Remove Seq2D type. Use Seq instead.

Author: Atry

array2D/src/main/scala/com/thoughtworks/deeplearning/array2D/layers/ToSeq.scala

@@ -7,7 +7,6 @@ import com.thoughtworks.deeplearning.Batch._
 import com.thoughtworks.deeplearning.Layer._
 import com.thoughtworks.deeplearning.array2D.utilities._
 import com.thoughtworks.deeplearning.{Batch, BatchId, BufferedLayer, Layer}
-import com.thoughtworks.deeplearning.seq2D.utilities.{Seq2D, Seq2DBatch}
 import org.nd4j.linalg.api.ndarray.INDArray
 import org.nd4j.linalg.factory.Nd4j
 import org.nd4j.linalg.ops.transforms.Transforms
@@ -19,8 +18,10 @@ final case class ToSeq[Input0 <: Batch](operand: Layer.Aux[Input0, Array2D#Batch
   override type Input = Input0
 
   final class BufferedBatch private[ToSeq] (override val input: BatchId.Aux[Input], upstream: Array2D#Batch)
-      extends ReferenceCount
-      with Seq2DBatch {
+      extends ReferenceCount {
+
+    override type Data = scala.Seq[scala.Seq[Eval[scala.Double]]]
+    override type Delta = (scala.Int, (scala.Int, Eval[scala.Double]))
 
     private def zeroDelta =
       upstream.value.map { upstreamData =>
@@ -44,19 +45,20 @@ final case class ToSeq[Input0 <: Batch](operand: Layer.Aux[Input0, Array2D#Batch
 
     override def backward(delta: Delta): Unit = {
       synchronized {
-        val (i, j, value) = delta.value
-        upstreamDelta.value(i, j) = upstreamDelta
-            .value(i, j) + value // Cannot use += because of https://issues.scala-lang.org/browse/SI-10021
+        val (i, (j, value)) = delta
+        // Cannot use += because of https://issues.scala-lang.org/browse/SI-10021
+        upstreamDelta.value(i, j) = upstreamDelta.value(i, j) + value.value
       }
     }
 
     override val value: Data = {
-      upstream.value.map { ndarray: INDArray =>
-        val doubleArray = ndarray.data.asDouble()
-        for (i <- (0 until ndarray.rows).view) yield {
-          doubleArray.view(i * ndarray.columns, (i + 1) * ndarray.columns)
+      val ndarray = upstream.value.value
+      val doubleArray = ndarray.data.asDouble()
+      for (i <- (0 until ndarray.rows).view) yield {
+        doubleArray.view(i * ndarray.columns, (i + 1) * ndarray.columns).map { doubleValue =>
+          Eval.now(doubleValue)
         }
-      }.memoize
+      }
     }
   }
 

array2D/src/main/scala/com/thoughtworks/deeplearning/array2D/package.scala

@@ -5,7 +5,6 @@ import com.thoughtworks.deeplearning.dsl._
 import com.thoughtworks.deeplearning.array2D.layers._
 import com.thoughtworks.deeplearning.array2D.optimizers.Optimizer
 import com.thoughtworks.deeplearning.double.utilities.Double
-import com.thoughtworks.deeplearning.seq2D.utilities.Seq2D
 import org.nd4j.linalg.api.ndarray.INDArray
 
 import scala.language.implicitConversions
@@ -16,7 +15,7 @@ import scala.language.implicitConversions
 package object array2D {
 
   /** @template */
-  type Array2D = utilities.Array2D
+  type Array2D = com.thoughtworks.deeplearning.array2D.utilities.Array2D
 
   implicit def `max(Array2D,Double)`[Left, Right, Input <: Batch]
     : max.Case.Aux[Layer.Aux[Input, Array2D#Batch], Layer.Aux[Input, Double#Batch], Layer.Aux[Input, Array2D#Batch]] =
@@ -133,7 +132,9 @@ package object array2D {
       Negative(differentiable)
     }
 
