HashedWDFeatures.cpp

Go to the documentation of this file.
00001 /*
00002  * This program is free software; you can redistribute it and/or modify
00003  * it under the terms of the GNU General Public License as published by
00004  * the Free Software Foundation; either version 3 of the License, or
00005  * (at your option) any later version.
00006  *
00007  * Written (W) 2010 Soeren Sonnenburg
00008  * Copyright (C) 2010 Berlin Institute of Technology
00009  */
00010 
00011 #include <shogun/features/HashedWDFeatures.h>
00012 #include <shogun/io/SGIO.h>
00013 
00014 using namespace shogun;
00015 
00016 CHashedWDFeatures::CHashedWDFeatures() :CDotFeatures()
00017 {
00018     SG_UNSTABLE("CHashedWDFeatures::CHashedWDFeatures()", "\n");
00019 
00020     strings = NULL;
00021 
00022     degree = 0;
00023     start_degree = 0;
00024     from_degree = 0;
00025     string_length = 0;
00026     num_strings = 0;
00027     alphabet_size = 0;
00028     w_dim = 0;
00029     partial_w_dim = 0;
00030     wd_weights = NULL;
00031     mask = 0;
00032     m_hash_bits = 0;
00033 
00034     normalization_const = 0.0;
00035 }
00036 
00037 CHashedWDFeatures::CHashedWDFeatures(CStringFeatures<uint8_t>* str,
00038         int32_t start_order, int32_t order, int32_t from_order,
00039         int32_t hash_bits) : CDotFeatures()
00040 {
00041     ASSERT(start_order>=0);
00042     ASSERT(start_order<order);
00043     ASSERT(order<=from_order);
00044     ASSERT(hash_bits>0);
00045     ASSERT(str);
00046     ASSERT(str->have_same_length());
00047     SG_REF(str);
00048 
00049     strings=str;
00050     string_length=str->get_max_vector_length();
00051     num_strings=str->get_num_vectors();
00052     CAlphabet* alpha=str->get_alphabet();
00053     alphabet_size=alpha->get_num_symbols();
00054     SG_UNREF(alpha);
00055 
00056     degree=order;
00057     start_degree=start_order;
00058     from_degree=from_order;
00059     m_hash_bits=hash_bits;
00060     set_wd_weights();
00061     set_normalization_const();
00062 }
00063 
00064 CHashedWDFeatures::CHashedWDFeatures(const CHashedWDFeatures& orig)
00065     : CDotFeatures(orig), strings(orig.strings),
00066     degree(orig.degree), start_degree(orig.start_degree),
00067     from_degree(orig.from_degree), m_hash_bits(orig.m_hash_bits),
00068     normalization_const(orig.normalization_const)
00069 {
00070     SG_REF(strings);
00071     string_length=strings->get_max_vector_length();
00072     num_strings=strings->get_num_vectors();
00073     CAlphabet* alpha=strings->get_alphabet();
00074     alphabet_size=alpha->get_num_symbols();
00075     SG_UNREF(alpha);
00076 
00077     set_wd_weights();
00078 }
00079 
00080 CHashedWDFeatures::~CHashedWDFeatures()
00081 {
00082     SG_UNREF(strings);
00083     SG_FREE(wd_weights);
00084 }
00085 
00086 float64_t CHashedWDFeatures::dot(int32_t vec_idx1, CDotFeatures* df, int32_t vec_idx2)
00087 {
00088     ASSERT(df);
00089     ASSERT(df->get_feature_type() == get_feature_type());
00090     ASSERT(df->get_feature_class() == get_feature_class());
00091     CHashedWDFeatures* wdf = (CHashedWDFeatures*) df;
00092 
00093     int32_t len1, len2;
00094     bool free_vec1, free_vec2;
00095 
00096     uint8_t* vec1=strings->get_feature_vector(vec_idx1, len1, free_vec1);
00097     uint8_t* vec2=wdf->strings->get_feature_vector(vec_idx2, len2, free_vec2);
00098 
00099     ASSERT(len1==len2);
00100 
00101     float64_t sum=0.0;
