Merge pull request #1 from andre-martins/master

Fix sequence budget factor.

lpsmap/ad3ext/FactorSequenceBudget.h

@@ -91,16 +91,9 @@ class FactorSequenceBudget : public GenericFactor {
     values[0].resize(num_states);
     path[0].resize(num_states);
     for (int l = 0; l < num_states; ++l) {
-      if (l == 0) {
-        // The state that counts for the budget.
-        values[0][l].resize(2);
-        path[0][l].resize(2);
-      } else {
-        values[0][l].resize(1);
-        path[0][l].resize(1);
-      }
+      values[0][l].resize(1);
+      path[0][l].resize(1);
       int bin = 0;
-      if (l == 0) ++bin;
       values[0][l][bin] = 
       GetNodeScore(
         0, l, variable_log_potentials,additional_log_potentials) 
@@ -117,23 +110,18 @@ class FactorSequenceBudget : public GenericFactor {
       path[i+1].resize(num_states);
       for (int k = 0; k < num_states; ++k) {
         int num_bins = (budget_ < i+1)? budget_+1 : i+2;
-        if (k == 0) {
-          // The state that counts for the budget.
-          // k == 0 and b = budget_ not allowed.
-          if (num_bins == budget_+1) --num_bins;
-          values[i+1][k].resize(num_bins+1);
-          path[i+1][k].resize(num_bins+1);
-        } else {
-          values[i+1][k].resize(num_bins);
-          path[i+1][k].resize(num_bins);
-        }
+        values[i+1][k].resize(num_bins);
+        path[i+1][k].resize(num_bins);
         for (int b = 0; b < num_bins; ++b) {
           double best_value = -std::numeric_limits<double>::infinity();
           int best = -1;
           for (int l = 0; l < num_states_[i]; ++l) {
-            if (l == 0 && b == 0) continue;
-            if (i > 0 && path[i][l][b] < 0) continue;
-            double val = values[i][l][b] +
+	    int bin = b;
+	    if (l == 0) --bin; // The state that counts for the budget.
+	    if (bin < 0) continue;
+	    if (bin >= path[i][l].size()) continue;
+	    if (i > 0 && path[i][l][bin] < 0) continue;
+            double val = values[i][l][bin] +
               GetEdgeScore(i+1, l, k, variable_log_potentials,
                            additional_log_potentials);
             if (best < 0 || val > best_value) {
@@ -142,12 +130,11 @@ class FactorSequenceBudget : public GenericFactor {
             }
           }
           int bin = b;
-          if (k == 0) ++bin;
           values[i+1][k][bin] = best_value +
             GetNodeScore(i+1, k, variable_log_potentials,
                          additional_log_potentials);
           path[i+1][k][bin] = best;
-          //cout << "path[" << i+1 << "][" << k << "][" << bin << "] = " << best << endl;
+	  assert(best >= 0);
         }
       }
     }
@@ -159,9 +146,12 @@ class FactorSequenceBudget : public GenericFactor {
     int num_bins = (budget_ < length)? budget_+1 : length+1;
     for (int b = 0; b < num_bins; ++b) {
       for (int l = 0; l < num_states_[length - 1]; ++l) {
-        if (l == 0 && b == 0) continue;
-        if (length > 1 && path[length-1][l][b] < 0) continue;
-        double val = values[length-1][l][b] +
+	int bin = b;
+	if (l == 0) --bin;
+	if (bin < 0) continue;
+	if (bin >= path[length-1][l].size()) continue;
+        if (length > 1 && path[length-1][l][bin] < 0) continue;
+        double val = values[length-1][l][bin] +
           GetEdgeScore(length, l, 0, variable_log_potentials,
                        additional_log_potentials);
         if (best < 0 || val > best_value) {
@@ -180,8 +170,8 @@ class FactorSequenceBudget : public GenericFactor {
     for (int i = length - 1; i > 0; --i) {
       //cout << "sequence[" << i << "] = " << (*sequence)[i] << endl;
       //cout << b << endl;
-      (*sequence)[i - 1] = path[i][(*sequence)[i]][b];
       if ((*sequence)[i] == 0) --b;
+      (*sequence)[i - 1] = path[i][(*sequence)[i]][b];
     }
     *value = best_value;
   }
@@ -341,4 +331,4 @@ class FactorSequenceBudget : public GenericFactor {
 
 } // namespace AD3
 
-#endif // FACTOR_SEQUENCE_BUDGET
+#endif // FACTOR_SEQUENCE_BUDGET

lpsmap/test.py

@@ -14,15 +14,17 @@ def main():
     transition = torch.zeros((n+1,2,2), requires_grad=True)
     transition[:, 1, 1] = 1./temperature
 
-    fg = TorchFactorGraph()
-    u = fg.variable_from(x/temperature)
+    for budget in range(12):
+        fg = TorchFactorGraph()
+        u = fg.variable_from(x/temperature)
 
-    fg.add(SequenceBudget(u, transition, 2))
-    #fg.add(Sequence(u, transition))
-    #fg.add(Budget(u, 2))
-    fg.solve()
+        fg.add(SequenceBudget(u, transition, budget))
+        #fg.add(Sequence(u, transition))
+        #fg.add(Budget(u, 2))
+        fg.solve()
 
-    print("solution: \n", u.value[:,0])
+        print("solution: \n", u.value[:,0])
+        print(sum(u.value[:,0]))
 
 if __name__ == '__main__':
-    main()
+    main()

setup.py

@@ -200,6 +200,8 @@ def install_for_development(self):
     Extension('lpsmap.ad3qp.base', ["lpsmap/ad3qp/base.pyx"]),
     Extension('lpsmap.ad3ext.sequence',
               ["lpsmap/ad3ext/sequence.pyx"]),
+    Extension('lpsmap.ad3ext.sequencebudget',
+              ["lpsmap/ad3ext/sequencebudget.pyx"]),
     Extension('lpsmap.ad3ext.tree',
               ["lpsmap/ad3ext/tree.pyx",
                "lpsmap/ad3ext/DependencyDecoder.cpp"]),