245 lines
9.3 KiB
C
245 lines
9.3 KiB
C
#include <stdio.h>
|
|
#include <stdlib.h>
|
|
#include "mtmn.h"
|
|
#include "pnet_model.h"
|
|
#include "rnet_model.h"
|
|
#include "onet_model.h"
|
|
#include "freertos/FreeRTOS.h"
|
|
|
|
static int first_time[3] = {0};
|
|
mtmn_net_t *pnet (dl_matrix3du_t *in)
|
|
{
|
|
#define PNET_LAYER_NUM (3)
|
|
static dl_matrix3d_t *filters[PNET_LAYER_NUM * 4] = {0};
|
|
char name[50];
|
|
|
|
// Get filters
|
|
if (0 == first_time[0])
|
|
{
|
|
first_time[0] = 1;
|
|
for (int i = 0; i < PNET_LAYER_NUM; i++)
|
|
{
|
|
sprintf(name, "pnet_l%d_dilate", i+1);
|
|
filters[i * 4 + 0] = (dl_matrix3d_t *)get_coeff_pnet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "pnet_l%d_depth", i+1);
|
|
filters[i * 4 + 1] = (dl_matrix3d_t *)get_coeff_pnet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "pnet_l%d_compress", i+1);
|
|
filters[i * 4 + 2] = (dl_matrix3d_t *)get_coeff_pnet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "pnet_l%d_bias", i+1);
|
|
filters[i * 4 + 3] = (dl_matrix3d_t *)get_coeff_pnet_model.getter_3d(name, NULL, 0);
|
|
}
|
|
}
|
|
dl_matrix3d_t **l1_f = &(filters[0]);
|
|
dl_matrix3d_t **l2_f = &(filters[4]);
|
|
dl_matrix3d_t **l3_f = &(filters[8]);
|
|
|
|
dl_matrix3d_t *l4_category_conv = (dl_matrix3d_t *)get_coeff_pnet_model.getter_3d("pnet_l4_category_weight", NULL, 0);
|
|
dl_matrix3d_t *l4_category_bias = (dl_matrix3d_t *)get_coeff_pnet_model.getter_3d("pnet_l4_category_bias", NULL, 0);
|
|
dl_matrix3d_t *l4_offset_conv = (dl_matrix3d_t *)get_coeff_pnet_model.getter_3d("pnet_l4_offset_weight", NULL, 0);
|
|
dl_matrix3d_t *l4_offset_bias = (dl_matrix3d_t *)get_coeff_pnet_model.getter_3d("pnet_l4_offset_bias", NULL, 0);
|
|
|
|
// Layer 1
|
|
dl_matrix3d_conv_config_t config;
|
|
config.stride_x = 2;
|
|
config.stride_y = 2;
|
|
config.padding = PADDING_VALID;
|
|
config.mode = DL_XTENSA_IMPL;
|
|
config.type = INPUT_UINT8;
|
|
dl_matrix3d_t *l1 = dl_matrix3d_mobilenet(in, l1_f[0], l1_f[1], l1_f[2], l1_f[3], NULL, &config);
|
|
|
|
// Layer 2
|
|
config.stride_x = 1;
|
|
config.stride_y = 1;
|
|
config.type = INPUT_FLOAT;
|
|
dl_matrix3d_t *l2 = dl_matrix3d_mobilenet(l1, l2_f[0], l2_f[1], l2_f[2], l2_f[3], NULL, &config);
|
|
|
|
// Layer 3
|
|
dl_matrix3d_t *l3 = dl_matrix3d_mobilenet(l2, l3_f[0], l3_f[1], l3_f[2], l3_f[3], NULL, &config);
|
|
|
|
mtmn_net_t *pnet_o = (mtmn_net_t *)calloc(1, sizeof(mtmn_net_t));
|
|
// Layer 4 category
|
|
pnet_o->category = dl_matrix3d_conv(l3, l4_category_conv, l4_category_bias, 1, 1, 0, 1);
|
|
dl_matrix3d_softmax(pnet_o->category);
|
|
|
|
// Layer 4 boxoffset
|
|
pnet_o->offset = dl_matrix3d_conv(l3, l4_offset_conv, l4_offset_bias, 1, 1, 0, 1);
|
|
|
|
dl_matrix3d_free(l1);
|
|
dl_matrix3d_free(l2);
|
|
dl_matrix3d_free(l3);
|
|
|
|
return pnet_o;
|
|
}
|
|
|
|
|
|
|
|
mtmn_net_t *rnet_with_score_verify(dl_matrix3du_t *in, float score_threshold)
|
|
{
|
|
#define RNET_LAYER_NUM (4)
|
|
static dl_matrix3d_t *filters[RNET_LAYER_NUM * 4];
|
|
char name[50];
|
|
|
|
// Get filters
|
|
if (0 == first_time[1])
|
|
{
|
|
first_time[1] = 1;
|
|
for (int i = 0; i < RNET_LAYER_NUM; i++)
|
|
{
|
|
