tommy/libpostal
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[optimization] for the FTRL and SGD optimizers, use the new *_array_sum_sq function to do L2 regularization, vs. the L2 norm which will use the linear algebra meaning

Browse Source
This commit is contained in:
Al
2018-01-22 01:56:32 -05:00
parent eb3fb37ad4
commit ccce4f793f
3 changed files with 10 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/ftrl.c
View File

@@ -208,7 +208,7 @@ double ftrl_reg_cost(ftrl_trainer_t *self, double_matrix_t *theta, uint32_array
if (row_idx >= i_start) { if (row_idx >= i_start) {
double *theta_i = double_matrix_get_row(theta, i); double *theta_i = double_matrix_get_row(theta, i);
l2_cost += double_array_l2_norm(theta_i, n); l2_cost += double_array_sum_sq(theta_i, n);
l1_cost += double_array_l1_norm(theta_i, n); l1_cost += double_array_l1_norm(theta_i, n);
} }
} }

2
src/stochastic_gradient_descent.c
View File

@@ -295,7 +295,7 @@ double stochastic_gradient_descent_reg_cost(sgd_trainer_t *self, uint32_array *u
double *theta_i = double_matrix_get_row(theta, row); double *theta_i = double_matrix_get_row(theta, row);
if (reg_type == REGULARIZATION_L2 && row >= i_start) { if (reg_type == REGULARIZATION_L2 && row >= i_start) {
cost += double_array_l2_norm(theta_i, n); cost += double_array_sum_sq(theta_i, n);
} else if (reg_type == REGULARIZATION_L1 && row >= i_start) { } else if (reg_type == REGULARIZATION_L1 && row >= i_start) {
cost += double_array_l1_norm(theta_i, n); cost += double_array_l1_norm(theta_i, n);
} }

8
src/vector_math.h
View File

@@ -210,6 +210,14 @@
return sqrt((double)result); \ return sqrt((double)result); \
} \ } \
\ \
static inline unsigned_type name##_sum_sq(type *array, size_t n) { \
unsigned_type result = 0; \
for (size_t i = 0; i < n; i++) { \
result += array[i] * array[i]; \
} \
return result; \
} \
\
static inline double name##_mean(type *array, size_t n) { \ static inline double name##_mean(type *array, size_t n) { \
unsigned_type sum = name##_sum(array, n); \ unsigned_type sum = name##_sum(array, n); \
return (double)sum / n; \ return (double)sum / n; \