Update functional option pattern

autograd_test.go

@@ -374,7 +374,7 @@ func Example_double() {
 
 func Example_linearRegression() {
 	// p318
-	s :=rand.Const()
+	s := rand.Const()
 	xrand := matrix.Rand(100, 1, s)
 	yrand := matrix.Rand(100, 1, s)
 
@@ -424,10 +424,11 @@ func Example_linearRegression() {
 
 func Example_mlp() {
 	s := rand.Const()
-	m := model.NewMLP([]int{10, 1}, model.MLPOpts{
-		Activation: F.ReLU,
-		Source:     s,
-	})
+	m := model.NewMLP([]int{10, 1},
+		model.MLPWithSource(s),
+		model.MLPWithActivation(F.ReLU),
+	)
+
 	o := optimizer.SGD{
 		LearningRate: 0.2,
 	}

layer/linear.go

@@ -9,36 +9,45 @@ import (
 	"github.com/itsubaki/autograd/variable"
 )
 
-type LinearOpts struct {
-	InSize int
-	NoBias bool
-	Source randv2.Source
-}
+type OptionFunc func(*LinearT)
 
-func Linear(outSize int, opts ...LinearOpts) *LinearT {
-	var s randv2.Source
-	if len(opts) != 0 && opts[0].Source != nil {
-		s = opts[0].Source
+func WithSource(s randv2.Source) OptionFunc {
+	return func(l *LinearT) {
+		l.s = s
 	}
+}
 
-	p := make(Parameters)
-	if len(opts) == 0 || !opts[0].NoBias {
-		p.Add("b", variable.Zero(1, outSize))
+func WithInSize(inSize int) OptionFunc {
+	return func(l *LinearT) {
+		l.Parameters.Add("w", initw(inSize, l.outSize, l.s))
 	}
-	if len(opts) != 0 && opts[0].InSize > 0 {
-		p.Add("w", initw(opts[0].InSize, outSize, s))
+}
+
+func WithNoBias() OptionFunc {
+	return func(l *LinearT) {
+		l.Parameters.Delete("b")
 	}
+}
 
-	return &LinearT{
+func Linear(outSize int, opts ...OptionFunc) *LinearT {
+	p := make(Parameters)
+	p.Add("b", variable.Zero(1, outSize))
+
+	l := &LinearT{
 		outSize:    outSize,
-		source:     s,
 		Parameters: p,
 	}
+
+	for _, opt := range opts {
+		opt(l)
+	}
+
+	return l
 }
 
 type LinearT struct {
 	outSize int
-	source  randv2.Source
+	s       randv2.Source
 	Parameters
 }
 
@@ -49,7 +58,7 @@ func (l *LinearT) First(x ...*variable.Variable) *variable.Variable {
 func (l *LinearT) Forward(x ...*variable.Variable) []*variable.Variable {
 	if _, ok := l.Parameters["w"]; !ok {
 		inSize := variable.Shape(x[0])[1]
-		l.Parameters.Add("w", initw(inSize, l.outSize, l.source))
+		l.Parameters.Add("w", initw(inSize, l.outSize, l.s))
 	}
 
 	return []*variable.Variable{

layer/linear_test.go

@@ -9,9 +9,7 @@ import (
 )
 
 func ExampleLinear() {
-	l := L.Linear(5, L.LinearOpts{
-		Source: rand.Const(),
-	})
+	l := L.Linear(5, L.WithSource(rand.Const()))
 
 	x := variable.New(1, 2, 3)
 	y := l.Forward(x)
@@ -28,10 +26,10 @@ func ExampleLinear() {
 }
 
 func ExampleLinear_inSize() {
-	l := L.Linear(5, L.LinearOpts{
-		InSize: 3,
-		Source: rand.Const(),
-	})
+	l := L.Linear(5,
+		L.WithSource(rand.Const()),
+		L.WithInSize(3),
+	)
 
 	x := variable.New(1, 2, 3)
 	y := l.Forward(x)
@@ -48,9 +46,7 @@ func ExampleLinear_inSize() {
 }
 
 func ExampleLinear_nobias() {
-	l := L.Linear(5, L.LinearOpts{
-		NoBias: true,
-	})
+	l := L.Linear(5, L.WithNoBias())
 
 	x := variable.New(1, 2, 3)
 	l.Forward(x)

layer/lstm.go

@@ -7,33 +7,38 @@ import (
 	"github.com/itsubaki/autograd/variable"
 )
 
