[parsing] Averaged perceptron model data structure for storing the finalized, averaged, sparse weights

This commit is contained in:
Al
2015-09-08 12:02:15 -07:00
parent 8d642b45b9
commit c80d8b8067
3 changed files with 281 additions and 0 deletions

219
src/averaged_perceptron.c Normal file
View File

@@ -0,0 +1,219 @@
#include "averaged_perceptron.h"
#define PERCEPTRON_SIGNATURE 0xCBCBCBCB
static inline bool averaged_perceptron_get_feature_id(averaged_perceptron_t *self, char *feature, uint32_t *feature_id) {
return trie_get_data(self->features, feature, feature_id);
}
inline double_array *averaged_perceptron_predict_scores(averaged_perceptron_t *self, cstring_array *features) {
if (self->scores == NULL || self->scores->n == 0) self->scores = double_array_new_zeros((size_t)self->num_classes);
double_array_set(self->scores->a, self->scores->n, 0.0);
double *scores = self->scores->a;
uint32_t i = 0;
char *feature;
uint32_t feature_id;
uint32_t *indptr = self->weights->indptr->a;
uint32_t *indices = self->weights->indices->a;
double *data = self->weights->data->a;
cstring_array_foreach(features, i, feature, {
if (!averaged_perceptron_get_feature_id(self, feature, &feature_id)) {
continue;
}
for (int col = indptr[feature_id]; col < indptr[feature_id+1]; col++) {
uint32_t class_id = indices[col];
scores[class_id] += data[col];
}
})
return self->scores;
}
inline double_array *averaged_perceptron_predict_scores_counts(averaged_perceptron_t *self, khash_t(str_uint32) *feature_counts) {
if (self->scores == NULL || self->scores->n == 0) self->scores = double_array_new_zeros((size_t)self->num_classes);
double_array_set(self->scores->a, self->scores->n, 0.0);
double *scores = self->scores->a;
uint32_t i = 0;
const char *feature;
uint32_t count;
uint32_t feature_id;
uint32_t *indptr = self->weights->indptr->a;
uint32_t *indices = self->weights->indices->a;
double *data = self->weights->data->a;
kh_foreach(feature_counts, feature, count, {
if (!averaged_perceptron_get_feature_id(self, (char *)feature, &feature_id)) {
continue;
}
for (int col = indptr[feature_id]; col < indptr[feature_id + 1]; col++) {
uint32_t class_id = indices[col];
scores[class_id] += data[col] * (double)count;
}
})
return self->scores;
}
inline uint32_t averaged_perceptron_predict(averaged_perceptron_t *self, cstring_array *features) {
double_array *scores = averaged_perceptron_predict_scores(self, features);
int64_t max_score = double_array_argmax(scores->a, scores->n);
return (uint32_t)max_score;
}
inline uint32_t averaged_perceptron_predict_counts(averaged_perceptron_t *self, khash_t(str_uint32) *feature_counts) {
double_array *scores = averaged_perceptron_predict_scores_counts(self, feature_counts);
int64_t max_score = double_array_argmax(scores->a, scores->n);
return (uint32_t)max_score;
}
averaged_perceptron_t *averaged_perceptron_read(FILE *f) {
if (f == NULL) return NULL;
uint32_t signature;
if (!file_read_uint32(f, &signature) || signature != PERCEPTRON_SIGNATURE) {
return NULL;
}
averaged_perceptron_t *perceptron = malloc(sizeof(averaged_perceptron_t));
if (!file_read_uint32(f, &perceptron->num_features) ||
!file_read_uint32(f, &perceptron->num_classes) ||
perceptron->num_classes == 0) {
return NULL;
}
perceptron->weights = sparse_matrix_read(f);
if (perceptron->weights == NULL) {
goto exit_perceptron_created;
}
perceptron->scores = double_array_new_zeros((size_t)perceptron->num_classes);
uint64_t classes_str_len;
if (!file_read_uint64(f, &classes_str_len)) {
goto exit_perceptron_created;
}
char_array *array = char_array_new_size(classes_str_len);
if (array == NULL) {
goto exit_perceptron_created;
}
if (!file_read_chars(f, array->a, classes_str_len)) {
char_array_destroy(array);
goto exit_perceptron_created;
}
array->n = classes_str_len;
perceptron->classes = cstring_array_from_char_array(array);
if (perceptron->classes == NULL) {
goto exit_perceptron_created;
}
perceptron->features = trie_read(f);
if (perceptron->features == NULL) {
goto exit_perceptron_created;
}
return perceptron;
exit_perceptron_created:
averaged_perceptron_destroy(perceptron);
return NULL;
}
averaged_perceptron_t *averaged_perceptron_load(char *filename) {
if (filename == NULL) return NULL;
FILE *f = fopen(filename, "rb");
if (f == NULL) return NULL;
averaged_perceptron_t *perceptron = averaged_perceptron_read(f);
fclose(f);
return perceptron;
}
bool averaged_perceptron_write(averaged_perceptron_t *self, FILE *f) {
if (self == NULL || f == NULL || self->weights == NULL || self->classes == NULL ||
self->features == NULL) {
return false;
}
if (!file_write_uint32(f, PERCEPTRON_SIGNATURE) ||
!file_write_uint32(f, self->num_features) ||
!file_write_uint32(f, self->num_classes)) {
return false;
}
if (!sparse_matrix_write(self->weights, f)) {
return false;
}
uint64_t classes_str_len = (uint64_t) cstring_array_used(self->classes);
if (!file_write_uint64(f, classes_str_len)) {
return false;
}
if (!file_write_chars(f, self->classes->str->a, classes_str_len)) {
return false;
}
if (!trie_write(self->features, f)) {
return false;
}
return true;
}
bool averaged_perceptron_save(averaged_perceptron_t *self, char *filename) {
if (self == NULL || filename == NULL) return false;
FILE *f = fopen(filename, "wb");
if (f == NULL) return false;
bool ret_val = averaged_perceptron_write(self, f);
fclose(f);
return ret_val;
}
void averaged_perceptron_destroy(averaged_perceptron_t *self) {
if (self == NULL) return;
if (self->features != NULL) {
trie_destroy(self->features);
}
if (self->classes != NULL) {
cstring_array_destroy(self->classes);
}
if (self->weights != NULL) {
sparse_matrix_destroy(self->weights);
}
if (self->scores != NULL) {
double_array_destroy(self->scores);
}
free(self);
}

