Element must be a scalar

DL4J
1

Hello i’m trying to get a model to predict an handwritten number image, however i am getting this error
java.lang.IllegalArgumentException: Element must be a scalar: element has shape [81600]

However i thought i had already done this by using ImagePreProcessingScaler class

Here is my code

public void TestExampleImage() throws IOException {
    MultiLayerNetwork loadedModel = MultiLayerNetwork.load(locationToSave, saveUpdater);
    BufferedImage myImage = ImageIO.read(new FileInputStream("handwritten4.jpeg"));

    NativeImageLoader loader = new NativeImageLoader(myImage.getHeight(), myImage.getWidth(), 1);
    ImagePreProcessingScaler imageScaler = new ImagePreProcessingScaler();

    INDArray img = loader.asRowVector(myImage);

    INDArray image = img.put(1,img);
    imageScaler.transform(image);


    INDArray output = loadedModel.output(image);
    log.info(output.toString());
}
2

I guess the error comes from the line

    INDArray image = img.put(1,img);

And you are doing it because you had an error similar to Wrong input size - expected matrix - #3 by will-maclean?

Use asMatrix instead of asRowVector and skip this put thing entirely.

3

Hey yes, i added that in following that error. But if i change asMatrix instead i get this error org.deeplearning4j.exception.DL4JInvalidInputException: Input that is not a matrix; expected matrix (rank 2), got rank 4 array with shape [1, 1, 340, 240].

It returns in a 4 rank instead

4

So you are using a model that expects flat input instead of a CNN type input. In this case the vector option is more likely what you want. Then you should just reshape your input using either reshape or expandDims

5

Hey sorry for slow reply, i used the reshaped as you recommended into a 2d format but now i am getting this error
Input size (81600 columns; shape = [1, 81600]) is invalid: does not match layer input size (layer # inputs = 784)

Here is my code

        MultiLayerNetwork loadedModel = MultiLayerNetwork.load(locationToSave, saveUpdater);
    BufferedImage myImage = ImageIO.read(new FileInputStream("handwritten4.jpeg"));

    NativeImageLoader loader = new NativeImageLoader(myImage.getHeight(), myImage.getWidth(), 1);
    ImagePreProcessingScaler imageScaler = new ImagePreProcessingScaler();

    INDArray img = loader.asMatrix(myImage);
    INDArray reshaped = img.reshape(1,81600);


    imageScaler.transform(reshaped);


    INDArray output = loadedModel.output(reshaped);
    log.info(output.toString());
6

It looks very much like you are trying to use a model that was trained on MNIST images (28x28) with your 320x240 image. And your model appears to have a dense input, so it can’t handle the different input size.

You will have to either downscale your image to be 28x28 or use a model that can deal with different input picture sizes.