crf1m_model.c

CRFsuite is a very fast implmentation of the Conditional Random Fields (CRF) algorithm. It handles t

    }

    return 0;
}

int crf1mmw_open_attrrefs(crf1mmw_t* writer, int num_attrs)
{
    uint32_t offset;
    FILE *fp = writer->fp;
    featureref_header_t* href = NULL;
    size_t size = CHUNK_SIZE + sizeof(uint32_t) * num_attrs;

    /* Check if we aren't writing anything at this moment. */
    if (writer->state != WSTATE_NONE) {
        return CRFERR_INTERNAL_LOGIC;
    }

    /* Allocate a feature reference array. */
    href = (featureref_header_t*)calloc(size, 1);
    if (href == NULL) {
        return CRFERR_OUTOFMEMORY;
    }

    /* Align the offset to a DWORD boundary. */
    offset = (uint32_t)ftell(fp);
    while (offset % 4 != 0) {
        uint8_t c = 0;
        fwrite(&c, sizeof(uint8_t), 1, fp);
        ++offset;
    }

    /* Store the current offset position to the file header. */
    writer->header.off_attrrefs = offset;
    fseek(fp, size, SEEK_CUR);

    /* Fill members in the feature reference header. */
    strncpy(href->chunk, CHUNK_ATTRREF, 4);
    href->size = 0;
    href->num = num_attrs;

    writer->href = href;
    writer->state = WSTATE_ATTRREFS;
    return 0;
}

int crf1mmw_close_attrrefs(crf1mmw_t* writer)
{
    uint32_t i;
    FILE *fp = writer->fp;
    featureref_header_t* href = writer->href;
    uint32_t begin = writer->header.off_attrrefs, end = 0;

    /* Make sure that we are writing attribute feature references. */
    if (writer->state != WSTATE_ATTRREFS) {
        return CRFERR_INTERNAL_LOGIC;
    }

    /* Store the current offset position. */
    end = (uint32_t)ftell(fp);

    /* Compute the size of this chunk. */
    href->size = (end - begin);

    /* Write the chunk header and offset array. */
    fseek(fp, begin, SEEK_SET);
    write_uint8_array(fp, href->chunk, 4);
    write_uint32(fp, href->size);
    write_uint32(fp, href->num);
    for (i = 0;i < href->num;++i) {
        write_uint32(fp, href->offsets[i]);
    }

    /* Move the file pointer to the tail. */
    fseek(fp, end, SEEK_SET);

    /* Uninitialize. */
    free(href);
    writer->href = NULL;
    writer->state = WSTATE_NONE;
    return 0;
}

int crf1mmw_put_attrref(crf1mmw_t* writer, int aid, const feature_refs_t* ref, int *map)
{
    int i, fid;
    uint32_t n = 0, offset = 0;
    FILE *fp = writer->fp;
    featureref_header_t* href = writer->href;

    /* Make sure that we are writing attribute feature references. */
    if (writer->state != WSTATE_ATTRREFS) {
        return CRFERR_INTERNAL_LOGIC;
    }

    /* Store the current offset to the offset array. */
    href->offsets[aid] = ftell(fp);

    /* Count the number of references to active features. */
    for (i = 0;i < ref->num_features;++i) {
        if (0 <= map[ref->fids[i]]) ++n;
    }

    /* Write the feature reference. */
    write_uint32(fp, (uint32_t)n);
    for (i = 0;i < ref->num_features;++i) {
        fid = map[ref->fids[i]];
        if (0 <= fid) write_uint32(fp, (uint32_t)fid);
    }

    return 0;
}

int crf1mmw_open_features(crf1mmw_t* writer)
{
    FILE *fp = writer->fp;
    feature_header_t* hfeat = NULL;

    /* Check if we aren't writing anything at this moment. */
    if (writer->state != WSTATE_NONE) {
        return CRFERR_INTERNAL_LOGIC;
    }

