java

import org.datavec.image.loader.NativeImageLoader;
import org.deeplearning4j.nn.graph.ComputationGraph;
import org.junit.Test;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.ops.transforms.Transforms;

import javax.swing.*;
import java.awt.*;
import java.awt.image.BufferedImage;
import java.io.File;

public class Debug8298 {

    @Test
    public void test() throws Exception {

        File f = new File("C:\\Temp\\Issue8298\\model.h5");
        ComputationGraph model = KerasModelImport.importKerasModelAndWeights(f.getAbsolutePath(), false);

        NativeImageLoader nil = new NativeImageLoader(256, 256, 1);
        File imgF = new File("C:\\Temp\\Issue8298\\n8DQ5.png");
        INDArray arr = nil.asMatrix(imgF);
        arr.divi(255);

        INDArray out = model.output(arr)[0];


        BufferedImage bi = toBI(out);

        JFrame frame = new JFrame();
        frame.getContentPane().setLayout(new FlowLayout());
        frame.getContentPane().add(new JLabel(new ImageIcon(bi)));
        frame.pack();
        frame.setVisible(true);
        frame.setTitle("DL4J");


        INDArray kerasPredict = Nd4j.createFromNpyFile(new File("C:/Temp/Issue8298/prediction.npy"));
        kerasPredict = kerasPredict.permute(0, 3, 1, 2);    //NHWC to NCHW
        System.out.println(kerasPredict.shapeInfoToString());
        BufferedImage biKeras = toBI(kerasPredict);
        JFrame frame2 = new JFrame();
        frame2.getContentPane().setLayout(new FlowLayout());
        frame2.getContentPane().add(new JLabel(new ImageIcon(biKeras)));
        frame2.pack();
        frame2.setVisible(true);
        frame2.setTitle("Keras");

        INDArray absDiff = Transforms.abs(out.sub(kerasPredict));
        System.out.println("Min diff: " + absDiff.minNumber());
        System.out.println("Max diff: " + absDiff.maxNumber());
        System.out.println("Avg diff: " + absDiff.meanNumber());



        Thread.sleep(100000);
    }

    private BufferedImage toBI(INDArray arr){
        int h = (int)arr.size(2);
        int w = (int)arr.size(3);
        BufferedImage bi = new BufferedImage(h, w, BufferedImage.TYPE_BYTE_GRAY);
        int[] ia = new int[1];
        for( int i=0; i<h; i++ ){
            for( int j=0; j<w; j++ ){
                int value = (int)(255 * arr.getDouble(0, 0, i, j));
                ia[0] = value;
                bi.getRaster().setPixel(i,j,ia);
            }
        }

        return bi;
    }
}

python

import keras
from keras.preprocessing.image import load_img
from keras.preprocessing.image import img_to_array
from keras import optimizers
from keras.models import model_from_json

file = open('C:/Temp/Issue8298/unet.json', 'r')
json = file.read()

model = model_from_json(json)

img = load_img('C:/Temp/Issue8298/n8DQ5.png', target_size=(256, 256), color_mode="grayscale")
print(type(img))
print(img.format)
print(img.mode)
print(img.size)
# img.show()
img_array = img_to_array(img)
img_array = np.expand_dims(img_array, axis=0)
img_array /= 255.0
# print(img_array)
print(img_array.dtype)
print(img_array.shape)
print(np.min(img_array))
print(np.max(img_array))
# model.compile(optimizer=optimizers.Adam(lr=0.005), loss='binary_crossentropy')
model.compile(optimizer=optimizers.Adam(lr=0.001), loss='mean_squared_error')
model.fit(img_array, img_array, epochs=20)

model.save("C:/Temp/Issue8298/model.h5")
out = model.predict(img_array)
np.save("C:/Temp/Issue8298/prediction.npy", out)