Add patience to --auto-stop

ffm-train.cpp

@@ -24,6 +24,7 @@ string train_help() {
 "-l <lambda>: set regularization parameter (default 0.00002)\n"
 "-k <factor>: set number of latent factors (default 4)\n"
 "-t <iteration>: set number of iterations (default 15)\n"
+"-b <patience>: set number of iterations for patience (default 1)\n"
 "-r <eta>: set learning rate (default 0.2)\n"
 "-s <nr_threads>: set number of threads (default 1)\n"
 "-p <path>: set path to the validation set\n"
@@ -70,6 +71,14 @@ Option parse_option(int argc, char **argv) {
             opt.param.nr_iters = atoi(args[i].c_str());
             if(opt.param.nr_iters <= 0)
                 throw invalid_argument("number of iterations should be greater than zero");
+        } else if(args[i].compare("-b") == 0)
+        {
+            if(i == argc-1)
+                throw invalid_argument("need to specify number of iterations after -b");
+            i++;
+            opt.param.nr_patience_iters = atoi(args[i].c_str());
+            if(opt.param.nr_patience_iters <= 0)
+                throw invalid_argument("number of iterations for patience should be greater than zero");
         } else if(args[i].compare("-k") == 0) {
             if(i == argc-1)
                 throw invalid_argument("need to specify number of factors after -k");

ffm.cpp

@@ -537,6 +537,7 @@ ffm_model ffm_train_on_disk(string tr_path, string va_path, ffm_parameter param)
     ffm_model model = init_model(tr.meta.n, tr.meta.m, param);
 
     bool auto_stop = param.auto_stop && !va_path.empty();
+    ffm_int patience = 0;
 
     ffm_long w_size = get_w_size(model);
     vector<ffm_float> prev_W(w_size, 0);
@@ -623,10 +624,14 @@ ffm_model ffm_train_on_disk(string tr_path, string va_path, ffm_parameter param)
 
             if(auto_stop) {
                 if(va_loss > best_va_loss) {
-                    memcpy(model.W, prev_W.data(), w_size*sizeof(ffm_float));
-                    cout << endl << "Auto-stop. Use model at " << iter-1 << "th iteration." << endl;
-                    break;
+                    patience++;
+                    if (patience == param.nr_patience_iters) {
+                        memcpy(model.W, prev_W.data(), w_size*sizeof(ffm_float));
+                        cout << endl << "Auto-stop. Use model at " << (iter - patience) << "th iteration." << endl;
+                        break;
+                    }
                 } else {
+                    patience = 0;
                     memcpy(prev_W.data(), model.W, w_size*sizeof(ffm_float));
                     best_va_loss = va_loss; 
                 }

ffm.h

@@ -34,6 +34,7 @@ struct ffm_parameter {
     ffm_int k = 4; // number of latent factors
     bool normalization = true;
     bool auto_stop = false;
+    ffm_int nr_patience_iters = 1;
 };
 
 void ffm_read_problem_to_disk(string txt_path, string bin_path);