Backward fail if CDivTable and Sum is used together

[zhichao-li]

Here's the sample code to reproduce this issue:
scala

    val input = Tensor[Float](Shape(2, 3).toSingle().toArray).rand(0, 1)
    val input1 = Input[Float]()
    val s = Sum[Float](dimension = 2, squeeze = false).inputs(input1)
    val div = CDivTable[Float]().inputs(input1, s)
    val graph = Graph(input = input1, output = div)
     val o = graph.forward(input)
     graph.backward(input, o)

requirement failed: invalid size eElement
java.lang.IllegalArgumentException: requirement failed: invalid size eElement
at scala.Predef$.require(Predef.scala:233)
at com.intel.analytics.bigdl.tensor.DenseTensor.view(DenseTensor.scala:172)
at com.intel.analytics.bigdl.nn.Sum.updateGradInput(Sum.scala:101)
at com.intel.analytics.bigdl.nn.Sum.updateGradInput(Sum.scala:44)
at com.intel.analytics.bigdl.nn.abstractnn.AbstractModule.backward(AbstractModule.scala:268)
at com.intel.analytics.bigdl.nn.StaticGraph.backwardExecution(StaticGraph.scala:137)
at com.intel.analytics.bigdl.nn.StaticGraph.backward(StaticGraph.scala:69)

[zhichao-li]

@qiuxin2012 @yiheng

[qiuxin2012]

@zhichao-li input tensor's size?

[zhichao-li]

@qiuxin2012 val input = Tensor[Float](Shape(2, 3).toSingle().toArray).rand(0, 1)

[zhichao-li]

It turn out that it's because of CDivTable lacking the ability to broadcast when size not match in backward stage. An enhancement we can make is to support the nice feature or throw exception explicitly in forward stage temporarily.