-type LSTMOpts struct {
-	Source randv2.Source
-}
+type LSTMOptionFunc func(*LSTMT)
 
-func LSTM(hiddenSize int, opts ...LSTMOpts) *LSTMT {
-	var s randv2.Source
-	if len(opts) != 0 && opts[0].Source != nil {
-		s = opts[0].Source
+func LSTMWithSource(s randv2.Source) LSTMOptionFunc {
+	return func(l *LSTMT) {
+		l.s = s
 	}
+}
 
-	l := make(Layers)
-	l.Add("x2f", Linear(hiddenSize, LinearOpts{Source: s}))
-	l.Add("x2i", Linear(hiddenSize, LinearOpts{Source: s}))
-	l.Add("x2o", Linear(hiddenSize, LinearOpts{Source: s}))
-	l.Add("x2u", Linear(hiddenSize, LinearOpts{Source: s}))
-	l.Add("h2f", Linear(hiddenSize, LinearOpts{Source: s, InSize: hiddenSize, NoBias: true}))
-	l.Add("h2i", Linear(hiddenSize, LinearOpts{Source: s, InSize: hiddenSize, NoBias: true}))
-	l.Add("h2o", Linear(hiddenSize, LinearOpts{Source: s, InSize: hiddenSize, NoBias: true}))
-	l.Add("h2u", Linear(hiddenSize, LinearOpts{Source: s, InSize: hiddenSize, NoBias: true}))
+func LSTM(hiddenSize int, opts ...LSTMOptionFunc) *LSTMT {
+	lstm := &LSTMT{
+		Layers: make(Layers),
+	}
 
-	return &LSTMT{
-		Layers: l,
+	for _, opt := range opts {
+		opt(lstm)
 	}
+
+	lstm.Layers.Add("x2f", Linear(hiddenSize, WithSource(lstm.s)))
+	lstm.Layers.Add("x2i", Linear(hiddenSize, WithSource(lstm.s)))
+	lstm.Layers.Add("x2o", Linear(hiddenSize, WithSource(lstm.s)))
+	lstm.Layers.Add("x2u", Linear(hiddenSize, WithSource(lstm.s)))
+	lstm.Layers.Add("h2f", Linear(hiddenSize, WithSource(lstm.s), WithInSize(hiddenSize), WithNoBias()))
+	lstm.Layers.Add("h2i", Linear(hiddenSize, WithSource(lstm.s), WithInSize(hiddenSize), WithNoBias()))
+	lstm.Layers.Add("h2o", Linear(hiddenSize, WithSource(lstm.s), WithInSize(hiddenSize), WithNoBias()))
+	lstm.Layers.Add("h2u", Linear(hiddenSize, WithSource(lstm.s), WithInSize(hiddenSize), WithNoBias()))
+
+	return lstm
 }
 
 type LSTMT struct {
 	h, c *variable.Variable
+	s    randv2.Source
 	Layers
 }
 

layer/lstm_test.go

@@ -9,9 +9,7 @@ import (
 )
 
 func ExampleLSTM() {
-	l := L.LSTM(2, L.LSTMOpts{
-		Source: rand.Const(),
-	})
+	l := L.LSTM(2, L.LSTMWithSource(rand.Const()))
 
 	x := variable.New(1)
 	y := l.Forward(x)
@@ -39,9 +37,7 @@ func ExampleLSTM() {
 }
 
 func ExampleLSTM_backward() {
-	l := L.LSTM(2, L.LSTMOpts{
-		Source: rand.Const(),
-	})
+	l := L.LSTM(2, L.LSTMWithSource(rand.Const()))
 
 	x := variable.New(1)
 	y := l.First(x)

layer/parameter.go

@@ -11,6 +11,10 @@ func (p Parameters) Add(name string, param Parameter) {
 	p[name] = param
 }
 
+func (p Parameters) Delete(name string) {
+	delete(p, name)
+}
+
 func (p Parameters) Params() Parameters {
 	return p
 }

layer/rnn.go

@@ -7,27 +7,31 @@ import (
 	"github.com/itsubaki/autograd/variable"
 )
 
