getMemoryReport produces an exception on a valid ComputationGraph

[bixbyr]

When calling getMemoryReport on a valid ComputationGraph which has a single input, and two outputs, get an exception.

// Network definition
class MultiTaskNetwork {

val l2RegParam = 10e-4
val momentumParam = 0.9

private val networkBaseConf = new NeuralNetConfiguration.Builder()
.trainingWorkspaceMode(WorkspaceMode.SEPARATE)
.learningRate(0.01)
.updater(new Nesterovs(momentumParam))
.graphBuilder()
.setInputTypes(InputType.convolutional(17,19,19))
.addInputs("input")
.addLayer("baseConvolution", new ConvolutionLayer.Builder(Array(3,3), Array(1,1)).convolutionMode(Same).nIn(17).nOut(256).l2Bias(l2RegParam).build, "input")
.addLayer("baseBatchNorm", new BatchNormalization.Builder().build, "baseConvolution")
.addLayer("baseReLu", new ActivationLayer.Builder().activation(RELU).build, "baseBatchNorm")

private val (networkTrunkConf, lastBlockOutput) = (1 to 20).foldLeft((networkBaseConf, "baseReLu")) {
case ((partialNetworkConf, inputName), blockNumber) =>
/*
Each residual block applies the following modules sequentially to its input:
(1) A convolution of 256 filters of kernel size 3 × 3 with stride 1
(2) Batch normalization
(3) A rectifier nonlinearity
(4) A convolution of 256 filters of kernel size 3 × 3 with stride 1
(5) Batch normalization
(6) A skip connection that adds the input to the block
(7) A rectifier nonlinearity
*/

  val confWithNextBlock = partialNetworkConf
    .addLayer(s"block${blockNumber}Conv1", new ConvolutionLayer.Builder(Array(3,3), Array(1,1)).convolutionMode(Same).nOut(256).l2Bias(l2RegParam).build, inputName)

// .addLayer(s"block${blockNumber}Conv1", new ConvolutionLayer.Builder(Array(3,3), Array(1,1)).convolutionMode(Same).nIn(17).nOut(256).l2Bias(l2RegParam).build, inputName)
.addLayer(s"block${blockNumber}BatchNorm1", new BatchNormalization.Builder().build, s"block${blockNumber}Conv1")
.addLayer(s"block${blockNumber}ReLu1", new ActivationLayer.Builder().activation(RELU).build, s"block${blockNumber}BatchNorm1")
.addLayer(s"block${blockNumber}Conv2", new ConvolutionLayer.Builder(Array(3,3), Array(1,1)).convolutionMode(Same).nOut(256).l2Bias(l2RegParam).build, s"block${blockNumber}ReLu1")
// .addLayer(s"block${blockNumber}Conv2", new ConvolutionLayer.Builder(Array(3,3), Array(1,1)).convolutionMode(Same).nIn(17).nOut(256).l2Bias(l2RegParam).build, s"block${blockNumber}ReLu1")
.addLayer(s"block${blockNumber}BatchNorm2", new BatchNormalization.Builder().build, s"block${blockNumber}Conv2")
.addVertex(s"block${blockNumber}SkipConnection", new ElementWiseVertex(ElementWiseVertex.Op.Add), inputName, s"block${blockNumber}BatchNorm2")
.addLayer(s"block${blockNumber}Output", new ActivationLayer.Builder().activation(RELU).build(), s"block${blockNumber}SkipConnection")
(confWithNextBlock, s"block${blockNumber}Output")
}

val fullNetworkConf = networkTrunkConf
/*
Policy head
(1) A convolution of 2 filters of kernel size 1 × 1 with stride 1
(2) Batch normalization
(3) A rectifier nonlinearity
(4) A fully connected linear layer that outputs a vector of size 192 + 1 = 362, corresponding to logit probabilities for all intersections and the pass move
*/
.addLayer("policyConv", new ConvolutionLayer.Builder(Array(1,1), Array(1,1)).convolutionMode(Same).nOut(2).l2Bias(l2RegParam).build, lastBlockOutput)
.addLayer("policyBatchNorm", new BatchNormalization.Builder().activation(RELU).build, "policyConv")
.addLayer("policyReLu", new ActivationLayer.Builder().activation(RELU).build, "policyBatchNorm")
.addLayer("policyFullyConnected", new DenseLayer.Builder().nOut(362).l2Bias(l2RegParam).build, "policyReLu")
.addLayer("policyOutput", new OutputLayer.Builder().lossFunction(MSE).nOut(362).build, "policyFullyConnected")
// Should this be the probit?

/*
Value head
(1) A convolution of 1 filter of kernel size 1 ×  1 with stride 1
(2) Batch normalization
(3) A rectifier nonlinearity
(4) A fully connected linear layer to a hidden layer of size 256
(5) A rectifier nonlinearity
(6) A fully connected linear layer to a scalar
(7) A tanh nonlinearity outputting a scalar in the range [− 1,  1]
 */
.addLayer("valueHeadConv", new ConvolutionLayer.Builder(Array(1,1), Array(1,1)).convolutionMode(Same).nOut(1).l2Bias(l2RegParam).build, lastBlockOutput)
.addLayer("valueBatchNorm", new BatchNormalization.Builder().activation(RELU).build, "valueHeadConv")
.addLayer("valueReLu", new BatchNormalization.Builder().activation(RELU).build, "valueBatchNorm")
.addLayer("valueFullyConnectedReLu", new DenseLayer.Builder().nOut(256).activation(RELU).l2Bias(l2RegParam).build, "valueReLu")
.addLayer("valueSingleNode", new DenseLayer.Builder().nIn(256).nOut(1).activation(TANH).l2Bias(l2RegParam).build, "valueFullyConnectedReLu")
.addLayer("valueOutput", new OutputLayer.Builder().lossFunction(NEGATIVELOGLIKELIHOOD).nOut(1).build, "valueSingleNode")
.setOutputs("valueOutput", "policyOutput")
.build()

val network = new ComputationGraph(fullNetworkConf)
}

// The getMemoryReport statement produces this exception
val multiTaskNetwork = new MultiTaskNetwork()
val memReport = multiTaskNetwork.fullNetworkConf.getMemoryReport(InputType.convolutional(17,19,19))

Exception in thread "main" java.lang.IllegalStateException: Invalid input type (layer index = -1, layer name="policyFullyConnected"): expected FeedForward input type. Got: InputTypeConvolutional(h=17,w=19,d=2)
at org.deeplearning4j.nn.conf.layers.FeedForwardLayer.getOutputType(FeedForwardLayer.java:36)
at org.deeplearning4j.nn.conf.layers.DenseLayer.getMemoryReport(DenseLayer.java:71)
at org.deeplearning4j.nn.conf.graph.LayerVertex.getMemoryReport(LayerVertex.java:133)
at org.deeplearning4j.nn.conf.ComputationGraphConfiguration.getMemoryReport(ComputationGraphConfiguration.java:495)
at com.starfruit.Main$.main(Main.scala:50)
at com.starfruit.Main.main(Main.scala)

Version info:
scalaVersion := "2.11.11"

val nd4jVersion = "0.9.1"
libraryDependencies += "org.deeplearning4j" % "deeplearning4j-core" % nd4jVersion
libraryDependencies += "org.nd4j" % "nd4s_2.11" % nd4jVersion
libraryDependencies += "org.nd4j" % "nd4j-native-platform" % nd4jVersion

[AlexDBlack]

#4326

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.