zkaiwu / dl4j

dl4j for beginner

Jdeepro

Introduction

This is a project for Java beginner and artificial intelligence, written in Latex, based on Java.

Jdeepro is a deep learning framework, providers tools and libraries written specially to take advantage of the Java™ Virtual Machine (JVM), written for the Java and Scala languages.

Just working as expected.

Expected to give you a production release soon. Welcome have a try and send me your patches.

Our vision

1. To be the first Numpy java impl.
2. Give you an consistent experience with Python numpy.
3. Basic guide for AI beginners.

2020/1/12 Demo Handwriting

2019/11/17 Gradient Descend

GradientDescent gd = LinearImp::SGD;
int N = 20;

NDArray x = np.random.uniform(0, 5, N).reshape(N,1);
NDArray ones = np.ones(new int[]{N, 1});

NDArray y = x.multiply(3).add(np.random.uniform(0, 3, N).reshape(N,1));
x = np.hstack(x, ones);

NDArray ret = gd.fit(x, y, 0.01, 1000, 1e-3);
System.out.println("#### test_linear_sgd: y = 2 * x + b ");
System.out.println(ret);

JPlot plot = new JPlot();
double [][]X = (double[][])np.getArray(x.T);
double [][]Y = (double[][])np.getArray(y.T);
double [] W = (double[])np.getArray(ret);

String caption = String.format("y=%f*x+%f", W[0], W[1]);

plot.figure();
plot.scatter(X[0], Y[0], "X-Y");
plot.plot(X[0], a->W[0]*a+W[1], caption);

plot.show();

Easy to use

You can easily change python code to java, like this: https://www.bilibili.com/video/av37947862?p=38

public void test_xor() {
    class np extends Numpy{};
    int[][] dat_X = {{1,0,0},{1,0,1},{1,1,0},{1,1,1}};
    NDArray X = np.array(dat_X);

    int[] dat_Y = {0,1,1,0};
    NDArray Y = np.array(dat_Y).V();

    NDArray V = np.random.rand(new int[]{3,4}).multiply(2).subtract(1);
    NDArray W = np.random.rand(new int[]{4,1}).multiply(2).subtract(1);

    double learn_rate = 0.11;
    Activation sigmoid = new Sigmoid();
    NDArray L1 = null, L2 = null;

    for (int i = 0; i < 20000; i++) {
        L1 = sigmoid.active(np.dot(X, V));
        L2 = sigmoid.active(np.dot(L1, W));

        NDArray L2_delta = Y.T.subtract(L2).multiply(sigmoid.deactive(L2));
        NDArray L1_delta = L2_delta.dot(W.T).multiply(sigmoid.deactive(L1));

        NDArray W_change = L1.T.dot(L2_delta).multiply(learn_rate);
        NDArray V_change = X.T.dot(L1_delta).multiply(learn_rate);

        W = W.add(W_change);
        V = V.add(V_change);

        // Error: np.mean(np.abs(Y.T-L2)) -> decrese to 0
        // System.out.println(np.mean(np.abs(Y.T.subtract(L2))));
    }

    double[] expected = {0,1,1,0};
    assertTrue(L2.same(Numpy.array(expected).V().T, 0.1));
}

y = 3*x1 + 4*x2

NDArray x1 = Numpy.linspace(0, 9, 10).V();
NDArray x2 = Numpy.linspace(4, 13, 10).V();
NDArray x = Numpy.concatenate(x1, x2);
System.out.println(x.T);

NDArray v = Numpy.array(new int[]{3, 4});
NDArray y = Numpy.dot(x.T, v.T);
double[] expected1 = {16.0, 23.0, 30.0, 37.0, 44.0, 51.0, 58.0, 65.0, 72.0, 79.0};
assertEquals(Numpy.array(expected1), y);

v = Numpy.array(new int[]{3, 4}).V();
y = Numpy.dot(x.T, v.T);
double[][] expected2 = {{16.0}, {23.0}, {30.0}, {37.0}, {44.0}, {51.0}, {58.0}, {65.0}, {72.0}, {79.0}};
assertEquals(Numpy.array(expected2), y);

Attendees

1. simbaba 2. jianyang 3. Tatsumi 4. xiaohui

示例代码 Sample

Similar to Python, give you an consistent experience, you can quikly shift from Python to Java.

Running result in jshell Environment

Similar Matplot：

JPlot plot = new JPlot();
double[] x = Range.arange(0, 10, 0.1);

Figure figure = plot.figure();
Axes axes = figure.add_subplot(2, 2, 1);
axes.plot(x, v->v, "y=x");

axes = figure.add_subplot(2, 2, 2);
axes.plot(x, v->v*v, "y=x^2");

axes = figure.add_subplot(2, 2, 3);
axes.plot(x, v->v*v*v, "y=x^3");

axes = figure.add_subplot(2, 2, 4);
axes.plot(x, v->Math.sin(v), "y=sin(x)");

plot.show();

Easy to use, you can quickly learn how to integrate to your code.

