/** toy example */
public static void test2() {
int N = 500;
double[][] m1 = new double[N][N];
double[][] m2 = new double[N][N];
double[][] m3 = new double[N][N];
// init
Random rand = new Random();
for (int i = 0; i < N; i++)
for (int j = 0; j < N; j++) {
m1[i][j] = 10 * (rand.nextDouble() - 0.2);
m2[i][j] = 20 * (rand.nextDouble() - 0.8);
}
// inverse
System.out.println("Start");
Matrix mat1 = new Matrix(m1);
Matrix mat2 = mat1.inverse();
Matrix mat3 = mat1.times(mat2);
double[][] m4 = mat3.getArray();
/*
for (int i = 0; i < m4.length; i++) {
int ss = 10;
for (int j = 0; j < ss; j++) {
System.out.printf("%f ", m4[i][j]);
}
System.out.print("\n");
}
*/
System.out.println("Done");
/*
// matrix *
System.out.println("Start");
for (int i = 0; i < N; i++) for (int j = 0; j < N; j++) {
double cell = 0;
for (int k = 0; k < N; k++)
cell += m1[i][k] * m2[k][j];
// System.out.printf("%f ", cell);
m3[i][j] = cell;
}
System.out.println("Done");
*/
}
public static void test1(String seed) {
/* read, init data & parameters */
for (int t = t0; t < T; t++) {
// String fileDir = "../../data/graph/" + Integer.toString(t) + ".csv"; // original
// co-voting dataset
// String fileDir = "./data/" + Integer.toString(t) + ".csv"; // artificial toy dataset
String fileDir =
"../../data_sm/nips_17/out/"
+ seed
+ "/"
+ Integer.toString(t)
+ ".train.csv"; // nips dataset (smaller)
Map<Integer, Double> freq = FileParser.readCSVDict(fileDir);
double[][] G = new double[n][n];
double[][] A = new double[n][n];
double[][] mu = new double[n][K];
double[][] mu_hat = new double[n][K];
double[][] mu_prime = new double[n][K];
double[][] mu_hat_prime = new double[n][K];
double[][] h = new double[n][K];
double[][] h_hat = new double[n][K];
FileParser.readCSVGraph(fileDir, freq, G, A);
for (int i = 0; i < n; i++)
for (int k = 0; k < K; k++) {
mu[i][k] = scale_0 * (rand.nextDouble() - 0.5);
mu_hat[i][k] = scale_0 * (rand.nextDouble() - 0.5);
mu_prime[i][k] = mu[i][k];
mu_hat_prime[i][k] = mu_hat[i][k];
h[i][k] = scale * (rand.nextDouble() - 0.5);
h_hat[i][k] = scale * (rand.nextDouble() - 0.5);
}
GS.add(G);
AS.add(A);
mu_s.add(mu);
mu_hat_s.add(mu_hat);
mu_prime_s.add(mu_prime);
mu_hat_prime_s.add(mu_hat_prime);
h_s.add(h);
h_prime_s.add(h);
h_hat_s.add(h_hat);
h_hat_prime_s.add(h);
/* for test */
delta_s.add(delta);
delta_prime_s.add(delta); // TODO previous: 0.1
v_s.add(0.1);
v_hat_s.add(0.1);
v_prime_s.add(0.1);
v_hat_prime_s.add(0.1);
System.out.println("done! t = " + t);
}
for (int t = t0; t < T; t++) {
for (int s = t0; s < T; s++) {
grad_mu_s.add(new double[n][K]);
grad_mu_hat_s.add(new double[n][K]);
grad_mu_prime_s.add(new double[n][K]);
grad_mu_hat_prime_s.add(new double[n][K]);
}
grad_h_hat_s.add(new double[n][K]);
grad_h_hat_prime_s.add(new double[n][K]);
}
/* end initialization */
/* outer for-loop */
double old_obj_1 = -1, old_obj_2 = -1;
for (int iter = 0; iter < MAX_ITER; iter++) {
// Scanner sc = new Scanner(System.in); int gu; gu = sc.nextInt();
System.out.println("====== iter = " + iter + " ======");
/** intrinsic feature * */
forward1(true, iter);
backward1(true);
compute_gradient1(iter);