61
src/averaged_perceptron.h Normal file
View File

@@ -0,0 +1,61 @@
/*
averaged_perceptron.h
---------------------
The averaged perceptron is a simple, efficient and effective method for
training sequence models.
The averaged perceptron is a linear model, meaning the score for a given class
is the dot product of weights and the feature values.
This implementation of the averaged perceptron uses a trie data structure to
store the mapping from features to ids, which can be quite memory efficient
as opposed to a hash table and allows us to store
The weights are stored as a sparse matrix in compressed sparse row format
(see sparse_matrix.h)
See [Collins, 2002] Discriminative Training Methods for Hidden Markov Models:
Theory and Experiments with Perceptron Algorithms
Paper: http://www.cs.columbia.edu/~mcollins/papers/tagperc.pdf
*/
#ifndef AVERAGED_PERCEPTRON_H
#define AVERAGED_PERCEPTRON_H
#include <stdlib.h>
#include <stdint.h>
#include <stdbool.h>
#include "collections.h"
#include "sparse_matrix.h"
#include "trie.h"
typedef struct averaged_perceptron {
uint32_t num_features;
uint32_t num_classes;
trie_t *features;
cstring_array *classes;
sparse_matrix_t *weights;
double_array *scores;
} averaged_perceptron_t;
averaged_perceptron_t *averaged_perceptron_read(FILE *f);
averaged_perceptron_t *averaged_perceptron_load(char *filename);
uint32_t averaged_perceptron_predict(averaged_perceptron_t *self, cstring_array *features);
uint32_t averaged_perceptron_predict_counts(averaged_perceptron_t *self, khash_t(str_uint32) *feature_counts);
double_array *averaged_perceptron_predict_scores(averaged_perceptron_t *self, cstring_array *features);
double_array *averaged_perceptron_predict_scores_counts(averaged_perceptron_t *self, khash_t(str_uint32) *feature_counts);
bool averaged_perceptron_write(averaged_perceptron_t *self, FILE *f);
bool averaged_perceptron_save(averaged_perceptron_t *self, char *filename);
averaged_perceptron_t *averaged_perceptron_read(FILE *f);
averaged_perceptron_t *averaged_perceptron_load(char *filename);
void averaged_perceptron_destroy(averaged_perceptron_t *self);
#endif

View File

@@ -31,6 +31,7 @@
static inline name *name##_new_zeros(size_t n) { \
name *vector = name##_new_size(n); \
memset(vector->a, 0, n * sizeof(type)); \
vector->n = n; \
return name##_new_value(n, (type)0); \
} \
\