WIP(example/translation): more update

example/translation/lang.go

@@ -88,5 +88,5 @@ func (l *Lang) SeqToString(seq []int) (retVal string) {
 		words = append(words, w)
 	}
 
-	return strings.Join(words, "")
+	return strings.Join(words, " ")
 }

example/translation/main.go

@@ -45,7 +45,7 @@ func newEncoder(vs nn.Path, inDim, hiddenDim int64) (retVal Encoder) {
 
 func (e Encoder) forward(xs ts.Tensor, state nn.GRUState) (retTs ts.Tensor, retState nn.GRUState) {
 
-	retTs = e.embedding.Forward(xs).MustView([]int64{1, -1}, false)
+	retTs = e.embedding.Forward(xs).MustView([]int64{1, -1}, true)
 	retState = e.gru.Step(retTs, state).(nn.GRUState)
 
 	return retTs, retState
@@ -83,6 +83,7 @@ func (d Decoder) forward(xs ts.Tensor, state nn.GRUState, encOutputs ts.Tensor,
 	state0 := stateTs.Idx([]ts.TensorIndexer{ts.NewSelect(0)})
 
 	catTs := ts.MustCat([]ts.Tensor{forwardTs, state0}, 1)
+	state0.MustDrop()
 
 	// NOTE. d.attn Ws shape : [512, 10]
 	appliedTs := catTs.Apply(d.attn)
@@ -104,15 +105,22 @@ func (d Decoder) forward(xs ts.Tensor, state nn.GRUState, encOutputs ts.Tensor,
 		shape := []int64{sz1, MaxLength - sz2, sz3}
 		zerosTs := ts.MustZeros(shape, gotch.Float, d.device)
 		encOutputsTs = ts.MustCat([]ts.Tensor{encOutputs, zerosTs}, 1)
+		zerosTs.MustDrop()
 	}
 
 	attnApplied := attnWeights.MustBmm(encOutputsTs, true).MustSqueeze1(1, true)
-	fmt.Printf("attnApplied shape: %v\n", attnApplied.MustSize())
-	fmt.Printf("xs shape: %v\n", forwardTs.MustSize())
+	attnWeights.MustDrop()
+	encOutputsTs.MustDrop()
 
-	xsTs := ts.MustCat([]ts.Tensor{forwardTs, attnApplied}, 1).Apply(d.attnCombine).MustRelu(true)
+	cTs := ts.MustCat([]ts.Tensor{forwardTs, attnApplied}, 1)
+	forwardTs.MustDrop()
+	attnApplied.MustDrop()
+	aTs := cTs.Apply(d.attnCombine)
+	cTs.MustDrop()
+	xsTs := aTs.MustRelu(true)
 
 	retState = d.gru.Step(xsTs, state).(nn.GRUState)
+	xsTs.MustDrop()
 
 	retTs = d.linear.Forward(retState.Value()).MustLogSoftmax(-1, gotch.Float, true)
 
@@ -139,14 +147,13 @@ func newModel(vs nn.Path, ilang Lang, olang Lang, hiddenDim int64) (retVal Model
 
 func (m *Model) trainLoss(input []int, target []int) (retVal ts.Tensor) {
 	state := m.encoder.gru.ZeroState(1)
-	fmt.Printf("state shape: %v\n", state.(nn.GRUState).Value().MustSize())
 	var encOutputs []ts.Tensor
 
 	for _, v := range input {
 		s := ts.MustOfSlice([]int64{int64(v)}).MustTo(m.device, true)
 
 		outTs, outState := m.encoder.forward(s, state.(nn.GRUState))
-
+		s.MustDrop()
 		encOutputs = append(encOutputs, outTs)
 		state.(nn.GRUState).Tensor.MustDrop()
 		state = outState
@@ -168,17 +175,13 @@ func (m *Model) trainLoss(input []int, target []int) (retVal ts.Tensor) {
 		state = outState
 
 		targetTs := ts.MustOfSlice([]int64{int64(s)}).MustTo(m.device, true)
-		fmt.Printf("targetTs shape: %v\n", targetTs.MustSize())
-		fmt.Printf("outTs shape: %v\n", outTs.MustSize())
-
-		// TODO: fix the error: input tensor should be 1D or 2D at /pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:35
-		ws := ts.NewTensor()
-		currLoss := outTs.MustNllLoss(targetTs, ws, int64(1), -100, false)
+		currLoss := outTs.MustView([]int64{1, -1}, false).MustNllLoss(targetTs, ts.NewTensor(), int64(1), -100, false)
 
 		loss.MustAdd_(currLoss)
 		currLoss.MustDrop()
 
-		_, output := outTs.MustTopK(1, -1, true, true)
+		noUseTs, output := outTs.MustTopK(1, -1, true, true)
+		noUseTs.MustDrop()
 
 		if m.decoderEos == outTs.Int64Values()[0] {
 			break
@@ -186,8 +189,13 @@ func (m *Model) trainLoss(input []int, target []int) (retVal ts.Tensor) {
 
 		prev.MustDrop()
 		prev = output
+		outTs.MustDrop()
 	}
 
+	state.(nn.GRUState).Tensor.MustDrop()
+	stackTs.MustDrop()
+	prev.MustDrop()
+
 	return loss
 
 }
@@ -293,12 +301,12 @@ func main() {
 		input := pair.Val1
 		target := pair.Val2
 		loss := model.trainLoss(input, target)
-		panic("reached")
 		opt.BackwardStep(loss)
 		lossStats.update(loss.Float64Values()[0] / float64(len(target)))
+		loss.MustDrop()
 
 		if i%1000 == 0 {
-			fmt.Printf("%v %v\n", i, lossStats.avgAndReset())
+			fmt.Printf("Trained %v samples -  Avg. Loss: %v\n", i, lossStats.avgAndReset())
 			for predIdx := 1; predIdx <= 5; predIdx++ {
 				idx := rand.Intn(len(pairs))
 				in := pairs[idx].Val1

nn/rnn.go

@@ -1,8 +1,6 @@
 package nn
 
 import (
-	"fmt"
-
 	"github.com/sugarme/gotch"
 	ts "github.com/sugarme/gotch/tensor"
 )
@@ -228,6 +226,12 @@ func NewGRU(vs Path, inDim, hiddenDim int64, cfg RNNConfig) (retVal GRU) {
 		}
 	}
 
+	if vs.Device().IsCuda() {
+		// NOTE. 3 is for GRU
+		// ref. rnn.cpp in Pytorch
+		ts.Must_CudnnRnnFlattenWeight(flatWeights, 4, inDim, 3, hiddenDim, cfg.NumLayers, cfg.BatchFirst, cfg.Bidirectional)
+	}
+
 	return GRU{
 		flatWeights: flatWeights,
 		hiddenDim:   hiddenDim,
@@ -256,14 +260,12 @@ func (g GRU) ZeroState(batchDim int64) (retVal State) {
 func (g GRU) Step(input ts.Tensor, inState State) (retVal State) {
 	unsqueezedInput := input.MustUnsqueeze(1, false)
 
-	fmt.Printf("unsqueezed input shape: %v\n", unsqueezedInput.MustSize())
-	fmt.Printf("inState Ts size: %v\n", inState.(GRUState).Tensor.MustSize())
-
 	output, state := g.SeqInit(unsqueezedInput, inState)
 
 	// NOTE: though we won't use `output`, it is a Ctensor created in C land, so
 	// it should be cleaned up here to prevent memory hold-up.
 	output.MustDrop()
+	unsqueezedInput.MustDrop()
 
 	return state
 }