VwRegressor.cpp

Go to the documentation of this file.

/*
 * Copyright (c) 2009 Yahoo! Inc.  All rights reserved.  The copyrights
 * embodied in the content of this file are licensed under the BSD
 * (revised) open source license.
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 3 of the License, or
 * (at your option) any later version.
 *
 * Written (W) 2011 Shashwat Lal Das
 * Copyright (C) 2011 Berlin Institute of Technology and Max-Planck-Society.
 */

#include <shogun/classifier/vw/VwRegressor.h>
#include <shogun/loss/SquaredLoss.h>
#include <shogun/io/IOBuffer.h>

using namespace shogun;

CVwRegressor::CVwRegressor()
    : CSGObject()
{
    weight_vectors = NULL;
    loss = new CSquaredLoss();
    init(NULL);
}

CVwRegressor::CVwRegressor(CVwEnvironment* env_to_use)
    : CSGObject()
{
    weight_vectors = NULL;
    loss = new CSquaredLoss();
    init(env_to_use);
}

CVwRegressor::~CVwRegressor()
{
    // TODO: the number of weight_vectors depends on num_threads
    // this should be reimplemented using SGVector (for reference counting)
    SG_FREE(weight_vectors);
    SG_UNREF(loss);
    SG_UNREF(env);
}

void CVwRegressor::init(CVwEnvironment* env_to_use)
{
    if (!env_to_use)
        env_to_use = new CVwEnvironment();

    env = env_to_use;
    SG_REF(env);

    // For each feature, there should be 'stride' number of
    // elements in the weight vector
    vw_size_t length = ((vw_size_t) 1) << env->num_bits;
    env->thread_mask = (env->stride * (length >> env->thread_bits)) - 1;

    // Only one learning thread for now
    vw_size_t num_threads = 1;
    weight_vectors = SG_MALLOC(float32_t*, num_threads);

    for (vw_size_t i = 0; i < num_threads; i++)
    {
        weight_vectors[i] = SG_CALLOC(float32_t, env->stride * length / num_threads);

        if (env->random_weights)
        {
            for (vw_size_t j = 0; j < length/num_threads; j++)
                weight_vectors[i][j] = CMath::random(-0.5, 0.5);
        }

        if (env->initial_weight != 0.)
            for (vw_size_t j = 0; j < env->stride*length/num_threads; j+=env->stride)
                weight_vectors[i][j] = env->initial_weight;

        if (env->adaptive)
            for (vw_size_t j = 1; j < env->stride*length/num_threads; j+=env->stride)
                weight_vectors[i][j] = 1;
    }
}

void CVwRegressor::dump_regressor(char* reg_name, bool as_text)
{
    CIOBuffer io_temp;
    int32_t f = io_temp.open_file(reg_name,'w');

    if (f < 0)
        SG_SERROR("Can't open: %s for writing! Exiting.\n", reg_name)

    const char* vw_version = env->vw_version;
    vw_size_t v_length = env->v_length;

    if (!as_text)
    {
        // Write version info
        io_temp.write_file((char*)&v_length, sizeof(v_length));
        io_temp.write_file(vw_version,v_length);

        // Write max and min labels
        io_temp.write_file((char*)&env->min_label, sizeof(env->min_label));
        io_temp.write_file((char*)&env->max_label, sizeof(env->max_label));

        // Write weight vector bits information
        io_temp.write_file((char *)&env->num_bits, sizeof(env->num_bits));
        io_temp.write_file((char *)&env->thread_bits, sizeof(env->thread_bits));

        // For paired namespaces forming quadratic features
        int32_t len = env->pairs.get_num_elements();
        io_temp.write_file((char *)&len, sizeof(len));

        for (int32_t k = 0; k < env->pairs.get_num_elements(); k++)
            io_temp.write_file(env->pairs.get_element(k), 2);

