001476214931490494, 0.0009803561238580223, 0.003969975025543781, -0.0034741159267098936, -0.0014762139946130218, 0.0009803570151630109, 0.003969976093553957},
		},
		{
			fn: LogSoftMaxB,
			output: New(
				Of(Float64),
				WithShape(3, 4),
				WithBacking([]float32{-1.5425355294551628, -1.4425355294551627, -1.3425355294551626, -1.2425355294551628, -1.5425355294551628, -1.4425355294551627, -1.3425355294551626, -1.2425355294551628, -1.5425355294551628, -1.4425355294551627, -1.3425355294551629, -1.2425355294551628}),
			),
			grad: New(
				Of(Float64),
				WithShape(3, 4),
				WithBacking([]float32{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           -1,
			expectedOutput: []float32{-0.011383822036598441, -0.003632778232153768, 0.0038817407844924366, 0.01113485948425977, -0.005597937295155945, -0.001445223403599799, 0.0020925260396803457, 0.004950634659075405, 0.00018794744628654992, 0.0007423314249541871, 0.00030331129486827163, -0.0012335901661089598},
		},
		{
			fn: SoftMaxB,
			output: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float32{0.47502081252106, 0.52497918747894, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.5249791874789399, 0.47502081252106, 0.52497918747894}),
			),
			grad: New(
				Of(Float64),
				WithShape(3, 2, 2),
				WithBacking([]float32{0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.10, 0.11, 0.12}),
			),
			axis:           -1,
			expectedOutput: []float32{-0.002493760401928919, 0.0024937604019289205, -0.0024937604019289183, 0.002493760401928922, -0.002493760401928915, 0.002493760401928922, -0.002493760401928912, 0.0024937604019289253, -0.0024937604019289183, 0.0024937604019289253, -0.0024937604019289183, 0.0024937604019289183},
		},
	}
	for i, tC := range testCases {
		t.Run(fmt.Sprintf("Example #%d - %v %v", i+1, tC.output.Shape(), tC.output.Dtype()), func(t *testing.T) {
			c := assert.New(t)

			dx, err := tC.fn(tC.output, tC.grad, tC.axis)
			t.Logf("output: %#v", tC.output.Data())

			c.NoError(err)
			c.NotNil(dx)

			c.Equal(tC.output.Shape(), dx.Shape())
			c.InDeltaSlice(tC.expectedOutput, dx.Data(), 1e-6)
		})
	}
}