Onnx support: modify unsqueeze function (#2910)

* modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function * modeify unsqueeze function
dding3 · Sep 30, 2019 · 45eabcd · 45eabcd
1 parent 95e3c47
commit 45eabcd
Show file tree

Hide file tree

Showing 10 changed files with 116 additions and 31 deletions.
diff --git a/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/nn/Unsqueeze.scala b/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/nn/Unsqueeze.scala
@@ -23,43 +23,55 @@ import com.intel.analytics.bigdl.tensor._
 import scala.reflect.ClassTag
 
 /**
- * Insert singleton dim (i.e., dimension 1) at position pos. For an input with dim = input.dim(),
+ * Insert singleton dim (i.e., dimension 1) at position array pos.
+ * For an input with dim = input.dim(),
  * there are dim + 1 possible positions to insert the singleton dimension.
+ * Dimension index are 1-based. 0 and negative pos correspond to unsqueeze() applied at
+ * pos = pos + input.dim() + 1
  *
- * @param pos The position will be insert singleton.
+ * @param pos The array of position will insert singleton.
  * @param numInputDims Optional. If in a batch model, set to the inputDims.
  */
 
 @SerialVersionUID(- 5180889605872472241L)
 class Unsqueeze[T: ClassTag](
-  val pos: Int,
+  val pos: Array[Int],
   var numInputDims: Int = Int.MinValue
 )(implicit ev: TensorNumeric[T]) extends AbstractModule[Tensor[_], Tensor[_], T]  {
+  def this( pos: Int, numInputDims: Int )(implicit ev: TensorNumeric[T]) = {
+    this(Array(pos), numInputDims)
+  }
+
+  def this( pos: Int )(implicit ev: TensorNumeric[T]) = {
+    this(Array(pos))
+  }
 
   def setNumInputDims(numInputDims: Int): Unit = {
     this.numInputDims = numInputDims
   }
 
-  private def getActualPosition(input: Tensor[_]) : Int = {
-    val dim = if (pos <= 0) {
-      input.dim() + pos + 1
-    } else {
-      pos
+  private def getActualPosition(input: Tensor[_]) : Array[Int] = {
+    for (index <- 0 until pos.length) {
+      // dimension index are 1-based
+      pos(index) = if (pos(index) <= 0) {
+        input.dim() + pos(index) + 1
+      }
+      else {
+        pos(index)
+      }
+      // get valid dimension offset for batchMode (if any)
+      val inputDim = input.dim() // data batch dim
+      numInputDims = if (numInputDims != Int.MinValue) numInputDims else inputDim // feature map dim
+      val offsetDim = inputDim - numInputDims
+      require(offsetDim >= 0, "input feature map dim (numInputDims) must be <= input:dim()," +
+        s" input feature map dim ${numInputDims}, inputdim ${inputDim}")
+      // the actual position; clearer error message for batchMode (if any)
+      val actualPos = pos(index) + offsetDim
+      require(actualPos >= 1 && actualPos <= (inputDim + 1), s"Invalid position: ${pos(index)}. " +
+        s"input:dim() is $input, input feature map dim (numInputDims) is $numInputDims.")
+      pos(index) = actualPos
     }
-
-    // get valid dimension offset for batchMode (if any)
-    val inputDim = input.dim() // data batch dim
-    numInputDims = if (numInputDims != Int.MinValue) numInputDims else inputDim // feature map dim
-    val offsetDim = inputDim - numInputDims
-    require(offsetDim >= 0, "input feature map dim (numInputDims) must be <= input:dim()," +
-      s" input feature map dim ${numInputDims}, inputdim ${inputDim}")
-
-    // the actual position; clearer error message for batchMode (if any)
-    val actualPos = dim + offsetDim
-    require(actualPos >= 1 && actualPos <= (inputDim + 1), s"Invalid position: $pos. " +
-      s"input:dim() is $input, input feature map dim (numInputDims) is $numInputDims.")
-
-    actualPos
+    pos
   }
 
   override def updateOutput(input: Tensor[_]): Tensor[_] = {
@@ -68,9 +80,8 @@ class Unsqueeze[T: ClassTag](
       output = input.emptyInstance()
     }
 