00102 
00103     for (int32_t i=0; i<len1; i++)
00104     {
00105         for (int32_t j=0; (i+j<len1) && (j<degree); j++)
00106         {
00107             if (vec1[i+j]!=vec2[i+j])
00108                 break;
00109             if (j>=start_degree)
00110                 sum += wd_weights[j]*wd_weights[j];
00111         }
00112     }
00113     strings->free_feature_vector(vec1, vec_idx1, free_vec1);
00114     wdf->strings->free_feature_vector(vec2, vec_idx2, free_vec2);
00115     return sum/CMath::sq(normalization_const);
00116 }
00117 
00118 float64_t CHashedWDFeatures::dense_dot(int32_t vec_idx1, const float64_t* vec2, int32_t vec2_len)
00119 {
00120     if (vec2_len != w_dim)
00121         SG_ERROR("Dimensions don't match, vec2_dim=%d, w_dim=%d\n", vec2_len, w_dim);
00122 
00123     float64_t sum=0;
00124     int32_t lim=CMath::min(degree, string_length);
00125     int32_t len;
00126     bool free_vec1;
00127     uint8_t* vec = strings->get_feature_vector(vec_idx1, len, free_vec1);
00128     uint32_t* val=SG_MALLOC(uint32_t, len);
00129 
00130     uint32_t offs=0;
00131 
00132     if (start_degree>0)
00133     {
00134         // compute hash for strings of length start_degree-1
00135         for (int32_t i=0; i+start_degree < len; i++)
00136             val[i]=CHash::MurmurHash3(&vec[i], start_degree, 0xDEADBEAF);
00137     }
00138     else
00139         SGVector<uint32_t>::fill_vector(val, len, 0xDEADBEAF);
00140 
00141     for (int32_t k=start_degree; k<lim; k++)
00142     {
00143         float64_t wd = wd_weights[k];
00144 
00145         uint32_t o=offs;
00146         uint32_t carry = 0;
00147         uint32_t chunk = 0;
00148 
00149         for (int32_t i=0; i+k < len; i++)
00150         {
00151             chunk++;
00152             CHash::IncrementalMurmurHash3(&(val[i]), &carry, &(vec[i+k]), 1);
00153             uint32_t h =
00154                     CHash::FinalizeIncrementalMurmurHash3(val[i], carry, chunk);
00155 #ifdef DEBUG_HASHEDWD
00156             SG_PRINT("vec[i]=%d, k=%d, offs=%d o=%d\n", vec[i], k,offs, o);
00157 #endif
00158             sum+=vec2[o+(h & mask)]*wd;
00159             val[i] = h;
00160             o+=partial_w_dim;
00161         }
00162         val[len-k-1] =
00163                 CHash::FinalizeIncrementalMurmurHash3(val[len-k-1], carry, chunk);
00164         offs+=partial_w_dim*len;
00165     }
00166     SG_FREE(val);
00167     strings->free_feature_vector(vec, vec_idx1, free_vec1);
00168 
00169     return sum/normalization_const;
00170 }
00171 
00172 void CHashedWDFeatures::add_to_dense_vec(float64_t alpha, int32_t vec_idx1, float64_t* vec2, int32_t vec2_len, bool abs_val)
00173 {
00174     if (vec2_len != w_dim)
00175         SG_ERROR("Dimensions don't match, vec2_dim=%d, w_dim=%d\n", vec2_len, w_dim);
00176 
00177     int32_t lim=CMath::min(degree, string_length);
00178     int32_t len;
00179     bool free_vec1;
00180     uint8_t* vec = strings->get_feature_vector(vec_idx1, len, free_vec1);
00181     uint32_t* val=SG_MALLOC(uint32_t, len);
00182 
00183     uint32_t offs=0;
00184 
00185     if (start_degree>0)
00186     {
00187         // compute hash for strings of length start_degree-1
00188         for (int32_t i=0; i+start_degree < len; i++)
00189             val[i]=CHash::MurmurHash3(&vec[i], start_degree, 0xDEADBEAF);
00190     }
00191     else
00192         SGVector<uint32_t>::fill_vector(val, len, 0xDEADBEAF);
00193 