    /* Allocate a feature chunk header. */
    hfeat = (feature_header_t*)calloc(sizeof(feature_header_t), 1);
    if (hfeat == NULL) {
        return CRFERR_OUTOFMEMORY;
    }

    writer->header.off_features = (uint32_t)ftell(fp);
    fseek(fp, CHUNK_SIZE, SEEK_CUR);

    strncpy(hfeat->chunk, CHUNK_FEATURE, 4);
    writer->hfeat = hfeat;

    writer->state = WSTATE_FEATURES;
    return 0;
}

int crf1mmw_close_features(crf1mmw_t* writer)
{
    FILE *fp = writer->fp;
    feature_header_t* hfeat = writer->hfeat;
    uint32_t begin = writer->header.off_features, end = 0;

    /* Make sure that we are writing attribute feature references. */
    if (writer->state != WSTATE_FEATURES) {
        return CRFERR_INTERNAL_LOGIC;
    }

    /* Store the current offset position. */
    end = (uint32_t)ftell(fp);

    /* Compute the size of this chunk. */
    hfeat->size = (end - begin);

    /* Write the chunk header and offset array. */
    fseek(fp, begin, SEEK_SET);
    write_uint8_array(fp, hfeat->chunk, 4);
    write_uint32(fp, hfeat->size);
    write_uint32(fp, hfeat->num);

    /* Move the file pointer to the tail. */
    fseek(fp, end, SEEK_SET);

    /* Uninitialize. */
    free(hfeat);
    writer->hfeat = NULL;
    writer->state = WSTATE_NONE;
    return 0;
}

int crf1mmw_put_feature(crf1mmw_t* writer, int fid, const crf1mm_feature_t* f)
{
    FILE *fp = writer->fp;
    feature_header_t* hfeat = writer->hfeat;

    /* Make sure that we are writing attribute feature references. */
    if (writer->state != WSTATE_FEATURES) {
        return CRFERR_INTERNAL_LOGIC;
    }

    /* We must put features #0, #1, ..., #(K-1) in this order. */
    if (fid != hfeat->num) {
        return CRFERR_INTERNAL_LOGIC;
    }

    write_uint32(fp, f->type);
    write_uint32(fp, f->src);
    write_uint32(fp, f->dst);
    write_float(fp, f->weight);
    ++hfeat->num;
    return 0;
}

crf1mm_t* crf1mm_new(const char *filename)
{
    FILE *fp = NULL;
    uint8_t* p = NULL;
    crf1mm_t *model = NULL;
    header_t *header = NULL;

    model = (crf1mm_t*)calloc(1, sizeof(crf1mm_t));
    if (model == NULL) {
        goto error_exit;
    }

    fp = fopen(filename, "rb");
    if (fp == NULL) {
        goto error_exit;
    }

    fseek(fp, 0, SEEK_END);
    model->size = (uint32_t)ftell(fp);
    fseek(fp, 0, SEEK_SET);

    model->buffer = (uint8_t*)malloc(model->size + 16);
    while ((uint32_t)model->buffer % 16 != 0) {
        ++model->buffer;
    }

    fread(model->buffer, 1, model->size, fp);
    fclose(fp);

    /* Write the file header. */
    header = (header_t*)calloc(1, sizeof(header_t));

    p = model->buffer;
    p += read_uint8_array(p, header->magic, sizeof(header->magic));
    p += read_uint32(p, &header->size);
    p += read_uint8_array(p, header->type, sizeof(header->type));
    p += read_uint32(p, &header->version);
    p += read_uint32(p, &header->num_features);
    p += read_uint32(p, &header->num_labels);
    p += read_uint32(p, &header->num_attrs);
    p += read_uint32(p, &header->off_features);
    p += read_uint32(p, &header->off_labels);
    p += read_uint32(p, &header->off_attrs);
    p += read_uint32(p, &header->off_labelrefs);
    p += read_uint32(p, &header->off_attrrefs);
    model->header = header;

    model->labels = cqdb_reader(
        model->buffer + header->off_labels,
        model->size - header->off_labels
        );

    model->attrs = cqdb_reader(
        model->buffer + header->off_attrs,
        model->size - header->off_attrs
        );

    return model;

error_exit:
    if (model != NULL) {
        free(model);
    }
    if (fp != NULL) {
        fclose(fp);
    }
    return NULL;
}

void crf1mm_close(crf1mm_t* model)
{
    if (model->labels != NULL) {
        cqdb_delete(model->labels);
    }
    if (model->attrs != NULL) {
        cqdb_delete(model->attrs);
    }
    free(model);
}

