skipping augmentedImageDatastore to train a net

Hi,

Following the example in "Train Deep Learning Network to Classify New Images",

How can I finetune my net without the augmentedImageDatastore step?

(I want to observe the performance when there arent any variations on the data)

Answer 1

[Cris LaPierre]

You will need to remove it from your network. Look into removeLayer

[omer wagner]

Could I use this on the same network (after performing the training on the Augmented), or that the layer is there in the final architecture?

no_aug_options = trainingOptions('sgdm', ...
    'MiniBatchSize',miniBatchSize, ...
    'MaxEpochs',6, ...
    'InitialLearnRate',3e-4, ...
    'Shuffle','every-epoch', ...
    'ValidationData',imdsValidation, ...
    'ValidationFrequency',valFrequency, ...
    'Verbose',false, ...
    'Plots','training-progress');

no_aug_net = trainNetwork(imdsTrain,lgraph,no_aug_options);

[Cris LaPierre]

The output of removeLayer is a new network. You decide when and where to use this new network.

• netUpdated = removeLayers(net,layerNames)

I believe you will need to retrain your network if you modify it.

Use analyzeNetwork to view the layers of your network.