00194     for (int32_t k=start_degree; k<lim; k++)
00195     {
00196         float64_t wd = alpha*wd_weights[k]/normalization_const;
00197 
00198         if (abs_val)
00199             wd=CMath::abs(wd);
00200 
00201         uint32_t o=offs;
00202         uint32_t carry = 0;
00203         uint32_t chunk = 0;
00204 
00205         for (int32_t i=0; i+k < len; i++)
00206         {
00207             chunk++;
00208             CHash::IncrementalMurmurHash3(&(val[i]), &carry, &(vec[i+k]), 1);
00209             uint32_t h = CHash::FinalizeIncrementalMurmurHash3(val[i], carry, chunk);
00210 
00211 #ifdef DEBUG_HASHEDWD
00212             SG_PRINT("offs=%d o=%d h=%d \n", offs, o, h);
00213             SG_PRINT("vec[i]=%d, k=%d, offs=%d o=%d\n", vec[i], k,offs, o);
00214 #endif
00215             vec2[o+(h & mask)]+=wd;
00216             val[i] = h;
00217             o+=partial_w_dim;
00218         }
00219         val[len-k-1] =
00220                 CHash::FinalizeIncrementalMurmurHash3(val[len-k-1], carry, chunk);
00221 
00222         offs+=partial_w_dim*len;
00223     }
00224 
00225     SG_FREE(val);
00226     strings->free_feature_vector(vec, vec_idx1, free_vec1);
00227 }
00228 
00229 void CHashedWDFeatures::set_wd_weights()
00230 {
00231     ASSERT(degree>0);
00232 
00233     mask=(uint32_t) (((uint64_t) 1)<<m_hash_bits)-1;
00234     partial_w_dim=1<<m_hash_bits;
00235     w_dim=partial_w_dim*string_length*(degree-start_degree);
00236 
00237     wd_weights=SG_MALLOC(float64_t, degree);
00238 
00239     for (int32_t i=0; i<degree; i++)
00240         wd_weights[i]=sqrt(2.0*(from_degree-i)/(from_degree*(from_degree+1)));
00241 
00242     SG_DEBUG("created HashedWDFeatures with d=%d (%d), alphabetsize=%d, "
00243             "dim=%d partial_dim=%d num=%d, len=%d\n",
00244             degree, from_degree, alphabet_size,
00245             w_dim, partial_w_dim, num_strings, string_length);
00246 }
00247 
00248 
00249 void CHashedWDFeatures::set_normalization_const(float64_t n)
00250 {
00251     if (n==0)
00252     {
00253         normalization_const=0;
00254         for (int32_t i=0; i<degree; i++)
00255             normalization_const+=(string_length-i)*wd_weights[i]*wd_weights[i];
00256 
00257         normalization_const=CMath::sqrt(normalization_const);
00258     }
00259     else
00260         normalization_const=n;
00261 
00262     SG_DEBUG("normalization_const:%f\n", normalization_const);
00263 }
00264 
00265 CFeatures* CHashedWDFeatures::duplicate() const
00266 {
00267     return new CHashedWDFeatures(*this);
00268 }
00269 
00270 
00271 int32_t CHashedWDFeatures::get_nnz_features_for_vector(int32_t num)
00272 {
00273     int32_t vlen=-1;
00274     bool free_vec;
00275     uint8_t* vec=strings->get_feature_vector(num, vlen, free_vec);
00276     strings->free_feature_vector(vec, num, free_vec);
00277     return degree*vlen;
00278 }
00279 
00280 void* CHashedWDFeatures::get_feature_iterator(int32_t vector_index)
00281 {
00282     SG_NOTIMPLEMENTED;
00283     return NULL;
00284 }
00285 
00286 bool CHashedWDFeatures::get_next_feature(int32_t& index, float64_t& value,
00287         void* iterator)
00288 {
00289     SG_NOTIMPLEMENTED;
00290     return NULL;
00291 }
00292 
00293 void CHashedWDFeatures::free_feature_iterator(void* iterator)
00294 {
00295     SG_NOTIMPLEMENTED;
00296 }
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Defines

SHOGUN Machine Learning Toolbox - Documentation