int crf1mm_get_num_attrs(crf1mm_t* model)
{
    return model->header->num_attrs;
}

int crf1mm_get_num_labels(crf1mm_t* model)
{
    return model->header->num_labels;
}

const char *crf1mm_to_label(crf1mm_t* model, int lid)
{
    if (model->labels != NULL) {
        return cqdb_to_string(model->labels, lid);
    } else {
        return NULL;
    }
}

int crf1mm_to_lid(crf1mm_t* model, const char *value)
{
    if (model->labels != NULL) {
        return cqdb_to_id(model->labels, value);
    } else {
        return -1;
    }
}

int crf1mm_to_aid(crf1mm_t* model, const char *value)
{
    if (model->attrs != NULL) {
        return cqdb_to_id(model->attrs, value);
    } else {
        return -1;
    }
}

const char *crf1mm_to_attr(crf1mm_t* model, int aid)
{
    if (model->attrs != NULL) {
        return cqdb_to_string(model->attrs, aid);
    } else {
        return NULL;
    }
}

int crf1mm_get_labelref(crf1mm_t* model, int lid, feature_refs_t* ref)
{
    uint8_t *p = model->buffer;
    uint32_t offset;

    p += model->header->off_labelrefs;
    p += CHUNK_SIZE;
    p += sizeof(uint32_t) * lid;
    read_uint32(p, &offset);

    p = model->buffer + offset;
    p += read_uint32(p, &ref->num_features);
    ref->fids = (int*)p;
    return 0;
}

int crf1mm_get_attrref(crf1mm_t* model, int aid, feature_refs_t* ref)
{
    uint8_t *p = model->buffer;
    uint32_t offset;

    p += model->header->off_attrrefs;
    p += CHUNK_SIZE;
    p += sizeof(uint32_t) * aid;
    read_uint32(p, &offset);

    p = model->buffer + offset;
    p += read_uint32(p, &ref->num_features);
    ref->fids = (int*)p;
    return 0;
}

int crf1mm_get_featureid(feature_refs_t* ref, int i)
{
    uint32_t fid;
    uint8_t* p = (uint8_t*)ref->fids;
    p += sizeof(uint32_t) * i;
    read_uint32(p, &fid);
    return (int)fid;
}

int crf1mm_get_feature(crf1mm_t* model, int fid, crf1mm_feature_t* f)
{
    uint8_t *p = NULL;
    uint32_t val = 0;
    uint32_t offset = model->header->off_features + CHUNK_SIZE;
    offset += FEATURE_SIZE * fid;
    p = model->buffer + offset;
    p += read_uint32(p, &val);
    f->type = val;
    p += read_uint32(p, &val);
    f->src = val;
    p += read_uint32(p, &val);
    f->dst = val;
    p += read_float(p, &f->weight);
    return 0;
}

void crf1mm_dump(crf1mm_t* crf1mm, FILE *fp)
{
    int j;
    uint32_t i;
    feature_refs_t refs;
    const header_t* hfile = crf1mm->header;

    /* Dump the file header. */
    fprintf(fp, "FILEHEADER = {\n");
    fprintf(fp, "  magic: %c%c%c%c\n",
        hfile->magic[0], hfile->magic[1], hfile->magic[2], hfile->magic[3]);
    fprintf(fp, "  size: %d\n", hfile->size);
    fprintf(fp, "  type: %c%c%c%c\n",
        hfile->type[0], hfile->type[1], hfile->type[2], hfile->type[3]);
    fprintf(fp, "  version: %d\n", hfile->version);
    fprintf(fp, "  num_features: %d\n", hfile->num_features);
    fprintf(fp, "  num_labels: %d\n", hfile->num_labels);
    fprintf(fp, "  num_attrs: %d\n", hfile->num_attrs);
    fprintf(fp, "  off_features: 0x%X\n", hfile->off_features);
    fprintf(fp, "  off_labels: 0x%X\n", hfile->off_labels);
    fprintf(fp, "  off_attrs: 0x%X\n", hfile->off_attrs);
    fprintf(fp, "  off_labelrefs: 0x%X\n", hfile->off_labelrefs);
    fprintf(fp, "  off_attrrefs: 0x%X\n", hfile->off_attrrefs);
    fprintf(fp, "}\n");
    fprintf(fp, "\n");

