How to get inputs of a layer?

[Dzandaa]

Hi,
You can get the outputs of a layer like:
NNAutoencoder.Layers[LayerCnt].Output.SizeX
NNAutoencoder.Layers[LayerCnt].Output.SizeY
NNAutoencoder.Layers[LayerCnt].Output.Depth
but is it possible to get it's inputs?

Like in NNAutoencoder.DebugStructure

Thank you.

[joaopauloschuler]

Hi @Dzandaa,
For a Layers[LayerCnt] layer, except for concatenating layers and input layer, you can get the input via:

NNAutoencoder.Layers[LayerCnt].PrevLayer.Output;

The PrevLayer gives you the previous layer. Then, you can get its output that is the input for LayerCnt layer.

Does this reply solve the question?

[Dzandaa]

Hello Joao,   Thank you very much for your answer.   As I don't create console programs, only GUI, I wanted to see the structure (input-output) of my layers in a TMemo.   And when initializing OpenCL, I had a crash (writeln) on Windows with Win32 gui application enabled    and EasyOpenCL.HideMessages() enabled.   I also wanted to see In Linux the console messages in a TMemo.   But I solved the problem by redirecting the stdout to a TMemo.   I created a denoising program that run on MNIST CIFAR and picture directory.   A strange thing, when I train on MNIST or CIFAR data, NeuralFit.TrainingAccuracy  is always zero, but the result is O.K.   When training on Pictures, NeuralFit.TrainingAccuracy is O.K.   Here is part of my code if you have any idea:     // ********************************* // ***** Create Neural Network ***** // *********************************    if(NNAutoencoder <> nil) then FreeAndNil(NNAutoencoder);   NNAutoencoder := THistoricalNets.Create();   NNAutoencoder.MessageProc := @Self.OnMessage; NNAutoencoder.ErrorProc := @Self.OnError;   NNAutoencoder.HideMessages(); // ***** Hide Messages *****   if HasOpenCL then begin   NNAutoencoder.EnableOpenCL(EasyOpenCL.PlatformIds[0], EasyOpenCL.Devices[0]); end;                                            ...   // Encoder NNAutoencoder.AddLayer([   TNNetInput.Create(XSize, YSize, ZSize),   TNNetConvolutionReLU.Create(32, 3, 1, 1, 1),   TNNetMaxPool.Create(2, 2, 0),   TNNetConvolutionReLU.Create(32, 3, 1, 1, 1),   TNNetMaxPool.Create(2, 2, 0),   // Decoder TNNetConvolutionReLU.Create(32, 3, 1, 1, 1),   TNNetUpsample.Create(),   TNNetConvolutionReLU.Create(32, 3, 1, 1, 1),   TNNetUpsample.Create(),   TNNetConvolutionReLU.Create(32, 3, 1, 1, 1),   TNNetConvolutionReLU.Create(ZSize, 3, 1, 1, 1)   ]);   NNAutoencoder.SetLearningRate(0.0010,0.9); NNAutoencoder.SetL2Decay(0.0);        // ***************************** // ***** Create Neural Fit ***** // ***************************** procedure TDenoizingForm.CreateNeuralFit; begin      if(NeuralFit <> nil) then FreeAndNil(NeuralFit);      NeuralFit := TNeuralDataLoadingFit.Create();        NeuralFit.OnAfterEpoch := @Self.OnAfterEpoch;      NeuralFit.OnAfterStep := @Self.OnAfterStep;      NeuralFit.OnStart := @Self.OnStart;        NeuralFit.MessageProc := @Self.OnMessage;      NeuralFit.ErrorProc := @Self.OnError;        NeuralFit.HideMessages(); // ***** Hide Messages *****        NeuralFit.LearningRateDecay := 0.0;      NeuralFit.L2Decay := 0.0;      NeuralFit.AvgWeightEpochCount := 1;      NeuralFit.InitialLearningRate := 0.01;      NeuralFit.ClipDelta := 0.01;      NeuralFit.FileNameBase := DataName+'-Denoising';      NeuralFit.EnableBipolar99HitComparison();        if HasOpenCL then      begin        NeuralFit.EnableOpenCL(EasyOpenCL.PlatformIds[0], EasyOpenCL.Devices[0]);      end;   end;   // ************************** // ***** Start Training ***** // ************************** ... NeuralFit.FitLoading(NNAutoencoder, {EpochSize=}100, 0, 0, {Batch=}64, {Epochs=1000, @GetTrainingData, nil, nil);     ...   // *************************** // ***** Get Training Data *** // ***************************  procedure TDenoizingForm.GetTrainingData(Idx: integer; ThreadId: integer; pInput, pOutput: TNNetVolume); var ImageId : integer; begin     ImageId := Random(ImageVolumes.Count);   pOutput.Copy(ImageVolumes[ImageId]);     pInput.Copy(pOutput);   pInput.AddGaussianNoise(FSNoise.Value);   end;     My wife is mathematician and Phd in Science. She work in a Spacial Center.   And she is looking to test CAI in comparison with Keras/Tensorflow on a new project for rock analysis.   Is it O.K. for you?   Thank you for that great library.   And sorry for my bad English :)   Have a nice day.   B->       Saturday, March 4, 2023, 1:17:28 PM, you wrote:

