Which label corresponds to which index in the ouput vector

step22 · May 25, 2022, 4:32pm

Hi,

@risk of asking the obvious:

I am training a simple classifier and use the generated tflite file in my own application.

What would be the simplest way to figure out which output index corresponds to which label.

I thought, the ei_classifier_inferencing_categories variable in the model_variables.h file of the exported c++ library would give me the right order … but seems like not …

Best Regards
step

louis · May 26, 2022, 3:43pm

Hello @step22,

From the ei_impulse_result_t you can get the label:

github.com

edgeimpulse/inferencing-sdk-cpp/blob/f0e62138e992e863c7447cc539f7e560c2190d7d/classifier/ei_classifier_types.h#L30

      
        
            * SOFTWARE.
            */
            
            
#ifndef _EDGE_IMPULSE_RUN_CLASSIFIER_TYPES_H_
            #define _EDGE_IMPULSE_RUN_CLASSIFIER_TYPES_H_
            
            
#include <stdint.h>
            #include "model-parameters/model_metadata.h"
            
            
typedef struct {
               const char *label;
               float value;
            } ei_impulse_result_classification_t;
            
            
typedef struct {
               const char *label;
               uint32_t x;
               uint32_t y;
               uint32_t width;
               uint32_t height;
               float value;

Regards,

Louis

step22 · May 27, 2022, 7:15am

Hi Louis,

thanks for your reply, but this actually does not answer my question.

my labels cannot be in the sdk, as they are somehow custom to my project.

I see this char * [] in model-parameters/model_variables.h which basically holds my cathegories:

const char* ei_classifier_inferencing_categories[] = { “cat1”, “cat2”, “cat3”, “cat4” };

but the mapping of the index in the output tensor of my network to the position of the label in the char*[] is kind of random.

I wonder how this is handled in your library …

Best Regards
step