@Test
public void test_create() {
    int[] dat = {1,2,3,4,5,6,7,8};
    NDArray a = new NDArray(dat, 2,4);
    assertTrue(a.toString(), Arrays.equals((int[])a.data(), (int[])dat));
}

@Test
public void test_transpose() {
    NDArray a = Numpy.arange(36).reshape(4,3,3);
    int[] expected = {0, 9, 18, 27};
    int[][] range = {{0}, {0}};
    assertEquals(Numpy.array(expected), a.T.slice(range));
    System.out.println(a.T);

    a = Numpy.arange(12).reshape(4,3);
    int[] expected2 = {1, 4, 7};
    int[][] range2 = {{1}, {0,-1}}; // {{1}, {0, 3}}
    assertEquals(Numpy.array(expected2), a.T.slice(range2));
}

@Test
public void test_api_at() {
    NDArray a = Numpy.arange(12).reshape(3,4);
    int[] expected = {4,5,6,7};
    assertEquals(Numpy.array(expected), a.at(1));
    assertEquals(Numpy.array(6), a.at(1,2));
}

@Test
public void test_api_atRange() {
    NDArray a = Numpy.arange(36).reshape(4,3,3);
    int[] expected = {8, 17};
    int[][] range = {{0,2}, {2}, {-1+3}}; // not support negative, but slice can
    assertEquals(Numpy.array(expected), a.atRange(range));
}

@Test
public void test_shape() {
    int[] dat = new int[]{2,4};
    NDArray a = Numpy.array(dat);
    assertEquals("(2,)", a.shape());
}

@Test
public void test_slice() {
    NDArray a = Numpy.arange(12).reshape(3,4);
    int[][] range1 = {{1, 3}, {-1}};
    NDArray b = a.slice(range1);
    assertEquals("array([7, 11])\n", b.toString());

    int[][] range2 = {{ALL}, {-2}};
    b = a.slice(range2);
    assertEquals("array([2, 6, 10])\n", b.toString());

    int[][] range3 = {{-2, -1}, {-2}};
    b = a.slice(range3);
    assertEquals("array([6])\n", b.toString());

    int[][] range4 = {{-1}, {-2}};
    b = a.slice(range4);
    assertEquals("10", b.toString());
}

@Test
public void test_broadcast() {
    NDArray a = Numpy.arange(12).reshape(3,4);
    NDArray b = Numpy.add(a, 2);

    assertEquals("array([[2 , 3 , 4 , 5 ]\n" +
        "       [6 , 7 , 8 , 9 ]\n" +
                "       [10, 11, 12, 13]])\n", b.toString());

    b = Numpy.sub(b, 2);
    assertEquals("array([[0 , 1 , 2 , 3 ]\n" +
                "       [4 , 5 , 6 , 7 ]\n" +
                "       [8 , 9 , 10, 11]])\n", b.toString());

    int[] dat = {10, 20, 30, 40};
    NDArray c = Numpy.add(a, dat);

    assertEquals("array([[10, 21, 32, 43]\n" +
        "       [14, 25, 36, 47]\n" +
                "       [18, 29, 40, 51]])\n", c.toString());
}

@Test
public void test_arange() {
    NDArray a = Numpy.arange(12).reshape(3,4);
    assertEquals("(3, 4)", a.shape());
}

@Test
public void test_random() {
    NDArray a = Numpy.random.rand(4,4);
    System.out.println(a);
    assertEquals("(4, 4)", a.shape());
}

@Test
public void test_ndarray_add() {
    NDArray a = Numpy.ones(3,4);
    NDArray b = Numpy.zeros(3,4);
    NDArray c = Numpy.add(a, b);
    System.out.println(c);
}

@Test
public void test_ndarray_dot() {
    NDArray a = Numpy.arange(12).reshape(3,4);
    NDArray b = Numpy.arange(16).reshape(4,4);
    NDArray c = Numpy.dot(a, b);

    assertEquals(
        "array([[56 , 62 , 68 , 74 ]\n"+
        "       [152, 174, 196, 218]\n"+
        "       [248, 286, 324, 362]])\n", c.toString());

    NDArray d = Numpy.arange(12).reshape(3,4);
    NDArray e = Numpy.dot(d, 2);
    assertEquals(e, Numpy.add(d, d));

    d = Numpy.arange(12).reshape(3,4);
    e = Numpy.array(2);
    assertEquals(Numpy.dot(d, 2), Numpy.dot(d, e));

    d = Numpy.arange(12).reshape(3, 4);
    e = Numpy.array(new int[]{1, 2, 3, 4});
    assertEquals(Numpy.array(new int[]{20, 60, 100}), Numpy.dot(d, e));
}

@Test
public void test_struct() {
    NDArray a = Numpy.arange(24).reshape(2,3,4);
    int[][][] array = (int[][][])Numpy.getArray(a);
    
    int[][][] real = new int[][][]{
        {{0, 1, 2, 3}, {4, 5, 6, 7}, {8, 9, 10, 11}}, 
        {{0, 1, 2, 3}, {4, 5, 6, 7}, {8, 9, 10, 11}}
    };
    
    assertEquals("3 dimens int[][][] array", Arrays.deepToString(array), Arrays.deepToString(real));
}