-    output
-      .asInstanceOf[Tensor[NumericWildcard]]
-      .addSingletonDimension(input.asInstanceOf[Tensor[NumericWildcard]], actualPos)
+    output.asInstanceOf[Tensor[NumericWildcard]]
+    .addMultiDimension(input.asInstanceOf[Tensor[NumericWildcard]], actualPos)
 
     output
   }
@@ -108,6 +119,17 @@ object Unsqueeze {
   def apply[@specialized(Float, Double) T: ClassTag](
     pos: Int,
     numInputDims: Int = Int.MinValue)(implicit ev: TensorNumeric[T]) : Unsqueeze[T] = {
-    new Unsqueeze[T](pos, numInputDims)
+    new Unsqueeze[T](Array(pos), numInputDims)
+  }
+
+  def apply[T: ClassTag](
+    posList: Array[Int],
+    numInputDims: Int)(implicit ev: TensorNumeric[T]) : Unsqueeze[T] = {
+    new Unsqueeze[T](posList, numInputDims)
+  }
+
+  def apply[T: ClassTag](
+    posList: Array[Int])(implicit ev: TensorNumeric[T]) : Unsqueeze[T] = {
+    new Unsqueeze[T](posList)
   }
 }
diff --git a/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/DenseTensor.scala b/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/DenseTensor.scala
@@ -2092,6 +2092,44 @@ private[tensor] class DenseTensor[@specialized T: ClassTag](
     this.set(t.storage(), t.storageOffset(), size, stride)
   }
 
+  override def addMultiDimension( t: Tensor[T], dims: Array[Int] = Array(1)): Tensor[T] = {
+    // increase 1 to the following pos after a previous smaller pos have one dimension inserted.
+    for (i <- 0 until dims.length) {
+      for (j <- i + 1 until dims.length) {
+        if (dims(j) > dims(i)) {
+          dims(j) = dims(j) + 1
+        }
+      }
+    }
+    var temp = t.clone()
+    var size = new Array[Int](t.dim())
+    var stride = new Array[Int](t.dim())
+
+    for ( i <- 0 until dims.length) {
+      require(dims(i) > 0 && dims(i) <= temp.dim() + 1, s"invalid dimension: ${dims(i)}. " +
+        s"Tensor is of ${temp.dim()} dimensions.")
+
+      size = new Array[Int](temp.dim() + 1)
+      stride = new Array[Int](temp.dim() + 1)
+      var d = 0
+      while (d < dims(i) - 1) {
+        size(d) = temp.size(d + 1)
+        stride(d) = temp.stride(d + 1)
+        d += 1
+      }
+      size(dims(i) - 1) = 1
+      stride(dims(i) - 1) = 1
+      d += 1
+      while (d < temp.dim + 1) {
+        size(d) = temp.size(d)
+        stride(d) = temp.stride(d)
+        d += 1
+      }
+      temp.set(temp.storage(), temp.storageOffset(), size, stride)
+    }
+    this.set(temp.storage(), temp.storageOffset(), size, stride)
+  }
+
   /**
    * Implements >= operator comparing each element in x with value
    *

diff --git a/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/DnnTensor.scala b/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/DnnTensor.scala
@@ -332,6 +332,7 @@ object DnnTensor {
     override def getType(): TensorDataType = ???
     override def diff(other: Tensor[T], count: Int, reverse: Boolean): Boolean = ???
     override def addSingletonDimension(t: Tensor[T], dim: Int): Tensor[T] = ???
+    override def addMultiDimension(t: Tensor[T], dims: Array[Int]): Tensor[T] = ???
     override def reshape(sizes: Array[Int]): Tensor[T] = ???
     override def save(path: String, overWrite: Boolean): DnnTensorUnsupportOperations.this.type = ???
     override def getTensorNumeric(): TensorNumeric[T] = ???

diff --git a/...dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/QuantizedTensorUnsupported.scala b/...dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/QuantizedTensorUnsupported.scala
@@ -525,6 +525,16 @@ abstract class QuantizedTensorUnsupported[T: ClassTag] extends Tensor[T] {
   override def addSingletonDimension(t: Tensor[T], dim: Int): Tensor[T] =
     throw new UnsupportedOperationException(errorString)
 
+  /**
+   * view this.tensor and add multiple Dimensions to `dim` dimension
+   *
+   * @param t source tensor
+   * @param dim the specific dimension array, default is [1]
+   * @return this
+   */
+  override def addMultiDimension(t: Tensor[T], dims: Array[Int]): Tensor[T] =
+    throw new UnsupportedOperationException(errorString)
+
   /**
    * create a new tensor without any change of the tensor
    *

diff --git a/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/SparseTensor.scala b/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/SparseTensor.scala
@@ -445,6 +445,10 @@ private[tensor] class SparseTensor[@specialized(Float, Double) T: ClassTag](
     throw new UnsupportedOperationException(s"SparseTensor: Unimplemented method")
   }
 
