|
| 1 | +program CaiResNet20; |
| 2 | +(* |
| 3 | + Coded by Joao Paulo Schwarz Schuler. |
| 4 | + https://github.com/joaopauloschuler/neural-api |
| 5 | +*) |
| 6 | +{$mode objfpc}{$H+} |
| 7 | + |
| 8 | +uses {$IFDEF UNIX} {$IFDEF UseCThreads} |
| 9 | + cthreads, {$ENDIF} {$ENDIF} |
| 10 | + Classes, SysUtils, CustApp, neuralnetwork, neuralvolume, |
| 11 | + Math, neuraldatasets, neuralfit, neuralthread; |
| 12 | + |
| 13 | +type |
| 14 | + TTestCNNAlgo = class(TCustomApplication) |
| 15 | + protected |
| 16 | + procedure DoRun; override; |
| 17 | + end; |
| 18 | + |
| 19 | + const |
| 20 | + // Padding and cropping constants. |
| 21 | + csPadding = 4; |
| 22 | + csCropSize = csPadding * 2; |
| 23 | + |
| 24 | +procedure CaiOptimizedResnetUnit(pNN: TNNet; pNeurons: integer); |
| 25 | +var |
| 26 | + PreviousLayer, ShortCut, LongPath: TNNetLayer; |
| 27 | + Stride: integer; |
| 28 | +begin |
| 29 | + PreviousLayer := pNN.GetLastLayer(); |
| 30 | + if PreviousLayer.Output.Depth = pNeurons |
| 31 | + then Stride := 1 |
| 32 | + else Stride := 2; |
| 33 | + LongPath := pNN.AddLayer([ |
| 34 | + TNNetConvolutionLinear.Create(pNeurons, {featuresize}3, {padding}1, Stride), |
| 35 | + TNNetReLU.Create(), |
| 36 | + TNNetConvolutionLinear.Create(pNeurons, {featuresize}3, {padding}1, {stride}1), |
| 37 | + TNNetReLUL.Create(-3, 3, 0) |
| 38 | + ]); |
| 39 | + if PreviousLayer.Output.Depth = pNeurons then |
| 40 | + begin |
| 41 | + pNN.AddLayer( TNNetSum.Create([PreviousLayer, LongPath]) ); |
| 42 | + end |
| 43 | + else |
| 44 | + begin |
| 45 | + ShortCut := pNN.AddLayerAfter([ |
| 46 | + TNNetConvolutionReLU.Create(pNeurons, {featuresize}3, {padding}1, Stride) |
| 47 | + ], PreviousLayer); |
| 48 | + pNN.AddLayer( TNNetSum.Create([ShortCut, LongPath]) ); |
| 49 | + end; |
| 50 | +end; |
| 51 | + |
| 52 | + procedure TTestCNNAlgo.DoRun; |
| 53 | + var |
| 54 | + NN: THistoricalNets; |
| 55 | + NeuralFit: TNeuralImageFit; |
| 56 | + ImgTrainingVolumes, ImgValidationVolumes, ImgTestVolumes: TNNetVolumeList; |
| 57 | + ModuleCount, ModuleNumber: integer; |
| 58 | + begin |
| 59 | + ModuleNumber := 3; |
| 60 | + if not CheckCIFARFile() then |
| 61 | + begin |
| 62 | + Terminate; |
| 63 | + exit; |
| 64 | + end; |
| 65 | + WriteLn('Creating Neural Network...'); |
| 66 | + NN := THistoricalNets.Create(); |
| 67 | + NN.AddLayer(TNNetInput.Create(32, 32, 3)); |
| 68 | + NN.AddLayer([ |
| 69 | + TNNetConvolutionLinear.Create({neurons=}16, {featuresize}3, {padding}1, {stride}1), |
| 70 | + TNNetReLU6.Create() |
| 71 | + ]); |
| 72 | + for ModuleCount := 1 to ModuleNumber do CaiOptimizedResnetUnit(NN, 16); |
| 73 | + for ModuleCount := 1 to ModuleNumber do CaiOptimizedResnetUnit(NN, 32); |
| 74 | + for ModuleCount := 1 to ModuleNumber do CaiOptimizedResnetUnit(NN, 64); |
| 75 | + |
| 76 | + NN.AddLayer([ |
| 77 | + TNNetAvgChannel.Create(), |
| 78 | + TNNetFullConnectLinear.Create(10), // The original implementation uses avg pooling. |
| 79 | + TNNetSoftMax.Create() |
| 80 | + ]); |
| 81 | + |
| 82 | + NN.DebugStructure(); |
| 83 | + CreateCifar10Volumes(ImgTrainingVolumes, ImgValidationVolumes, |
| 84 | + ImgTestVolumes, csEncodeRGB, {ValidationSampleSize=}2000); |
| 85 | + |
| 86 | + // Add padding to dataset |
| 87 | + WriteLn |
| 88 | + ( |
| 89 | + 'Original image size: ', |
| 90 | + ImgTrainingVolumes[0].SizeX,',', |
| 91 | + ImgTrainingVolumes[0].SizeY,' px.' |
| 92 | + ); |
| 93 | + ImgTrainingVolumes.AddPadding(csPadding); |
| 94 | + ImgValidationVolumes.AddPadding(csPadding); |
| 95 | + ImgTestVolumes.AddPadding(csPadding); |
| 96 | + WriteLn |
| 97 | + ( |
| 98 | + 'New image size after padding: ', |
| 99 | + ImgTrainingVolumes[0].SizeX,',', |
| 100 | + ImgTrainingVolumes[0].SizeY,' px.' |
| 101 | + ); |
| 102 | + |
| 103 | + NeuralFit := TNeuralImageFit.Create; |
| 104 | + // Enable cropping while fitting. |
| 105 | + NeuralFit.HasImgCrop := true; |
| 106 | + NeuralFit.MaxCropSize := csCropSize; |
| 107 | + NeuralFit.FileNameBase := 'SimpleImageClassifier-'+IntToStr(GetProcessId()); |
| 108 | + NeuralFit.InitialLearningRate := 0.001; |
| 109 | + NeuralFit.LearningRateDecay := 0.01; |
| 110 | + NeuralFit.StaircaseEpochs := 10; |
| 111 | + NeuralFit.Inertia := 0.9; |
| 112 | + NeuralFit.L2Decay := 0.00001; |
| 113 | + NeuralFit.Fit(NN, ImgTrainingVolumes, ImgValidationVolumes, ImgTestVolumes, {NumClasses=}10, {batchsize=}32, {epochs=}50); |
| 114 | + NeuralFit.Free; |
| 115 | + ReadLn(); |
| 116 | + NN.Free; |
| 117 | + ImgTestVolumes.Free; |
| 118 | + ImgValidationVolumes.Free; |
| 119 | + ImgTrainingVolumes.Free; |
| 120 | + Terminate; |
| 121 | + end; |
| 122 | + |
| 123 | +var |
| 124 | + Application: TTestCNNAlgo; |
| 125 | +begin |
| 126 | + Application := TTestCNNAlgo.Create(nil); |
| 127 | + Application.Title:='CIFAR-10 Classification Example'; |
| 128 | + Application.Run; |
| 129 | + Application.Free; |
| 130 | +end. |
0 commit comments