-    def toSeq: Layer.Aux[Input, Seq2D#Batch] = {
+    def toSeq: Layer.Aux[
+      Input,
+      Batch.Aux[scala.Seq[scala.Seq[Eval[scala.Double]]], (scala.Int, (scala.Int, Eval[scala.Double]))]] = {
       ToSeq(differentiable)
     }
 

array2D/src/main/scala/com/thoughtworks/deeplearning/seq2D/layers/Get.scala

@@ -15,7 +15,7 @@ import scala.language.implicitConversions
 package object double {
 
   /** @template */
-  type Double = utilities.Double
+  type Double = com.thoughtworks.deeplearning.double.utilities.Double
 
   implicit def liftNativeDoubleToLayer[InputData, InputDelta](implicit inputType: Type[InputData, InputDelta])
     : ToLayer.Aux[scala.Double, Batch.Aux[InputData, InputDelta], Eval[scala.Double], Eval[scala.Double]] =

array2D/src/main/scala/com/thoughtworks/deeplearning/seq2D/package.scala

@@ -5,8 +5,8 @@ import com.thoughtworks.deeplearning.Layer._
 import com.thoughtworks.deeplearning.Batch._
 import com.thoughtworks.deeplearning.hlist._
 import com.thoughtworks.deeplearning.boolean._
-import com.thoughtworks.deeplearning.seq2D._
 import com.thoughtworks.deeplearning.double._
+import com.thoughtworks.deeplearning.seq._
 import com.thoughtworks.deeplearning.array2D._
 import com.thoughtworks.deeplearning.dsl._
 import com.thoughtworks.deeplearning.dsl.layers.{Identity, Literal}
@@ -26,6 +26,8 @@ import scala.util.Random
   */
 object FortuneTeller {
 
+  type Seq2D = Seq[Seq[Double]]
+
   type Nullable[A <: Type[_, _]] = HNil :+: A :+: CNil
 
   type InputField[A <: Type[_, _]] = HNil :+: A :+: CNil
@@ -84,19 +86,19 @@ object FortuneTeller {
     } {
       _.head.choice { _ =>
 //        probabilityLossNetwork.compose()
-        probabilityLossNetwork.compose(min(exp(-rowSeq(0, 0)), 1.0))
-//        max(1.0 - rowSeq(0, 0), 0.0)
+        probabilityLossNetwork.compose(min(exp(-rowSeq(0)(0)), 1.0))
+//        max(1.0 - rowSeq(0)(0), 0.0)
       } { inr =>
         val expectedValue = inr.head
-        (rowSeq(0, 0) + abs(rowSeq(0, 1) - expectedValue)): rowAndExpectedLabel.To[Double]
+        (rowSeq(0)(0) + abs(rowSeq(0)(1) - expectedValue)): rowAndExpectedLabel.To[Double]
       }
     }
 
     val loss1 = expectedLabelField1.choice { _ =>
       0.0 // Drop out
     } { expectedEnum =>
-      val score0 = rowSeq(0, 2)
-      val score1 = rowSeq(0, 3)
+      val score0 = rowSeq(0)(2)
+      val score1 = rowSeq(0)(3)
       val sum = score0 + score1 + 0.00000001
       val probability0 = score0 / sum
       val probability1 = score1 / sum
@@ -110,15 +112,15 @@ object FortuneTeller {
     val loss2 = expectedLabelField2.choice { _ =>
       0.0 // Drop out
     } { expectedDouble =>
-      abs(expectedDouble.head - rowSeq(0, 4) + 1.0)
+      abs(expectedDouble.head - rowSeq(0)(4) + 1.0)
     }
 
     val loss3 = expectedLabelField3.choice { _ =>
       0.0 // Drop out
     } { expectedEnum =>
-      val score0 = rowSeq(0, 5)
-      val score1 = rowSeq(0, 6)
-      val score2 = rowSeq(0, 7)
+      val score0 = rowSeq(0)(5)
+      val score1 = rowSeq(0)(6)
+      val score2 = rowSeq(0)(7)
       val sum = score0 + score1 + score2 + 0.00000001
       val probability0 = score0 / sum
       val probability1 = score1 / sum
@@ -142,10 +144,10 @@ object FortuneTeller {
 
