Benchmark Validation in App does not match Arduino Uno R4 WiFi
Hello.
Forgive me for what might be an incredibly "noob" question, but i'm not ashamed to admit that I am, in fact, a noob.
I've been trying and failing and trying and moderately succeeding over the last few days using NanoEdge AI Studio creating an audio classification model.
I'm using an Arduino Uno R4 Wifi. I also have a MAX4466 connected to A0, GND, and 3.3v
My entire process up until now.
1. Get a sketch running on my arduino which simply does a micValue = analogRead(A0) in a loop for 1024 times. For each read, i so a Serial.print(micValue); Serial.print(" "). After 1024 of those, I do a Serial.println("") to go to the next line.
2. With that running on the Arduino, I opened NanoEdge Studio, created my project, recorded a bunch of signals of background_noise, silence, me speaking and another sound that i want to classify. It's not a HUGE dataset, only a couple of minutes, but i'm increasingly adding more to each Class.
3. I ran the benchmark.
4. The benchmark produced some results in the 90+% range.
5. I performed the validation step with me speaking, some background noise, silence, and the beep and it produced a mostly accurate prediction based on what sound i was currently forcing. When i wasn't speaking, SILENCE was the most probable. When I was speaking, SPEECH was the most probably, when i did some shuffling and other stuff background_noise was the most probable and when i played the other sound, that was the most probable.
6. I "deployed" the library (and saved it to my computer)
I added the library in Arduino and can include NanoEdgeAI.h and knowledge.h fine.
I copied this code from the example file.
float input_user_buffer[DATA_INPUT_USER * AXIS_NUMBER]; // Buffer of input values
float output_class_buffer[CLASS_NUMBER]; // Buffer of class probabilities
uint16_t id_class = 0;
const char *id2class[CLASS_NUMBER + 1] = { // Buffer for mapping class id to class name
"unknown",
"speech",
"silence",
"braun_infusomat",
"background_noise",
}; get_microphone_data();
neai_classification(input_user_buffer, output_class_buffer, &id_class);
Serial.print(id_class);
Serial.print(": ");
Serial.println(id2class[id_class]);
for(int x =0; x <= CLASS_NUMBER; x++)
{
Serial.println( output_class_buffer[x]);
}
delay(100);However, the results are nowhere near the accuracy of the test performed on the validation tab in NanoEdgeAI Studio.
Any thoughts on what i might be doing wrong?