        // ngram and skips information
        io_temp.write_file((char*)&env->ngram, sizeof(env->ngram));
        io_temp.write_file((char*)&env->skips, sizeof(env->skips));
    }
    else
    {
        // Write as human readable form
        char buff[512];
        int32_t len;

        len = sprintf(buff, "Version %s\n", vw_version);
        io_temp.write_file(buff, len);
        len = sprintf(buff, "Min label:%f max label:%f\n", env->min_label, env->max_label);
        io_temp.write_file(buff, len);
        len = sprintf(buff, "bits:%d thread_bits:%d\n", (int32_t)env->num_bits, (int32_t)env->thread_bits);
        io_temp.write_file(buff, len);

        if (env->pairs.get_num_elements() > 0)
        {
            len = sprintf(buff, "\n");
            io_temp.write_file(buff, len);
        }

        len = sprintf(buff, "ngram:%d skips:%d\nindex:weight pairs:\n", (int32_t)env->ngram, (int32_t)env->skips);
        io_temp.write_file(buff, len);
    }

    uint32_t length = 1 << env->num_bits;
    vw_size_t num_threads = env->num_threads();
    vw_size_t stride = env->stride;

    // Write individual weights
    for(uint32_t i = 0; i < length; i++)
    {
        float32_t v;
        v = weight_vectors[i%num_threads][stride*(i/num_threads)];
        if (v != 0.)
        {
            if (!as_text)
            {
                io_temp.write_file((char *)&i, sizeof (i));
                io_temp.write_file((char *)&v, sizeof (v));
            }
            else
            {
                char buff[512];
                int32_t len = sprintf(buff, "%d:%f\n", i, v);
                io_temp.write_file(buff, len);
            }
        }
    }

    io_temp.close_file();
}

void CVwRegressor::load_regressor(char* file)
{
    CIOBuffer source;
    int32_t fd = source.open_file(file, 'r');

    if (fd < 0)
        SG_SERROR("Unable to open file for loading regressor!\n")

    // Read version info
    vw_size_t v_length;
    source.read_file((char*)&v_length, sizeof(v_length));
    char* t = SG_MALLOC(char, v_length);
    source.read_file(t,v_length);
    if (strcmp(t,env->vw_version) != 0)
    {
        SG_FREE(t);
        SG_SERROR("Regressor source has an incompatible VW version!\n")
    }
    SG_FREE(t);

    // Read min and max label
    source.read_file((char*)&env->min_label, sizeof(env->min_label));
    source.read_file((char*)&env->max_label, sizeof(env->max_label));

    // Read num_bits, multiple sources are not supported
    vw_size_t local_num_bits;
    source.read_file((char *)&local_num_bits, sizeof(local_num_bits));

    if ((vw_size_t) env->num_bits != local_num_bits)
        SG_SERROR("Wrong number of bits in regressor source!\n")

    env->num_bits = local_num_bits;

    vw_size_t local_thread_bits;
    source.read_file((char*)&local_thread_bits, sizeof(local_thread_bits));

    env->thread_bits = local_thread_bits;

    int32_t len;
    source.read_file((char *)&len, sizeof(len));

    // Read paired namespace information
    DynArray<char*> local_pairs;
    for (; len > 0; len--)
    {
        char pair[3];
        source.read_file(pair, sizeof(char)*2);
        pair[2]='\0';
        local_pairs.push_back(pair);
    }

    env->pairs = local_pairs;

    // Initialize the weight vector
    if (weight_vectors)
        SG_FREE(weight_vectors);
    init(env);

    vw_size_t local_ngram;
    source.read_file((char*)&local_ngram, sizeof(local_ngram));
    vw_size_t local_skips;
    source.read_file((char*)&local_skips, sizeof(local_skips));

    env->ngram = local_ngram;
    env->skips = local_skips;

    // Read individual weights
    vw_size_t stride = env->stride;
    while (true)
    {
        uint32_t hash;
        ssize_t hash_bytes = source.read_file((char *)&hash, sizeof(hash));
        if (hash_bytes <= 0)
            break;

        float32_t w = 0.;
        ssize_t weight_bytes = source.read_file((char *)&w, sizeof(float32_t));
        if (weight_bytes <= 0)
            break;

        vw_size_t num_threads = env->num_threads();

        weight_vectors[hash % num_threads][(hash*stride)/num_threads]
            = weight_vectors[hash % num_threads][(hash*stride)/num_threads] + w;
    }
    source.close_file();
}