   def array2DToRow(implicit input: Array2D): input.To[PredictionResult] = {
     val rowSeq = input.toSeq
-    val field0: input.To[Double :: Double :: HNil] = min(rowSeq(0, 0), 1.0) :: rowSeq(0, 1) :: HNil
-    val field1: input.To[Enum0Prediction] = rowSeq(0, 2) :: rowSeq(0, 3) :: HNil
-    val field2: input.To[Double] = rowSeq(0, 4)
-    val field3 = rowSeq(0, 5) :: rowSeq(0, 6) :: rowSeq(0, 7) :: HNil
+    val field0: input.To[Double :: Double :: HNil] = min(rowSeq(0)(0), 1.0) :: rowSeq(0)(1) :: HNil
+    val field1: input.To[Enum0Prediction] = rowSeq(0)(2) :: rowSeq(0)(3) :: HNil
+    val field2: input.To[Double] = rowSeq(0)(4)
+    val field3 = rowSeq(0)(5) :: rowSeq(0)(6) :: rowSeq(0)(7) :: HNil
     field0 :: field1 :: field2 :: field3 :: HNil
   }
   val array2DToRowNetwork = array2DToRow

array2D/src/main/scala/com/thoughtworks/deeplearning/seq2D/utilities/Seq2DBatch.scala

@@ -11,7 +11,7 @@ import com.thoughtworks.deeplearning.Layer._
 import com.thoughtworks.deeplearning.Batch._
 import com.thoughtworks.deeplearning.hlist._
 import com.thoughtworks.deeplearning.boolean._
-import com.thoughtworks.deeplearning.seq2D._
+import com.thoughtworks.deeplearning.seq._
 import com.thoughtworks.deeplearning.double._
 import com.thoughtworks.deeplearning.array2D._
 import com.thoughtworks.deeplearning.dsl._
@@ -123,7 +123,7 @@ final class XorSpec extends FreeSpec with Matchers {
 
   def decode(implicit row: Array2D): row.To[XorSpec.Output] = {
     val rowSeq = row.toSeq
-    rowSeq(0, 0) :: rowSeq(0, 1) :: rowSeq(0, 2) :: HNil
+    rowSeq(0)(0) :: rowSeq(0)(1) :: rowSeq(0)(2) :: HNil
   }
 
   val decodeNetwork = decode
@@ -150,7 +150,7 @@ final class XorSpec extends FreeSpec with Matchers {
       0.0
     } {
       _.choice { expectedValue =>
-        val value = predictionResult(0, 0)
+        val value = predictionResult(0)(0)
         -expectedValue * log(value) - (1.0 - expectedValue) * log(1.0 - value)
       } { _ =>
         `throw`(new IllegalArgumentException)
@@ -161,7 +161,7 @@ final class XorSpec extends FreeSpec with Matchers {
       0.0
     } {
       _.choice { expectedValue =>
-        val value = predictionResult(0, 1)
+        val value = predictionResult(0)(1)
         -expectedValue * log(value) - (1.0 - expectedValue) * log(1.0 - value)
       } { _ =>
         `throw`(new IllegalArgumentException)
@@ -172,7 +172,7 @@ final class XorSpec extends FreeSpec with Matchers {
       0.0
     } {
       _.choice { expectedValue =>
-        val value = predictionResult(0, 2)
+        val value = predictionResult(0)(2)
         -expectedValue * log(value) - (1.0 - expectedValue) * log(1.0 - value)
       } { _ =>
         `throw`(new IllegalArgumentException)
@@ -195,7 +195,7 @@ final class XorSpec extends FreeSpec with Matchers {
     val field0 = Random.nextBoolean()
     val field1 = Random.nextBoolean()
     val field2 = field0 ^ field1
-    val Seq(dropout0, dropout1, dropout2) = Seq.fill(3)(false).updated(Random.nextInt(3), true)
+    val scala.Seq(dropout0, dropout1, dropout2) = scala.Seq.fill(3)(false).updated(Random.nextInt(3), true)
     def input(isDropout: scala.Boolean, value: scala.Boolean) = {
       if (isDropout) {
         Inl(HNil)