…

Hi @Dzandaa, For a Layers[LayerCnt] layer, except for concatenating layers and input layer, you can get the input via: NNAutoencoder.Layers[LayerCnt].PrevLayer.Output; The PrevLayer gives you the previous layer. Then, you can get its output that is the input for LayerCnt layer. Does this reply solve the question? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.

--  Best regards,  Dzandaa                           ***@***.***

[joaopauloschuler]

Denoising is one of the most interesting applications for neural networks in my opinion. I should dedicate more time for it myself. In the case that you publish your source code in full, I'll be happy to add a link to it. I also think that the area of autoencoding super interesting.

I've been playing with autoencoders with 64x64 images. In the case that you are interested, this is my current architecture:

    FAutoencoder.AddLayer([
      TNNetInput.Create(64, 64, 3),
      TNNetConvolution.Create({Features=}32 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}2,{SuppressBias=}1), //32x32
      TNNetConvolution.Create({Features=}32 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetConvolution.Create({Features=}32 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}2,{SuppressBias=}1), //16x16
      TNNetConvolution.Create({Features=}32 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetConvolution.Create({Features=}64 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}2,{SuppressBias=}1), //8x8
      TNNetConvolution.Create({Features=}64 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetConvolution.Create({Features=}128 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}2,{SuppressBias=}1), //4x4
      TNNetConvolution.Create({Features=}128 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),

      TNNetUpsample.Create(), //8x8
      TNNetConvolution.Create({Features=}128 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetConvolution.Create({Features=}128 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetUpsample.Create(), //16x16
      TNNetConvolution.Create({Features=}32 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetConvolution.Create({Features=}128 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetUpsample.Create(), //32x32
      TNNetConvolution.Create({Features=}32 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetConvolution.Create({Features=}128 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetUpsample.Create(), //64x64
      TNNetConvolution.Create({Features=}32 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetConvolution.Create({Features=}32 * NeuronMultiplier,{FeatureSize=}3,{Padding=}1,{Stride=}1,{SuppressBias=}1),
      TNNetConvolutionLinear.Create({Features=}3,{FeatureSize=}1,{Padding=}0,{Stride=}1,{SuppressBias=}0),
      TNNetReLUL.Create(-40, +40, 0) // Protection against overflow
    ]);

For the encoder/decoder, I personally prefer avoiding maxpoolings. I usually use convolutions with stride=2 instead. I also avoid ReLUs. But this is my personal preference. You are free to use the layer of your preference. These are my other parameters:

  FFit.LearningRateDecay := 0.0;
  FFit.L2Decay := 0.0;
  FFit.AvgWeightEpochCount := 1;
  FFit.InitialLearningRate := 0.0001;
  FFit.ClipDelta := 0.01;
  FFit.FileNameBase := FBaseName+'autoencoder';
  FFit.EnableBipolar99HitComparison();

I'm curious to know what your wife finds comparing CAI against Keras. Feel free to share good and bad news.

Regarding "I train on MNIST or CIFAR data, NeuralFit.TrainingAccuracy is always zero", maybe, the problem is around the preprocessing... Not sure.

Keep posting. I enjoy reading.

[Dzandaa]

Hello Joao,
If you don't mind,
I prefer to exchange ideas on CAI on the Pascal Lazarus forum or through messaging.
I don't really like what Microsoft is doing with Github