+  override def addMultiDimension(t: Tensor[T], dims: Array[Int]): Tensor[T] = {
+    throw new UnsupportedOperationException(s"SparseTensor: Unimplemented method")
+  }
+
   override def reshape(sizes: Array[Int]): Tensor[T] = {
     throw new UnsupportedOperationException(s"SparseTensor: Unimplemented method")
   }

diff --git a/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/Tensor.scala b/spark/dl/src/main/scala/com/intel/analytics/bigdl/dllib/tensor/Tensor.scala
@@ -717,6 +717,15 @@ trait Tensor[T] extends Serializable with TensorMath[T] with Activity {
    */
   def addSingletonDimension(t: Tensor[T] = this, dim: Int = 1): Tensor[T]
 
+  /**
+   * view this.tensor and add multiple Dimensions to `dim` dimension
+   *
+   * @param t source tensor
+   * @param dim the specific dimension array, default is [1]
+   * @return this
+   */
+  def addMultiDimension(t: Tensor[T] = this, dims: Array[Int] = Array(1)): Tensor[T]
+
   /**
    * create a new tensor without any change of the tensor
    *

diff --git a/spark/dl/src/main/scala/com/intel/analytics/bigdl/python/api/PythonBigDL.scala b/spark/dl/src/main/scala/com/intel/analytics/bigdl/python/api/PythonBigDL.scala
@@ -1537,10 +1537,10 @@ class PythonBigDL[T: ClassTag](implicit ev: TensorNumeric[T]) extends Serializab
       ip)
   }
 
-  def createUnsqueeze(pos: Int,
+  def createUnsqueeze(pos: JList[Int],
     numInputDims: Int = Int.MinValue)
   : Unsqueeze[T] = {
-    Unsqueeze[T](pos,
+    Unsqueeze[T](pos.asScala.toArray,
       numInputDims)
   }
 

diff --git a/spark/dl/src/test/scala/com/intel/analytics/bigdl/torch/MultiCriterionSpec.scala b/spark/dl/src/test/scala/com/intel/analytics/bigdl/torch/MultiCriterionSpec.scala
@@ -18,20 +18,22 @@ package com.intel.analytics.bigdl.torch
 import com.intel.analytics.bigdl.nn._
 import com.intel.analytics.bigdl._
 import com.intel.analytics.bigdl.tensor.Tensor
+import com.intel.analytics.bigdl.utils.RandomGenerator.RNG
 
 import scala.util.Random
 
 @com.intel.analytics.bigdl.tags.Serial
 class MultiCriterionSpec extends TorchSpec {
     "A MultiCriterion Module " should "generate correct output and grad with Tensor input" in {
     torchCheck()
+    RNG.setSeed(10)
     val module = new MultiCriterion[Double]()
     val nll = new ClassNLLCriterion[Double]()
     val nll2 = new MSECriterion[Double]()
     module.add(nll, 0.5)
     module.add(nll2)
 
-    val input = Tensor[Double](5).apply1(e => Random.nextDouble())
+    val input = Tensor[Double](5).rand()
     val target = Tensor[Double](5)
     target(Array(1)) = 1
     target(Array(2)) = 2

diff --git a/spark/dl/src/test/scala/com/intel/analytics/bigdl/torch/TorchSpec.scala b/spark/dl/src/test/scala/com/intel/analytics/bigdl/torch/TorchSpec.scala
@@ -22,5 +22,4 @@ class TorchSpec extends FlatSpec with BeforeAndAfter with Matchers {
       cancel("Torch is not installed")
     }
   }
-
 }
diff --git a/spark/dl/src/test/scala/com/intel/analytics/bigdl/torch/UnsqueezeSpec.scala b/spark/dl/src/test/scala/com/intel/analytics/bigdl/torch/UnsqueezeSpec.scala
@@ -134,6 +134,6 @@ class UnsqueezeSpec extends TorchSpec {
   "A Unsqueeze(0)" should "generate correct output and grad" in {
     val layer = new Unsqueeze[Double](0)
     val input = Tensor[Double](2, 2).rand()
-    layer.forward(input).size() should be(Array(2, 2, 1))
+    layer.forward(input).size() should be (Array(2, 2, 1))
   }
 }