sprintf(name, "rnet_l%d_dilate", i+1);
|
|
filters[i * 4 + 0] = (dl_matrix3d_t *)get_coeff_rnet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "rnet_l%d_depth", i+1);
|
|
filters[i * 4 + 1] = (dl_matrix3d_t *)get_coeff_rnet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "rnet_l%d_compress", i+1);
|
|
filters[i * 4 + 2] = (dl_matrix3d_t *)get_coeff_rnet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "rnet_l%d_bias", i+1);
|
|
filters[i * 4 + 3] = (dl_matrix3d_t *)get_coeff_rnet_model.getter_3d(name, NULL, 0);
|
|
}
|
|
}
|
|
|
|
dl_matrix3d_t **l1_f = &(filters[0]);
|
|
dl_matrix3d_t **l2_f = &(filters[4]);
|
|
dl_matrix3d_t **l3_f = &(filters[8]);
|
|
dl_matrix3d_t **l4_f = &(filters[12]);
|
|
|
|
dl_matrix3d_t *l5_category_fc = (dl_matrix3d_t *)get_coeff_rnet_model.getter_3d("rnet_l5_category_weight", NULL, 0);
|
|
dl_matrix3d_t *l5_category_bias = (dl_matrix3d_t *)get_coeff_rnet_model.getter_3d("rnet_l5_category_bias", NULL, 0);
|
|
dl_matrix3d_t *l5_offset_fc = (dl_matrix3d_t *)get_coeff_rnet_model.getter_3d("rnet_l5_offset_weight", NULL, 0);
|
|
dl_matrix3d_t *l5_offset_bias = (dl_matrix3d_t *)get_coeff_rnet_model.getter_3d("rnet_l5_offset_bias", NULL, 0);
|
|
|
|
// Layer 1
|
|
dl_matrix3d_conv_config_t config;
|
|
config.stride_x = 2;
|
|
config.stride_y = 2;
|
|
config.padding = PADDING_VALID;
|
|
config.mode = DL_XTENSA_IMPL;
|
|
config.type = INPUT_UINT8;
|
|
dl_matrix3d_t *l1 = dl_matrix3d_mobilenet(in, l1_f[0], l1_f[1], l1_f[2], l1_f[3], NULL, &config);
|
|
|
|
// Layer 2
|
|
config.type = INPUT_FLOAT;
|
|
dl_matrix3d_t *l2 = dl_matrix3d_mobilenet(l1, l2_f[0], l2_f[1], l2_f[2], l2_f[3], NULL, &config);
|
|
|
|
// Layer 3
|
|
config.stride_x = 1;
|
|
config.stride_y = 1;
|
|
dl_matrix3d_t *l3 = dl_matrix3d_mobilenet(l2, l3_f[0], l3_f[1], l3_f[2], l3_f[3], NULL, &config);
|
|
|
|
// Layer 4
|
|
dl_matrix3d_t *l4 = dl_matrix3d_mobilenet(l3, l4_f[0], l4_f[1], l4_f[2], l4_f[3], NULL, &config);
|
|
|
|
mtmn_net_t *rnet_o = (mtmn_net_t *)calloc(1, sizeof(mtmn_net_t));
|
|
|
|
// Layer 5 category
|
|
rnet_o->category = dl_matrix3d_fc(l4, l5_category_fc, l5_category_bias);
|
|
dl_matrix3d_softmax(rnet_o->category);
|
|
|
|
if (rnet_o->category->item[1] < score_threshold)
|
|
{
|
|
dl_matrix3d_free(rnet_o->category);
|
|
free(rnet_o);
|
|
rnet_o = NULL;
|
|
goto rnet_exit;
|
|
}
|
|
// Layer 5 boxoffset
|
|
rnet_o->offset = dl_matrix3d_fc(l4, l5_offset_fc, l5_offset_bias);
|
|
|
|
rnet_exit:
|
|
dl_matrix3d_free(l1);
|
|
dl_matrix3d_free(l2);
|
|
dl_matrix3d_free(l3);
|
|
dl_matrix3d_free(l4);
|
|
|
|
return rnet_o;
|
|
}
|
|
|
|
mtmn_net_t *onet_with_score_verify(dl_matrix3du_t *in, float score_threshold)
|
|
{
|
|
#define ONET_LAYER_NUM (5)
|
|
#define ONET_COEF_PER_LAYER (5)
|
|
dl_matrix3d_t *filters[ONET_LAYER_NUM * ONET_COEF_PER_LAYER];
|
|
char name[50];
|
|
|
|
// Get filters
|
|
for (int i = 0; i < ONET_LAYER_NUM; i++)
|
|
{
|
|
sprintf(name, "onet_l%d_dilate", i+1);
|
|
filters[i * ONET_COEF_PER_LAYER + 0] = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "onet_l%d_depth", i+1);
|
|
filters[i * ONET_COEF_PER_LAYER + 1] = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "onet_l%d_compress", i+1);
