Slide 33
Slide 33 text
MNIST CNN (DNN部リソース)
CNV3x3
CNV3x3
MaxPol
Affine
CNV3x3
CNV3x3
MaxPol
Affine
// sub-networks for convolution(3x3)
bb::NeuralNetSparseMicroMlp<6, 16>sub0_smm0(1 * 3 * 3, 192);
bb::NeuralNetSparseMicroMlp<6, 16>sub0_smm1(192, 32);
bb::NeuralNetGroup<>sub0_net;
sub0_net.AddLayer(&sub0_smm0);
sub0_net.AddLayer(&sub0_smm1);
// sub-networks for convolution(3x3)
bb::NeuralNetSparseMicroMlp<6, 16>sub1_smm0(32 * 3 * 3, 192);
bb::NeuralNetSparseMicroMlp<6, 16>sub1_smm1(192, 32);
bb::NeuralNetGroup<>sub1_net;
sub1_net.AddLayer(&sub1_smm0);
sub1_net.AddLayer(&sub1_smm1);
// sub-networks for convolution(3x3)
bb::NeuralNetSparseMicroMlp<6, 16>sub3_smm0(32 * 3 * 3, 192);
bb::NeuralNetSparseMicroMlp<6, 16>sub3_smm1(192, 32);
bb::NeuralNetGroup<>sub3_net;
sub3_net.AddLayer(&sub3_smm0);
sub3_net.AddLayer(&sub3_smm1);
// sub-networks for convolution(3x3)
bb::NeuralNetSparseMicroMlp<6, 16>sub4_smm0(32 * 3 * 3, 192);
bb::NeuralNetSparseMicroMlp<6, 16>sub4_smm1(192, 32);
bb::NeuralNetGroup<>sub4_net;
sub4_net.AddLayer(&sub4_smm0);
sub4_net.AddLayer(&sub4_smm1);
// main-networks
bb::NeuralNetRealToBinaryinput_real2bin(28 * 28, 28 * 28);
bb::NeuralNetLoweringConvolution<>layer0_conv(&sub0_net, 1, 28, 28, 32, 3, 3);
bb::NeuralNetLoweringConvolution<>layer1_conv(&sub1_net, 32, 26, 26, 32, 3, 3);
bb::NeuralNetMaxPooling<>layer2_maxpol(32, 24, 24, 2, 2);
bb::NeuralNetLoweringConvolution<>layer3_conv(&sub3_net, 32, 12, 12, 32, 3, 3);
bb::NeuralNetLoweringConvolution<>layer4_conv(&sub4_net, 32, 10, 10, 32, 3, 3);
bb::NeuralNetMaxPooling<>layer5_maxpol(32, 8, 8, 2, 2);
bb::NeuralNetSparseMicroMlp<6, 16>layer6_smm(32 * 4 * 4, 480);
bb::NeuralNetSparseMicroMlp<6, 16>layer7_smm(480, 80);
bb::NeuralNetBinaryToRealoutput_bin2real(80, 10);
xc7z020clg400-1
33
DNN部のみ
250MHz / (28x28) = 318,877fps