/*********************************************
* Sample particles for a given document
*
* doc:
*********************************************/
LatentSeq DecodeGraph(const Doc doc){
// ----------------------------------------
// init
int nsent = doc.size();
LatentSeq latseq;
// ----------------------------------------
// for each sentence in doc, each latval, compute
// the posterior prob p(R|cvec, sent)
vector<float> U;
for (unsigned sidx = 0; sidx < nsent; sidx ++){
final_hlist.clear();
for (int val = 0; val < nlatvar; val ++){
ComputationGraph cg;
BuildSentGraph(doc[sidx], sidx, cg, val);
float prob = as_scalar(cg.forward());
U.push_back(prob);
cg.clear();
}
// normalize and get the argmax
log_normalize(U);
// greedy decoding
int max_idx = argmax(U);
// get the corresponding context vector
final_h = final_hlist[max_idx];
//
U.clear();
// cerr << "max_latval = " << max_idx << endl;
latseq.push_back(max_idx);
}
// cerr << "====" << endl;
return latseq;
}
int main(int argc, char** argv) {
cnn::Initialize(argc, argv);
/* if (argc == 2) {
ifstream in("");
boost::archive::text_iarchive ia(in);
ia >> m;
}*/
// parameters
ifstream fin("C:\\Data\\msr_train.txt");
ifstream fin_mt("C:\\Data\\mtscore\\All_train_score.txt");
unsigned INPUT_SIZE_MT = 7;
unsigned INPUT_SIZE = 22 * 22;
unsigned DATA_SIZE = 4076;
unsigned OUTPUT_SIZE = 2;
const unsigned HIDDEN1_SIZE = 128;
const unsigned HIDDEN2_SIZE = 64;
const unsigned ITERATIONS = 50000;
fin >> INPUT_SIZE >> DATA_SIZE;
Model m;
//SimpleSGDTrainer sgd(&m);
MomentumSGDTrainer sgd(&m);
Parameters* P_W1 = m.add_parameters({ HIDDEN1_SIZE, INPUT_SIZE + INPUT_SIZE_MT });
Parameters* P_b1 = m.add_parameters({ HIDDEN1_SIZE });
Parameters* P_W2 = m.add_parameters({ HIDDEN2_SIZE, HIDDEN1_SIZE });
Parameters* P_b2 = m.add_parameters({ HIDDEN2_SIZE });
Parameters* P_V = m.add_parameters({ OUTPUT_SIZE, HIDDEN2_SIZE });
Parameters* P_a = m.add_parameters({ OUTPUT_SIZE });
vector<cnn::real> x_values;// (INPUT_SIZE * DATA_SIZE);
x_values.clear();
vector<cnn::real> y_values;// (OUTPUT_SIZE * DATA_SIZE);
y_values.clear();
for (int i = 0; i < DATA_SIZE; ++i)
{
int label;
fin >> label;
if (label == 0)
{
y_values.push_back(cnn::real(0));
y_values.push_back(cnn::real(1));
}
else
{
y_values.push_back(cnn::real(1));
y_values.push_back(cnn::real(0));
}
for (int j = 0; j < INPUT_SIZE_MT; ++j)
{
double x;
fin_mt >> x;
x_values.push_back(cnn::real(x));
}
for (int j = 0; j < INPUT_SIZE; ++j)
{
double x;
fin >> x;
x_values.push_back(cnn::real(x));
}
}
fin.close();
fin_mt.close();
cerr << x_values.size() << '\n' << y_values.size() << '\n';
Dim x_dim({ INPUT_SIZE + INPUT_SIZE_MT }, DATA_SIZE), y_dim({ OUTPUT_SIZE }, DATA_SIZE);
cerr << "x_dim=" << x_dim << ", y_dim=" << y_dim << endl;
//Load dev data
//ifstream f_test("C:\\Data\\msr_train.txt");
ifstream f_test("C:\\Data\\msr_test.txt");
ifstream f_test_mt("C:\\Data\\mtscore\\All_test_score.txt");