    /* Dump the labels. */
    fprintf(fp, "LABELS = {\n");
    for (i = 0;i < hfile->num_labels;++i) {
        const char *str = crf1mm_to_label(crf1mm, i);
#if 0
        int check = crf1mm_to_lid(crf1mm, str);
        if (i != check) {
            fprintf(fp, "WARNING: inconsistent label CQDB\n");
        }
#endif
        fprintf(fp, "  %5d: %s\n", i, str);
    }
    fprintf(fp, "}\n");
    fprintf(fp, "\n");

    /* Dump the attributes. */
    fprintf(fp, "ATTRIBUTES = {\n");
    for (i = 0;i < hfile->num_attrs;++i) {
        const char *str = crf1mm_to_attr(crf1mm, i);
#if 0
        int check = crf1mm_to_aid(crf1mm, str);
        if (i != check) {
            fprintf(fp, "WARNING: inconsistent attribute CQDB\n");
        }
#endif
        fprintf(fp, "  %5d: %s\n", i, str);
    }
    fprintf(fp, "}\n");
    fprintf(fp, "\n");

    /* Dump the transition features. */
    fprintf(fp, "TRANSITIONS = {\n");
    for (i = 0;i < hfile->num_labels;++i) {
        crf1mm_get_labelref(crf1mm, i, &refs);
        for (j = 0;j < refs.num_features;++j) {
            crf1mm_feature_t f;
            int fid = crf1mm_get_featureid(&refs, j);
            const char *from = NULL, *to = NULL;

            crf1mm_get_feature(crf1mm, fid, &f);
            from = crf1mm_to_label(crf1mm, f.src);
            to = crf1mm_to_label(crf1mm, f.dst);
            fprintf(fp, "  (%d) %s --> %s: %f\n", f.type, from, to, f.weight);
        }
    }
    fprintf(fp, "}\n");
    fprintf(fp, "\n");

    /* Dump the transition features. */
    fprintf(fp, "TRANSITIONS_FROM_BOS = {\n");
    crf1mm_get_labelref(crf1mm, hfile->num_labels, &refs);
    for (j = 0;j < refs.num_features;++j) {
        crf1mm_feature_t f;
        int fid = crf1mm_get_featureid(&refs, j);
        const char *to = NULL;

        crf1mm_get_feature(crf1mm, fid, &f);
        to = crf1mm_to_label(crf1mm, f.dst);
        fprintf(fp, "  (%d) BOS --> %s: %f\n", f.type, to, f.weight);
    }
    fprintf(fp, "}\n");
    fprintf(fp, "\n");

    /* Dump the transition features. */
    fprintf(fp, "TRANSITIONS_TO_EOS = {\n");
    crf1mm_get_labelref(crf1mm, hfile->num_labels+1, &refs);
    for (j = 0;j < refs.num_features;++j) {
        crf1mm_feature_t f;
        int fid = crf1mm_get_featureid(&refs, j);
        const char *from = NULL;

        crf1mm_get_feature(crf1mm, fid, &f);
        from = crf1mm_to_label(crf1mm, f.src);
        fprintf(fp, "  (%d) %s --> EOS: %f\n", f.type, from, f.weight);
    }
    fprintf(fp, "}\n");
    fprintf(fp, "\n");

    /* Dump the transition features. */
    fprintf(fp, "STATE_FEATURES = {\n");
    for (i = 0;i < hfile->num_attrs;++i) {
        crf1mm_get_attrref(crf1mm, i, &refs);
        for (j = 0;j < refs.num_features;++j) {
            crf1mm_feature_t f;
            int fid = crf1mm_get_featureid(&refs, j);
            const char *attr = NULL, *to = NULL;

            crf1mm_get_feature(crf1mm, fid, &f);
#if 0
            if (f.src != i) {
                fprintf(fp, "WARNING: an inconsistent attribute reference.\n");
            }
#endif
            attr = crf1mm_to_attr(crf1mm, f.src);
            to = crf1mm_to_label(crf1mm, f.dst);
            fprintf(fp, "  (%d) %s --> %s: %f\n", f.type, attr, to, f.weight);
        }
    }
    fprintf(fp, "}\n");
    fprintf(fp, "\n");
}