-type RNNOpts struct {
-	Source randv2.Source
+type RNNOptionFunc func(*RNNT)
+
+func RNNWithSource(s randv2.Source) RNNOptionFunc {
+	return func(l *RNNT) {
+		l.s = s
+	}
 }
+func RNN(hiddenSize int, opts ...RNNOptionFunc) *RNNT {
+	rnn := &RNNT{
+		Layers: make(Layers),
+	}
 
-func RNN(hiddenSize int, opts ...RNNOpts) *RNNT {
-	var s randv2.Source
-	if len(opts) != 0 && opts[0].Source != nil {
-		s = opts[0].Source
+	for _, opt := range opts {
+		opt(rnn)
 	}
 
-	l := make(Layers)
-	l.Add("x2h", Linear(hiddenSize, LinearOpts{Source: s}))
-	l.Add("h2h", Linear(hiddenSize, LinearOpts{Source: s, InSize: hiddenSize, NoBias: true}))
+	rnn.Layers.Add("x2h", Linear(hiddenSize, WithSource(rnn.s)))
+	rnn.Layers.Add("h2h", Linear(hiddenSize, WithSource(rnn.s), WithInSize(hiddenSize), WithNoBias()))
 
-	return &RNNT{
-		Layers: l,
-	}
+	return rnn
 }
 
 type RNNT struct {
 	h *variable.Variable
+	s randv2.Source
 	Layers
 }
 

layer/rnn_test.go

@@ -9,9 +9,7 @@ import (
 )
 
 func ExampleRNN() {
-	l := L.RNN(2, L.RNNOpts{
-		Source: rand.Const(),
-	})
+	l := L.RNN(2, L.RNNWithSource(rand.Const()))
 
 	x := variable.New(1)
 	y := l.Forward(x)
@@ -29,9 +27,7 @@ func ExampleRNN() {
 }
 
 func ExampleRNN_backward() {
-	l := L.RNN(2, L.RNNOpts{
-		Source: rand.Const(),
-	})
+	l := L.RNN(2, L.RNNWithSource(rand.Const()))
 
 	x := variable.New(1)
 	y := l.First(x)

model/lstm.go

@@ -7,30 +7,36 @@ import (
 	"github.com/itsubaki/autograd/variable"
 )
 
-type LSTMOpts struct {
-	Source randv2.Source
+type LSTMOptionFunc func(*LSTM)
+
+func LSTMWithSource(s randv2.Source) LSTMOptionFunc {
+	return func(l *LSTM) {
+		l.s = s
+	}
 }
 
 type LSTM struct {
+	s randv2.Source
 	Model
 }
 
-func NewLSTM(hiddenSize, outSize int, opts ...LSTMOpts) *LSTM {
-	var s randv2.Source
-	if len(opts) > 0 && opts[0].Source != nil {
-		s = opts[0].Source
+func NewLSTM(hiddenSize, outSize int, opts ...LSTMOptionFunc) *LSTM {
+	lstm := &LSTM{
+		Model: Model{
+			Layers: make([]L.Layer, 0),
+		},
 	}
 
-	layers := []L.Layer{
-		L.LSTM(hiddenSize, L.LSTMOpts{Source: s}),
-		L.Linear(outSize, L.LinearOpts{Source: s}),
+	for _, opt := range opts {
+		opt(lstm)
 	}
 
-	return &LSTM{
-		Model: Model{
-			Layers: layers,
-		},
-	}
+	lstm.Layers = append(lstm.Layers, []L.Layer{
+		L.LSTM(hiddenSize, L.LSTMWithSource(lstm.s)),
+		L.Linear(outSize, L.WithSource(lstm.s)),
+	}...)
+
+	return lstm
 }
 
 func (m *LSTM) ResetState() {

model/lstm_test.go

@@ -21,9 +21,7 @@ func ExampleLSTM() {
 }
 
 func ExampleLSTM_backward() {
-	m := model.NewLSTM(1, 1, model.LSTMOpts{
-		Source: rand.Const(),
-	})
+	m := model.NewLSTM(1, 1, model.LSTMWithSource(rand.Const()))
 
 	x := variable.New(1, 2)
 	y := m.Forward(x)