example external feature function

5 files changed, 72 insertions, 1 deletions

diff --git a/Makefile.am b/Makefile.am
index c0826532..3e0103a8 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -14,7 +14,8 @@ SUBDIRS = \
   dpmert \
   pro-train \
   rampion \
-  minrisk
+  minrisk \
+  example_extff
 
 #gi/pyp-topics/src gi/clda/src gi/posterior-regularisation/prjava
 
diff --git a/configure.ac b/configure.ac
index 967b657c..03a0ee87 100644
--- a/configure.ac
+++ b/configure.ac
@@ -125,6 +125,7 @@ AC_CONFIG_FILES([klm/util/Makefile])
 AC_CONFIG_FILES([klm/lm/Makefile])
 AC_CONFIG_FILES([mira/Makefile])
 AC_CONFIG_FILES([dtrain/Makefile])
+AC_CONFIG_FILES([example_extff/Makefile])
 
 AC_CONFIG_FILES([python/setup.py])
 
diff --git a/example_extff/Makefile.am b/example_extff/Makefile.am
new file mode 100644
index 00000000..ac2694ca
--- /dev/null
+++ b/example_extff/Makefile.am
@@ -0,0 +1,5 @@
+AM_CPPFLAGS = -DBOOST_TEST_DYN_LINK -W -Wno-sign-compare $(GTEST_CPPFLAGS) -I.. -I../mteval -I../utils -I../klm -I../decoder
+
+lib_LTLIBRARIES = libff_example.la
+libff_example_la_SOURCES = ff_example.cc
+libff_example_la_LDFLAGS = -version-info 1:0:0 -module
diff --git a/example_extff/README.md b/example_extff/README.md
new file mode 100644
index 00000000..f2aba487
--- /dev/null
+++ b/example_extff/README.md
@@ -0,0 +1,8 @@
+This is an example of an _external_ feature function which is loaded as a dynamically linked library at run time to compute feature functions over derivations in a hypergraph. To load feature external feature functions, you can specify them in your `cdec.ini` configuration file as follows:
+
+	feature_function=External /path/to/libmy_feature.so
+
+Any extra options are passed to the external library.
+
+*Note*: the build system uses [GNU Libtool](http://www.gnu.org/software/libtool/) to create the shared library. This may be placed in a hidden directory called `./libs`.
+
diff --git a/example_extff/ff_example.cc b/example_extff/ff_example.cc
new file mode 100644
index 00000000..51ebf364
--- /dev/null
+++ b/example_extff/ff_example.cc
@@ -0,0 +1,56 @@
+#include "ff.h"
+#include <iostream>
+#include <sstream>
+
+using namespace std;
+
+// example of a "stateful" feature made available as an external library
+// This feature looks nodes and their daughters and fires an indicator based
+// on the arities of the rules involved.
+// (X (X a) b (X c)) - this is a 2 arity parent with children of 0 and 0 arity
+//                     so you get MAF_2_0_0=1
+class ParentChildrenArityFeatures : public FeatureFunction {
+ public:
+  ParentChildrenArityFeatures(const string& param) : fids(16, vector<int>(256, -1)) {
+    SetStateSize(1); // number of bytes extra state required by this Feature
+  }
+  virtual void FinalTraversalFeatures(const void* context,
+                                      SparseVector<double>* features) const {
+    // Goal always is arity 1, so there's no discriminative value of
+    // computing a feature
+  }
+ protected:
+  virtual void TraversalFeaturesImpl(const SentenceMetadata& smeta,
+                                     const Hypergraph::Edge& edge,
+                                     const std::vector<const void*>& ant_contexts,
+                                     FeatureVector* features,
+                                     FeatureVector* estimated_features,
+                                     void* context) const {
+    unsigned child_arity_code = 0;
+    for (unsigned j = 0; j < ant_contexts.size(); ++j) {
+      child_arity_code <<= 4;
+      child_arity_code |= *reinterpret_cast<const unsigned char*>(ant_contexts[j]);
+    }
+    int& fid = fids[edge.Arity()][child_arity_code]; // reference!
+    if (fid < 0) {
+      ostringstream feature_string;
+      feature_string << "MAF_" << edge.Arity();
+      for (unsigned j = 0; j < ant_contexts.size(); ++j)
+        feature_string << '_' << 
+          static_cast<int>(*reinterpret_cast<const unsigned char*>(ant_contexts[j]));
+      fid = FD::Convert(feature_string.str());
+    }
+    features->set_value(fid, 1.0);
+    *reinterpret_cast<unsigned char*>(context) = edge.Arity(); // save state
+  }
+ private:
+  mutable vector<vector<int> > fids;
+};
+
+// IMPORTANT: this function must be implemented by any external FF library
+// if your library has multiple features, you can use str to configure things
+extern "C" FeatureFunction* create_ff(const string& str) {
+  return new ParentChildrenArityFeatures(str);
+}
+
+