vector<cnn::real> x_test_values;// (INPUT_SIZE * DATA_SIZE);
x_test_values.clear();
vector<cnn::real> y_test_values;// (OUTPUT_SIZE * DATA_SIZE);
y_test_values.clear();
unsigned TEST_SIZE;
f_test >> INPUT_SIZE >> TEST_SIZE;
for (int i = 0; i < TEST_SIZE; ++i)
{
int label;
f_test >> label;
if (label == 0)
{
y_test_values.push_back(cnn::real(0));
y_test_values.push_back(cnn::real(1));
}
else
{
y_test_values.push_back(cnn::real(1));
y_test_values.push_back(cnn::real(0));
}
for (int j = 0; j < INPUT_SIZE_MT; ++j)
{
double x;
f_test_mt >> x;
x_test_values.push_back(cnn::real(x));
}
for (int j = 0; j < INPUT_SIZE; ++j)
{
double x;
f_test >> x;
x_test_values.push_back(cnn::real(x));
}
}
f_test_mt.close();
f_test.close();
double max = 0;
int ki = 0;
for (unsigned iter = 0; iter < ITERATIONS; ++iter) {
//for (unsigned iter = 0; true; ++iter) {
// train the parameters
{
ComputationGraph cg;
Expression W1 = parameter(cg, P_W1);
Expression b1 = parameter(cg, P_b1);
Expression W2 = parameter(cg, P_W2);
Expression b2 = parameter(cg, P_b2);
Expression V = parameter(cg, P_V);
Expression a = parameter(cg, P_a);
// set x_values to change the inputs to the network
Expression x = input(cg, x_dim, &x_values);
// set y_values expressing the output
Expression y = input(cg, y_dim, &y_values);
Expression h1 = rectify(W1*x + b1);
Expression h2 = rectify(W2*h1 + b2);
Expression y_pred = softmax(V*h2 + a);
Expression loss = binary_log_loss(y_pred, y);
Expression sum_loss = sum_batches(loss);
//cg.PrintGraphviz();
float my_loss = as_scalar(cg.forward());
cg.backward();
sgd.update(1e-2);
sgd.update_epoch();
cerr << "ITERATIONS = " << iter << '\t';
cerr << "E = " << my_loss << '\t'; //P = 1, iter = 6000, l_rate = 1
}
//DEV SCORE
double l = 0;
for (int i = 0; i < TEST_SIZE; ++i)
{
ComputationGraph cgr;
Expression Wr1 = parameter(cgr, P_W1);
Expression br1 = parameter(cgr, P_b1);
Expression Wr2 = parameter(cgr, P_W2);
Expression br2 = parameter(cgr, P_b2);
Expression Vr = parameter(cgr, P_V);
Expression ar = parameter(cgr, P_a);
vector<cnn::real> x(INPUT_SIZE + INPUT_SIZE_MT);
for (int j = 0; j < INPUT_SIZE + INPUT_SIZE_MT; ++j)
{
x[j] = x_test_values[j + i * (INPUT_SIZE + INPUT_SIZE_MT)];
}
Expression xr = input(cgr, { INPUT_SIZE + INPUT_SIZE_MT }, &x);
vector<cnn::real> y(2);
y[0] = 0; y[1] = 1;
Expression yr = input(cgr, { OUTPUT_SIZE }, &y);
Expression hr1 = rectify(Wr1*xr + br1);
Expression hr2 = rectify(Wr2*hr1 + br2);
Expression y_predr = softmax(Vr*hr2 + ar);
Expression lossr = dot_product(y_predr, yr);
double t = as_scalar(cgr.forward()) > 0.5? 0 : 1;
l += (t == y_test_values[i * 2]) ? 1 : 0;
//cerr << ((t == y_test_values[i * 2]) ? 1 : 0) << ' ';
}
l /= TEST_SIZE;
cerr << "P = " << l << '\t';
if (l > max)
{
max = l;
ki = iter;
}
cerr << "max acc = " << max << "\tat\t" << ki << '\n';
}
//boost::archive::text_oarchive oa(cout);
//oa << m;
system("pause");
return 0;
}