Neural networks in VBA

[drgs]

If you guys have some experience from Pytorch or other major deep learning frameworks, and have some time to spare, give my simple neural network project (maybe the first one in VBA?) a test run. Big plus if you can test it with OpenBLAS dll which you put inside the same folder as the Excel workbook and edit the dll path in BlasFunctions module -- OpenBLAS provides prebuilt binary packages here https://sourceforge.net/projects/ope...files/v0.3.27/:

https://github.com/personalityson/VBANN

Public Sub SetupAndTrainSequential()
    Const MODEL_NAME As String = "MySequentialModel"
    Dim lBatchSize As Long
    Dim lNumEpochs As Long
    Dim lInputSize As Long
    Dim lLabelSize As Long
    Dim oFullSet As TensorDataset
    Dim oTrainingSet As SubsetDataset
    Dim oTrainingLoader As DataLoader
    Dim oTestSet As SubsetDataset
    Dim oTestLoader As DataLoader
    Dim oModel As Sequential

    lInputSize = 8
    lLabelSize = 1
    lBatchSize = 16
    lNumEpochs = 40

    'Prepare training data
    Set oFullSet = ImportDatasetFromWorksheet(ThisWorkbook, "Concrete", Array(lInputSize, lLabelSize), True, False)
    RandomSplit oFullSet, 0.8, oTrainingSet, oTestSet
    Set oTrainingLoader = DataLoader(oTrainingSet, lBatchSize)
    Set oTestLoader = DataLoader(oTestSet, lBatchSize)
    
    'Setup and train
    Set oModel = Sequential(L2Loss(), SGDM())
    oModel.Add InputNormalizationLayer(oTrainingLoader)
    oModel.Add FullyConnectedLayer(lInputSize, 32)
    oModel.Add LeakyReLULayer()
    oModel.Add FullyConnectedLayer(32, 16)
    oModel.Add LeakyReLULayer()
    oModel.Add FullyConnectedLayer(16, 8)
    oModel.Add LeakyReLULayer()
    oModel.Add FullyConnectedLayer(8, lLabelSize)
    oModel.Fit oTrainingLoader, oTestLoader, lNumEpochs

    'Compute test loss
    MsgBox oModel.Loss(oTestLoader)

    'Save to worksheet
    Serialize MODEL_NAME, oModel

    'Load from worksheet
    Set oModel = Unserialize(MODEL_NAME)

    'Compute test loss again with unserialized model
    MsgBox oModel.Loss(oTestLoader)

    Beep
End Sub

Public Sub SetupAndTrainXGBoost()
    Const MODEL_NAME As String = "MyXGBoostModel"
    Dim lInputSize As Long
    Dim lLabelSize As Long
    Dim lNumRounds As Long
    Dim lMaxDepth As Long
    Dim dblLearningRate As Double
    Dim X As Tensor
    Dim T As Tensor
    Dim oFullSet As TensorDataset
    Dim oTrainingSet As SubsetDataset
    Dim oTestSet As SubsetDataset
    Dim oModel As XGBoost

    lInputSize = 8
    lLabelSize = 1
    dblLearningRate = 0.1
    lMaxDepth = 6
    lNumRounds = 100

    'Prepare training data
    Set oFullSet = ImportDatasetFromWorksheet(ThisWorkbook, "Concrete", Array(lInputSize, lLabelSize), True, True)
    RandomSplit oFullSet, 0.8, oTrainingSet, oTestSet
    Set X = oTrainingSet.Cache.Tensor(1)
    Set T = oTrainingSet.Cache.Tensor(2)

    'Setup and train
    Set oModel = XGBoost(L2Loss(), dblLearningRate, lMaxDepth)
    oModel.Fit X, T, lNumRounds

    'Compute test loss
    Set X = oTestSet.Cache.Tensor(1)
    Set T = oTestSet.Cache.Tensor(2)
    MsgBox oModel.Loss(X, T)

    'Save to worksheet
    Serialize MODEL_NAME, oModel

    'Load from worksheet
    Set oModel = Unserialize(MODEL_NAME)

    'Compute test loss again with unserialized model
    MsgBox oModel.Loss(X, T)

    Beep
End Sub

Thanks

[reexre]

Hi
I am interested in NN.
I have implemented a simple one in VB6 that can evolve using backpropagation or Neuroevolution. (all from scratch)
https://www.vbforums.com/showthread....Neural-Network
I would like to try yours because I think it is more advanced (Deep). And maybe it could also be used for "style transfer" on images.

I have (sort of) never worked in VBA, so, do you have a VB6 version ?

[drgs]

Hi
I am interested in NN.
I have implemented a simple one in VB6 that can evolve using backpropagation or Neuroevolution. (all from scratch)
https://www.vbforums.com/showthread....Neural-Network
I would like to try yours because I think it is more advanced (Deep). And maybe it could also be used for "style transfer" on images.

I have (sort of) never worked in VBA, so, do you have a VB6 version ?

Hi,
For me it is the opposite -- I have never worked with VB6. I do rely on saving models to worksheets and reading data from worksheets, so I guess it won't work without modifications.

I only have very basic layers to put together a MLP, but by its modular design it's only a matter of adding new layer classes...
(Currently I'm trying to integrate an extra dll for vector math (Intel Performance Primitives) to speed things up, so I ended up rewriting a few thing.)

For style transfer on images you need at least convolutional layers + pooling layers, I think. Plus, a way to read images from disc, with color channels and everything.
It's possible, but not really what I intended it for

Btw, curious to have a peek at your code, but cant seem to find any source...

[reexre]

Thank you for reply

...
Btw, curious to have a peek at your code, but cant seem to find any source...

it's in the #2 post NNandGA.zip

[yomero]

If you guys have some experience from Pytorch or other major deep learning frameworks, and have some time to spare, give my simple neural network project (maybe the first one in VBA?) a test run. Big plus if you can test it with OpenBLAS dll which you put inside the same folder as the Excel workbook and edit the dll path in BlasFunctions module -- OpenBLAS provides prebuilt binary packages here https://sourceforge.net/projects/ope...files/v0.3.27/:

https://github.com/personalityson/VBANN

Interesting approaching...

I try to test the openBlas option but it don't work or I couldn't do it to work.

The last procedure before crash is this

https://imgur.com/a/R52H6eK

What I'm missing?

I'm using Office 2019 on Windows 7 x86