#
# Inception V3 model from:
#
#    https://arxiv.org/pdf/1512.00567.pdf
#
# and in Tensorflow implementation
#
InceptionV3(input, labelDim, bnTimeConst) =
{
    # 299 x 299 x 3
    conv_1 = ConvBNReLULayer{32, (3:3), (2:2), pad = false, bnTimeConst = bnTimeConst}(input)
    # 149 x 149 x 32
    conv_2 = ConvBNReLULayer{32, (3:3), (1:1), pad = false, bnTimeConst = bnTimeConst}(conv_1)
    # 147 x 147 x 32
    conv_3 = ConvBNReLULayer{64, (3:3), (1:1), pad = true, bnTimeConst = bnTimeConst}(conv_2)
    # 147 x 147 x 64
    pool_1 = MaxPoolingLayer{(3:3), stride = (2:2), pad = false}(conv_3)
    # 73 x 73 x 64
    conv_4 = ConvBNReLULayer{80, (1:1), (1:1), pad = true, bnTimeConst = bnTimeConst}(pool_1)
    # 73 x 73 x 80
    conv_5 = ConvBNReLULayer{192, (3:3), (1:1), pad = false, bnTimeConst = bnTimeConst}(conv_4)
    # 71 x 71 x 192
    pool_2 = MaxPoolingLayer{(3:3), stride = (2:2), pad = false}(conv_5)
    # 35 x 35 x 192

    #
    # Inception Blocks
    #
    mixed_1 = InceptionBlock1{64, (48:64), (64:96:96), 32, bnTimeConst}(pool_2)
    # 35 x 35 x 256
    mixed_2 = InceptionBlock1{64, (48:64), (64:96:96), 64, bnTimeConst}(mixed_1)
    # 35 x 35 x 288
    mixed_3 = InceptionBlock1{64, (48:64), (64:96:96), 64, bnTimeConst}(mixed_2)
    # 35 x 35 x 288
    mixed_4 = InceptionBlock2{384, (64:96:96), bnTimeConst}(mixed_3)
    # 17 x 17 x 768
    mixed_5 = InceptionBlock3{192, (128:128:192), (128:128:128:128:192), 192, bnTimeConst}(mixed_4)
    # 17 x 17 x 768
    mixed_6 = InceptionBlock3{192, (160:160:192), (160:160:160:160:192), 192, bnTimeConst}(mixed_5)
    # 17 x 17 x 768
    mixed_7 = InceptionBlock3{192, (160:160:192), (160:160:160:160:192), 192, bnTimeConst}(mixed_6)
    # 17 x 17 x 768
    mixed_8 = InceptionBlock3{192, (192:192:192), (192:192:192:192:192), 192, bnTimeConst}(mixed_7)
    # 17 x 17 x 768
    mixed_9 = InceptionBlock4{(192:320), (192:192:192:192), bnTimeConst}(mixed_8)
    # 17 x 17 x 1280
    mixed_10 = InceptionBlock5{320, (384:384:384), (448:384:384:384), 192, bnTimeConst}(mixed_9)
    # 8 x 8 x 2048
    mixed_11 = InceptionBlock5{320, (384:384:384), (448:384:384:384), 192, bnTimeConst}(mixed_10)
    # 8 x 8 x 2048

    #
    # Prediction
    #
    pool_3 = AveragePoolingLayer{(8:8), pad = false}(mixed_11)
    # 1 x 1 x 2048
    drop = Dropout(pool_3)
    # 1 x 1 x 2048
    z = LinearLayer{labelDim, init = 'heNormal'}(drop)

    #
    # Auxiliary
    #
    # 17 x 17 x 768
    aux_pool_1 = AveragePoolingLayer{(5:5), stride = (3:3), pad = false}(mixed_8)
    # 5 x 5 x 768
    aux_conv_1 = ConvBNReLULayer{128, (1:1), (1:1), pad=true, bnTimeConst = bnTimeConst}(aux_pool_1)
    # 5 x 5 x 128
    aux_conv_2 = ConvBNReLULayer{768, (5:5), (1:1), pad=false, bnTimeConst = bnTimeConst}(aux_conv_1)
    # 1 x 1 x 768    
    aux = LinearLayer{labelDim, init = 'heNormal'}(aux_conv_2)
}

#
# Inception V3 model with normalized input, to use the below function
# remove "ImageNet1K_mean.xml" from each reader.
#
InceptionV3Norm(input, labelDim, bnTimeConst) = 
{
    # Normalize inputs to -1 and 1.
    featMean  = 128
    featScale = 1/128
    Normalize{m,f} = x => f .* (x - m)
            
    inputNorm = Normalize{featMean, featScale}(input)
    model     = InceptionV3(inputNorm, labelDim, bnTimeConst)
}.model