double new_obj_1 = 0;
/* gradient descent: inner for-loop here */
int inner_iter_1 = 0;
while (inner_iter_1 < INNER_ITER) {
/* update variational parameters \hat{h} using gradient descent */
for (int t = 0; t < T - t0; t++) {
double[][] h_hat_t = h_hat_s.get(t);
double[][] grad_h_hat_t = grad_h_hat_s.get(t);
for (int i = 0; i < n; i++)
for (int k = 0; k < K; k++) {
h_hat_t[i][k] += lr_1 * grad_h_hat_t[i][k];
}
h_hat_s.set(t, h_hat_t);
}
/* update \hat{\mu} and \hat{V}, since both are function of \hat{h} */
forward1(false, iter);
backward1(false);
double obj1 = compute_objective1();
if (inner_iter_1 % 10 == 0)
System.out.println("(1) iter = " + inner_iter_1 + ", obj 1 = " + obj1);
if (inner_iter_1 != 0 && obj1 < new_obj_1) {
lr_1 *= 0.8;
break;
}
new_obj_1 = obj1;
inner_iter_1 += 1;
}
if (inner_iter_1 == INNER_ITER) lr_1 *= 2;
/* sample */
for (int t = 0; t < T - t0; t++) {
double[][] samples =
Operations.sample_multivariate_normal(mu_hat_s.get(t), v_hat_s.get(t), N_SAMPLES);
double[][] h_t = new double[n][K];
for (int i = 0; i < n; i++)
for (int k = 0; k < K; k++) {
h_t[i][k] = samples[i][k];
}
h_s.set(t, h_t);
}
/** impression feature * */
forward2(true);
backward2(true);
compute_gradient2(iter);
double new_obj_2 = 0;
/* gradient descent: inner for-loop here */
int inner_iter_2 = 0;
while (inner_iter_2 < INNER_ITER) {
/* update \hat{h}' using gradient descent */
for (int t = 0; t < T - t0; t++) {
double[][] h_hat_prime_t = h_hat_prime_s.get(t);
double[][] grad_h_hat_prime_t = grad_h_hat_prime_s.get(t);
for (int i = 0; i < n; i++)
for (int k = 0; k < K; k++) {
h_hat_prime_t[i][k] += lr_2 * grad_h_hat_prime_t[i][k];
}
h_hat_prime_s.set(t, h_hat_prime_t);
}
/* update \hat{\mu}' and \hat{V}', since both are function of \hat{h}' */
forward2(false);
backward2(false);
double obj2 = compute_objective2();
if (inner_iter_2 % 10 == 0)
System.out.println("(2) iter = " + inner_iter_2 + ", obj 2 = " + obj2);
if (inner_iter_2 != 0 && obj2 < new_obj_2) {
lr_2 *= 0.8;
break;
}
new_obj_2 = obj2;
inner_iter_2 += 1;
}
if (inner_iter_2 == INNER_ITER) lr_2 *= 2;
/* sample */
for (int t = 0; t < T - t0; t++) {
double[][] samples =
Operations.sample_multivariate_normal(
mu_hat_prime_s.get(t), v_hat_prime_s.get(t), N_SAMPLES);
double[][] h_prime_t = new double[n][K];
for (int i = 0; i < n; i++)
for (int k = 0; k < K; k++) {
h_prime_t[i][k] = samples[i][k];
}
h_prime_s.set(t, h_prime_t);
}
/** output * */
for (int t = 0; t < T - t0; t++) {
double[][] h_t = h_s.get(t);
double[][] h_prime_t = h_prime_s.get(t);
/* output filename:
* ./res/<seed>_<sigma>/h_<time>_<iter>.txt
*/
FileParser.output(
h_t, "./res/" + seed + "_" + delta_str + "/h_" + (t + t0) + "_" + iter + ".txt");
FileParser.output(
h_prime_t,
"./res/" + seed + "_" + delta_str + "/h_p_" + (t + t0) + "_" + iter + ".txt");
}
/* check convergence */
double diff_1 = -(new_obj_1 - old_obj_1) / old_obj_1;
double diff_2 = -(new_obj_2 - old_obj_2) / old_obj_2;
if (iter != 0 && diff_1 < 1e-6 && diff_2 < 1e-6) {
System.out.println("diff_1 = " + diff_1);
System.out.println("diff_2 = " + diff_2);
break;
}
old_obj_1 = new_obj_1;
old_obj_2 = new_obj_2;
}
}