|
|
filters[i * ONET_COEF_PER_LAYER + 2] = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "onet_l%d_bias", i+1);
|
|
filters[i * ONET_COEF_PER_LAYER + 3] = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d(name, NULL, 0);
|
|
sprintf(name, "onet_l%d_prelu_alpha", i+1);
|
|
filters[i * ONET_COEF_PER_LAYER + 4] = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d(name, NULL, 0);
|
|
}
|
|
|
|
dl_matrix3d_t **l1_f = &(filters[ONET_COEF_PER_LAYER * 0]);
|
|
dl_matrix3d_t **l2_f = &(filters[ONET_COEF_PER_LAYER * 1]);
|
|
dl_matrix3d_t **l3_f = &(filters[ONET_COEF_PER_LAYER * 2]);
|
|
dl_matrix3d_t **l4_f = &(filters[ONET_COEF_PER_LAYER * 3]);
|
|
dl_matrix3d_t **l5_f = &(filters[ONET_COEF_PER_LAYER * 4]);
|
|
|
|
dl_matrix3d_t *l6_category_fc = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d("onet_l6_category_weight", NULL, 0);
|
|
dl_matrix3d_t *l6_category_bias = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d("onet_l6_category_bias", NULL, 0);
|
|
dl_matrix3d_t *l6_offset_fc = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d("onet_l6_offset_weight", NULL, 0);
|
|
dl_matrix3d_t *l6_offset_bias = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d("onet_l6_offset_bias", NULL, 0);
|
|
dl_matrix3d_t *l6_landmark_fc = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d("onet_l6_landmark_weight", NULL, 0);
|
|
dl_matrix3d_t *l6_landmark_bias = (dl_matrix3d_t *)get_coeff_onet_model.getter_3d("onet_l6_landmark_bias", NULL, 0);
|
|
|
|
// Layer 1
|
|
dl_matrix3d_conv_config_t config;
|
|
config.stride_x = 2;
|
|
config.stride_y = 2;
|
|
config.padding = PADDING_VALID;
|
|
config.mode = DL_C_IMPL;
|
|
config.type = INPUT_UINT8;
|
|
|
|
dl_matrix3d_t *l1 = dl_matrix3d_mobilenet(in, l1_f[0], l1_f[1], l1_f[2], l1_f[3], l1_f[4], &config);
|
|
|
|
// Layer 2
|
|
config.type = INPUT_FLOAT;
|
|
dl_matrix3d_t *l2 = dl_matrix3d_mobilenet(l1, l2_f[0], l2_f[1], l2_f[2], l2_f[3], l2_f[4], &config);
|
|
|
|
// Layer 3
|
|
dl_matrix3d_t *l3 = dl_matrix3d_mobilenet(l2, l3_f[0], l3_f[1], l3_f[2], l3_f[3], l3_f[4], &config);
|
|
|
|
// Layer 4
|
|
config.stride_x = 1;
|
|
config.stride_y = 1;
|
|
dl_matrix3d_t *l4 = dl_matrix3d_mobilenet(l3, l4_f[0], l4_f[1], l4_f[2], l4_f[3], l4_f[4], &config);
|
|
|
|
// Layer 5
|
|
dl_matrix3d_t *l5 = dl_matrix3d_mobilenet(l4, l5_f[0], l5_f[1], l5_f[2], l5_f[3], l5_f[4], &config);
|
|
|
|
mtmn_net_t *onet_o = (mtmn_net_t *)calloc(1, sizeof(mtmn_net_t));
|
|
|
|
// Layer 6 category
|
|
onet_o->category = dl_matrix3d_fc(l5, l6_category_fc, l6_category_bias);
|
|
dl_matrix3d_softmax(onet_o->category);
|
|
if (onet_o->category->item[1] < score_threshold)
|
|
{
|
|
dl_matrix3d_free(onet_o->category);
|
|
free(onet_o);
|
|
onet_o = NULL;
|
|
goto onet_exit;
|
|
}
|
|
|
|
// Layer 6 boxoffset
|
|
onet_o->offset = dl_matrix3d_fc(l5, l6_offset_fc, l6_offset_bias);
|
|
|
|
// Layer 6 landmark
|
|
onet_o->landmark = dl_matrix3d_fc(l5, l6_landmark_fc, l6_landmark_bias);
|
|
|
|
|
|
onet_exit:
|
|
dl_matrix3d_free(l1);
|
|
dl_matrix3d_free(l2);
|
|
dl_matrix3d_free(l3);
|
|
dl_matrix3d_free(l4);
|
|
dl_matrix3d_free(l5);
|
|
|
|
return onet_o;
|
|
}
|
|
|