Question/Issue:
[Describe the question or issue in detail]
For my project, I downloaded the source code of the mfcc processing block and modified the ‘dsp.py’ and ‘parameters.json’ files to apply a digital filter and normalization to the raw data before inputing it to the mfcc extraction process.
I was able to use the custom block to build and download my model. However, when I want to compile a sketch using my model in arduino, I get the error:
undefined reference to extract_mfcc_custom_features(ei::ei_signal_t*, ei::ei_matrix*, void*, float)'collect2.exe: error: ld returned 1 exit status
While trying to figure out the issue, I built another version of the same model, but with the original mfcc processing block and downloaded it to see what was different in the arduino library files. What I found out, is that ‘extract_mfcc_custom_features’ is only used in the ‘model_variables.h’ file of my custom model, where it is referred as it follows:
int extract_mfcc_custom_features(signal_t *signal, matrix_t *output_matrix, void *config_ptr, const float frequency);
const uint8_t ei_dsp_blocks_size = 1;
ei_model_dsp_t ei_dsp_blocks[ei_dsp_blocks_size] = {
{ // DSP block 33
33,
560, // output size
&extract_mfcc_custom_features, // DSP function pointer
(void*)&ei_dsp_config_33, // pointer to config struct
ei_dsp_config_33_axes, // array of offsets into the input stream, one for each axis
ei_dsp_config_33_axes_size, // number of axes
1, // version
nullptr, // factory function
}
};
And in the ‘model_variables.h’ file of the model using the original mfcc block it is defined as:
const uint8_t ei_dsp_blocks_size = 1;
ei_model_dsp_t ei_dsp_blocks[ei_dsp_blocks_size] = {
{ // DSP block 33
33,
560, // output size
&extract_mfcc_features, // DSP function pointer
(void*)&ei_dsp_config_33, // pointer to config struct
ei_dsp_config_33_axes, // array of offsets into the input stream, one for each axis
ei_dsp_config_33_axes_size, // number of axes
1, // version
nullptr, // factory function
}
};
Thus, I tried renaming ‘extract_mfcc_custom_features’ to ‘extract_mfcc_features’ and erasing the lines
int extract_mfcc_custom_features(signal_t *signal, matrix_t *output_matrix, void *config_ptr, const float frequency);
since these weren’t in the file of the model using the original mfcc block.
After that, the arduino sketch compiled successfully, but now the inference is not running, as displayed in my serial monitor:
Collecting data…
Running inference…
ERR: MFCC failed (-1002)
ERR: Failed to run DSP process (-1002)
ERR: Failed to run classifier (-5)
Does anyone know how to solve this discrepancy that happens when using a model with a custom block?
The code of my arduino sketch for using the model, which worked fine with the library of the model using the original mfcc block, is the following, which collects raw sensor data using a piezoelectric sensor connected to the esp32 every 2 seconds and inputs it to the model:
// If your target is limited in memory remove this macro to save 10K RAM
#define EIDSP_QUANTIZE_FILTERBANK 0
#include <ArticulationRecog_inferencing.h> // Edge Impulse model
#define SAMPLE_RATE 1000 // 1 kHz sampling rate
#define SAMPLE_DURATION 2 // 2 seconds
#define NUM_SAMPLES (SAMPLE_RATE * SAMPLE_DURATION)
#define SENSOR_PIN 32 // Analog input pin
static float input_data[NUM_SAMPLES]; // Buffer for sensor data
static bool debug_nn = false; // Debug flag
void setup() {
Serial.begin(115200);
analogReadResolution(12); // Use 12-bit ADC resolution (0-4095)
ei_printf("Edge Impulse Inferencing Demo\n");
ei_printf("Sampling at %d Hz, collecting %d samples (%d seconds)\n", SAMPLE_RATE, NUM_SAMPLES, SAMPLE_DURATION);
run_classifier_init(); // Initialize Edge Impulse classifier
ei_printf("\nStarting continuous inference in 2 seconds...\n");
delay(2000);
}
void loop() {
Serial.println(“Collecting data…”);
// Collect 2 seconds of sensor data at 1 kHz
for (int i = 0; i < NUM_SAMPLES; i++) {
input_data[i] = (float)analogRead(SENSOR_PIN) / 4095.0; // Normalize input
delayMicroseconds(1000); // 1 ms delay for 1 kHz sampling
}
Serial.println("Running inference...");
// Run inference
signal_t signal;
signal.total_length = NUM_SAMPLES;
signal.get_data = &raw_adc_signal_get_data;
ei_impulse_result_t result = {0};
EI_IMPULSE_ERROR res = run_classifier(&signal, &result, debug_nn);
if (res != EI_IMPULSE_OK) {
ei_printf("ERR: Failed to run classifier (%d)\n", res);
return;
}
// Print predictions
ei_printf("Predictions:\n");
for (size_t i = 0; i < EI_CLASSIFIER_LABEL_COUNT; i++) {
ei_printf(" %s: ", result.classification[i].label);
ei_printf_float(result.classification[i].value);
ei_printf("\n");
}
#if EI_CLASSIFIER_HAS_ANOMALY == 1
ei_printf(" Anomaly score: “);
ei_printf_float(result.anomaly);
ei_printf(”\n");
#endif
delay(500); // Short delay before next inference cycle
}
/**
- Get raw ADC signal data for Edge Impulse
*/
static int raw_adc_signal_get_data(size_t offset, size_t length, float *out_ptr) {
memcpy(out_ptr, &input_data[offset], length * sizeof(float));
return 0;
}
Insufficient available memory on your device.
Failed memory allocation (