WorkingDir  = "."
DataDir     = "./data/"
ModelDir    = "$WorkingDir$/DeepCrossing"
ConfigDir   = "."

MBSize=100
LRate=0.0001

CROSSDim1 = 512
CROSSDim2 = 512
CROSSDIM3 = 256
CROSSDIM4 = 128
CROSSDIM5 = 64
MAXEPOCHS = 50

trainFile = "train.bin"
validFile = "valid.bin"

command = train
precision = float
train = [
    action = train
    numMBsToShowResult=500
    deviceId = Auto
    minibatchSize =$MBSize$
    modelPath = $ModelDir$/DeepCrossing.net
    traceLevel = 3

    SGD = [
		epochSize=0
		learningRatesPerSample = $LRate$    #run 1 to 2 epochs, check the result and adjust this value
        momentumPerMB = 0
		maxEpochs=$MAXEPOCHS$

		gradUpdateType=none
		gradientClippingWithTruncation=true
		clippingThresholdPerSample=1#INF
			
		AutoAdjust=[
			reduceLearnRateIfImproveLessThan=0
			loadBestModel=true
			increaseLearnRateIfImproveMoreThan=1000000000
			learnRateDecreaseFactor=0.5
			autoAdjustLR=AdjustAfterEpoch
			learnRateAdjustInterval=2
		]

	]

    BrainScriptNetworkBuilder=[
	    
		// Helper Macro
        ResUnit(InnerDim, OutDim, INNODE)
		{
			L1 = DenseLayer {InnerDim, activation = ReLU, init = "gaussian"}( INNODE )
			L2 = DenseLayer {OutDim, activation = Pass, init = "gaussian"}( L1 )
			ResUnit = ReLU(Plus(INNODE, L2))
		}.ResUnit

        // Constants
        T1Dim = XXX
		T2Dim = XXX
		T3Dim = XXX
		D1Dim = XXX
		D2Dim = XXX
		D3Dim = XXX
		S1Dim = XXX
		S2Dim = XXX

        // Inputs
		T1 = SparseInput(T1Dim)
		T2 = SparseInput(T2Dim)
		T3 = SparseInput(T3Dim)
		D1 = Input(D1Dim)
		D2 = Input(D2Dim)
		D3 = Input(D3Dim)
		S1 = SparseInput(S1Dim)
		S2 = SparseInput(S2Dim)
        

		// Embedding Constants
		EDim = XXX # Previously we stated 128 as a potential default
		
		// Embedding Layer
		T1E = DenseLayer {EDim, activation = ReLU, init = "gaussian"}( T1 )
		T2E = DenseLayer {EDim, activation = ReLU, init = "gaussian"}( T2 )
		T3E = DenseLayer {EDim, activation = ReLU, init = "gaussian"}( T3 )
		S2E = DenseLayer {EDim, activation = ReLU, init = "gaussian"}( S2E )
        
		// Stacking
        c = Splice(( T1E : T2E : T3E : D1 : D2 : D3 : S1 : S2E ))
		
		// ResNet Constants		
		AllDim = EDim + EDim + EDim + EDim + D1Dim + D2Dim + D3Dim + S1Dim
		Inner1 = XXX
		Inner2 = XXX
		Inner3 = XXX
		Inner4 = XXX
		Inner5 = XXX
		
		// ResNet Layers
		r1 = ResNet( Inner1, AllDim, c )
		r2 = ResNet( Inner1, AllDim, r1 )
		r3 = ResNet( Inner1, AllDim, r2 )
		r4 = ResNet( Inner1, AllDim, r3 )
		r5 = ResNet( Inner1, AllDim, r4 )

		// Scoring Layer
		s = DenseLayer {1, activation = Sigmoid, init = "gaussian"}( r5 )
		CE = logistic(Label, s)
		
		// Criteria, output, and evaluation nodes
        criterionNodes = (CE)
		evaluationNodes = (err)
		outputNodes = (s)
    ]

    reader = [
        readerType = "CNTKTextFormatReader"
        file = "$DataDir$/$trainFile$"
    ] 
    cvReader = [
        readerType = "CNTKTextFormatReader"
        file = "$DataDir$/$validFile$"
    ]    
]