Merge pull request #310 from frx/streaming_vw

Some VW improvements and python examples
shogun-toolbox · Aug 24, 2011 · a90530f · a90530f
2 parents 7e47856 + 7cf81e6
commit a90530f
Show file tree

Hide file tree

Showing 12 changed files with 145 additions and 11 deletions.
diff --git a/examples/undocumented/python_modular/streaming_vw_createcache_modular.py b/examples/undocumented/python_modular/streaming_vw_createcache_modular.py
@@ -0,0 +1,49 @@
+from modshogun import StreamingVwFile
+from modshogun import StreamingVwCacheFile
+from modshogun import T_SVMLIGHT
+from modshogun import StreamingVwFeatures
+from modshogun import VowpalWabbit
+
+def create_cache():
+    """Creates a binary cache from an ascii data file."""
+
+    # Open the input file as a StreamingVwFile
+    input_file = StreamingVwFile("../data/fm_train_sparsereal.dat")
+    # Default file name will be vw_cache.dat.cache
+    input_file.set_write_to_cache(True)
+
+    # 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)
+    vw.set_no_training(True)
+
+    # Train (in this case does nothing but run over all examples)
+    vw.train()
+
+def train_from_cache():
+    """Train using the generated cache file."""
+
+    # Open the input cache file as a StreamingVwCacheFile
+    input_file = StreamingVwCacheFile("vw_cache.dat.cache");
+
+    # The rest is exactly as for normal input
+    features = StreamingVwFeatures(input_file, True, 1024);
+    vw = VowpalWabbit(features)
+    vw.train()
+
+if __name__ == "__main__":
+    print "Creating cache..."
+    create_cache()
+    print "Done."
+    print
+
+    print "Training using the cache file..."
+    print
+    train_from_cache()
+
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/IO.i b/src/interfaces/modular/IO.i
@@ -27,6 +27,7 @@
 %rename(SerializableXmlFile) CSerializableXmlFile;
 %rename(SimpleFile) CSimpleFile;
 %rename(MemoryMappedFile) CMemoryMappedFile;
+%rename(VwParser) CVwParser;
 
 %include <shogun/io/File.h>
 %include <shogun/io/StreamingFile.h>
@@ -70,6 +71,7 @@ namespace shogun
 
 %include <shogun/io/AsciiFile.h>
 %include <shogun/io/StreamingAsciiFile.h>
+%include <shogun/classifier/vw/VwParser.h>
 %include <shogun/io/StreamingVwFile.h>
 %include <shogun/io/StreamingVwCacheFile.h>
 %include <shogun/io/BinaryFile.h>

diff --git a/src/interfaces/modular/IO_includes.i b/src/interfaces/modular/IO_includes.i
@@ -9,6 +9,7 @@
 #include <shogun/io/StreamingFileFromSimpleFeatures.h>
 #include <shogun/io/AsciiFile.h>
 #include <shogun/io/StreamingAsciiFile.h>
+#include <shogun/classifier/vw/VwParser.h>
 #include <shogun/io/StreamingVwFile.h>
 #include <shogun/io/StreamingVwCacheFile.h>
 #include <shogun/io/BinaryFile.h>

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/VwParser.h b/src/shogun/classifier/vw/VwParser.h
@@ -61,7 +61,7 @@ class CVwParser: public CSGObject
 	/**
 	 * Destructor
 	 */
-	~CVwParser();
+	virtual ~CVwParser();
 
 	/**
 	 * Get the environment

diff --git a/src/shogun/classifier/vw/cache/VwCacheReader.h b/src/shogun/classifier/vw/cache/VwCacheReader.h
@@ -91,6 +91,37 @@ class CVwCacheReader: public CSGObject
 	 */
 	virtual CVwEnvironment* get_env();
 
+	/**
+	 * Update min and max labels seen in the environment
+	 *
+	 * @param label current label based on which to update
+	 */
+	virtual void set_mm(float64_t label)
+	{
+		env->min_label = CMath::min(env->min_label, label);
+		if (label != FLT_MAX)
+			env->max_label = CMath::max(env->max_label, label);
+	}
+
+	/**
+	 * A dummy function performing no operation in case training
+	 * is not to be performed.
+	 *
+	 * @param label label
+	 */
+	virtual void noop_mm(float64_t label) { }
+
+	/**
+	 * Function which is actually called to update min and max labels
+	 * Should be set to one of the functions implemented for this.
+	 *
+	 * @param label label based on which to update
+	 */
+	virtual void set_minmax(float64_t label)
+	{
+		set_mm(label);
+	}
+
 	/**
 	 * Function to read one example from the cache
 	 *

diff --git a/src/shogun/classifier/vw/cache/VwNativeCacheReader.cpp b/src/shogun/classifier/vw/cache/VwNativeCacheReader.cpp
@@ -97,6 +97,8 @@ char* CVwNativeCacheReader::bufread_label(VwLabel* const ld, char* c)
 {
 	ld->label = *(float32_t*)c;
 	c += sizeof(ld->label);
+	set_minmax(ld->label);
+
 	ld->weight = *(float32_t*)c;
 	c += sizeof(ld->weight);
 	ld->initial = *(float32_t*)c;

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)
 	{

diff --git a/src/shogun/io/StreamingVwFile.cpp b/src/shogun/io/StreamingVwFile.cpp
@@ -72,4 +72,5 @@ void CStreamingVwFile::init()
 
 	set_parser_type(T_VW);
 	write_to_cache = false;
+	SG_REF(env);
 }