Added VW python example, removed doc warning and added a missing sum_…

…feat_sq init.
shogun-toolbox · Aug 24, 2011 · 0c3aaf8 · 0c3aaf8
1 parent 7e47856
commit 0c3aaf8
Show file tree

Hide file tree

Showing 6 changed files with 60 additions and 10 deletions.
diff --git a/examples/undocumented/python_modular/streaming_vw_modular.py b/examples/undocumented/python_modular/streaming_vw_modular.py
@@ -0,0 +1,26 @@
+from modshogun import StreamingVwFile
+from modshogun import T_SVMLIGHT
+from modshogun import StreamingVwFeatures
+from modshogun import VowpalWabbit
+
+def run_vw():
+    """Runs the VW algorithm on a toy dataset in SVMLight format."""
+
+    # Open the input file as a StreamingVwFile
+    input_file = StreamingVwFile("../data/fm_train_sparsereal.dat")
+
+    # Tell VW that the file is in SVMLight format
+    # Supported types are T_DENSE, T_SVMLIGHT and T_VW
+    input_file.set_parser_type(T_SVMLIGHT)
+
+    # Create a StreamingVwFeatures object, `True' indicating the examples are labelled
+    features = StreamingVwFeatures(input_file, True, 1024)
+
+    # Create a VW object from the features
+    vw = VowpalWabbit(features)
+
+    # Train
+    vw.train()
+
+if __name__ == "__main__":
+    run_vw()
diff --git a/src/interfaces/modular/Classifier.i b/src/interfaces/modular/Classifier.i
@@ -109,6 +109,8 @@
 %include <shogun/classifier/vw/VowpalWabbit.h>
 %include <shogun/classifier/svm/DomainAdaptationSVMLinear.h>
 
+enum E_VW_PARSER_TYPE { T_VW, T_SVMLIGHT, T_DENSE };
+
 #ifdef HAVE_PYTHON
 %pythoncode %{
   class SVM(CSVM):

diff --git a/src/shogun/classifier/vw/VowpalWabbit.cpp b/src/shogun/classifier/vw/VowpalWabbit.cpp
@@ -135,18 +135,22 @@ bool CVowpalWabbit::train_machine(CFeatures* feat)
 		{
 			example = features->get_example();
 
-			if (example->pass != current_pass)
+			// Check if we shouldn't train (generally used for cache creation)
+			if (!no_training)
 			{
-				env->eta *= env->eta_decay_rate;
-				current_pass = example->pass;
-			}
+				if (example->pass != current_pass)
+				{
+					env->eta *= env->eta_decay_rate;
+					current_pass = example->pass;
+				}
 
-			predict_and_finalize(example);
+				predict_and_finalize(example);
 
-			learner->train(example, example->eta_round);
-			example->eta_round = 0.;
+				learner->train(example, example->eta_round);
+				example->eta_round = 0.;
 
-			output_example(example);
+				output_example(example);
+			}
 
 			features->release_example();
 		}
@@ -216,6 +220,7 @@ void CVowpalWabbit::init(CStreamingVwFeatures* feat)
 	SG_REF(reg);
 
 	quiet = false;
+	no_training = false;
 	dump_interval = exp(1.);
 	reg_name = NULL;
 	reg_dump_text = true;

diff --git a/src/shogun/classifier/vw/VowpalWabbit.h b/src/shogun/classifier/vw/VowpalWabbit.h
@@ -62,6 +62,16 @@ class CVowpalWabbit: public COnlineLinearMachine
 	 */
 	void reinitialize_weights();
 
+	/** 
+	 * Set whether one desires to not train and only
+	 * make passes over all examples instead.
+	 *
+	 * This is useful if one wants to create a cache file from data.
+	 * 
+	 * @param dont_train true if one doesn't want to train
+	 */
+	void set_no_training(bool dont_train) { no_training = dont_train; }
+
 	/**
 	 * Set whether learning is adaptive or not
 	 *
@@ -265,6 +275,9 @@ class CVowpalWabbit: public COnlineLinearMachine
 	/// Whether to display statistics or not
 	bool quiet;
 
+	/// Whether we should just run over examples without training
+	bool no_training;
+
 	/// Multiplication factor for number of examples to dump after
 	float32_t dump_interval;
 

diff --git a/src/shogun/classifier/vw/cache/VwNativeCacheReader.h b/src/shogun/classifier/vw/cache/VwNativeCacheReader.h
@@ -21,13 +21,16 @@
 namespace shogun
 {
 
-/// Packed structure for efficient storage
+#ifndef DOXYGEN_SHOULD_SKIP_THIS	
+// Packed structure for efficient storage
 struct one_float
 {
-	/// The float to store
+	// The float to store
 	float32_t f;
 } __attribute__((packed));
 
+#endif // DOXYGEN_SHOULD_SKIP_THIS
+
 /** @brief Class CVwNativeCacheReader reads from a cache exactly as
  * that which has been produced by VW's default cache format.
  *

diff --git a/src/shogun/features/StreamingVwFeatures.cpp b/src/shogun/features/StreamingVwFeatures.cpp
@@ -98,6 +98,7 @@ void CStreamingVwFeatures::setup_example(VwExample* ae)
 	VwFeature temp = {1,constant_hash & env->mask};
 	ae->indices.push(constant_namespace);
 	ae->atomics[constant_namespace].push(temp);
+	ae->sum_feat_sq[constant_namespace] = 0;
 
 	if(env->stride != 1)
 	{