tflite_runtime freezes while using this classifier outside Docker

[charon25]

My current project is quite similar to yours, but as I needed another way of getting and sending the audio, I chosed to pick parts of it, and not use it in a Docker container.

But as I use tflite_runtime in the same way as this project, it sometimes completely stops working on invoke() (and any subsequent method, for example if I try to recreate it with tflite.Intepreter()), and I can't even use Ctrl+C to stop the program, I need to kill it.

Do you have any idea why ? My test code is the following :

import tflite_runtime.interpreter as tflite
import glob
import os
from os import path
import librosa
import numpy as np
import sqlite3
import time
from datetime import datetime, timedelta

DB_FILE = os.getenv("DB_PATH", "/home/mendel/udp/audio_db.db")  # path and filename of database
WAV_PATH = os.getenv("WAV_PATH", "/home/mendel/udp/audio/") # folder with wav files
MODEL_FILE = os.getenv("MODEL_FILE", "/home/mendel/classifier/sound_edgetpu.tflite") # path and filename  with model file
LABEL_FILE = os.getenv("LABEL_FILE", "/home/mendel/classifier/labels.txt") #path and filename of associated model class names
AUTO_DELETE = os.getenv("AUTO_DELETE", "false") #if equal to true, files above the threshold will automatically be deleted 
ct = os.getenv("CERTAINTY_THRESHOLD", "50") # minimum value to consider prediction to be acceptable
if ct.isnumeric():
    CERTAINTY_THRESHOLD = int(ct)
else:
    CERTAINTY_THRESHOLD = 70

# Load labels from a file
sound_names = [line.rstrip('\n') for line in open(LABEL_FILE)]
print("Loaded {0} labels for model.".format(len(sound_names)))

# just extract the features
def extract_features_only(filename):
    features = np.empty((0,193))
    X, sample_rate = librosa.load(filename)
    stft = np.abs(librosa.stft(X))
    mfccs = np.mean(librosa.feature.mfcc(y=X, sr=sample_rate, n_mfcc=40).T,axis=0)
    chroma = np.mean(librosa.feature.chroma_stft(S=stft, sr=sample_rate).T,axis=0)
    mel = np.mean(librosa.feature.melspectrogram(X, sr=sample_rate).T,axis=0)
    contrast = np.mean(librosa.feature.spectral_contrast(S=stft, sr=sample_rate).T,axis=0)
    tonnetz = np.mean(librosa.feature.tonnetz(y=librosa.effects.harmonic(X), sr=sample_rate).T,axis=0)
    ext_features = np.hstack([mfccs,chroma,mel,contrast,tonnetz])
    features = np.vstack([features,ext_features])
    return features

try:
    interpreter = tflite.Interpreter(model_path=MODEL_FILE, experimental_delegates=[tflite.load_delegate('libedgetpu.so.1')])
except ValueError:
    print("Interpreter error. Is Edge TPU plugged in?")

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
interpreter.allocate_tensors()

FILE = "/home/mendel/udp_classifier/audio/1622637563.wav"

n = 0
while True:
    predict_x = extract_features_only(FILE)
    interpreter.set_tensor(input_details[0]['index'], predict_x.astype(np.float32))
    start_time = time.time()
    interpreter.invoke()
    end_time = time.time()
    duration = str(round(1000*(end_time - start_time), 1))
    print("Interpreter duration: {} ms".format(duration))
    tflite_model_predictions = interpreter.get_tensor(output_details[0]['index'])
    # print(tflite_model_predictions[0])
    # get the indices of the top 2 predictions, invert into descending order
    ind = np.argpartition(tflite_model_predictions[0], -2)[-2:]
    ind[np.argsort(tflite_model_predictions[0][ind])]
    ind = ind[::-1]
    top_certainty = int(tflite_model_predictions[0,ind[0]] * 100)
    second_certainty = int(tflite_model_predictions[0,ind[1]] * 100)
    print("Top guess: ", sound_names[ind[0]], " (", top_certainty, "%)")
    print("2nd guess: ", sound_names[ind[1]], " (", second_certainty, "%)")
    n += 1
    print(n)
    time.sleep(0.5)

It is supposed to classify the same sound over and over, but it will always freeze after a few iterations (around 10 in my tests). For example, here it stopped after 10 :

(classifier_env) mendel@fun-rabbit:~/udp_classifier$ python test_tflite.py 
Loaded 10 labels for model.
Interpreter duration: 10.3 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
1
Interpreter duration: 1.5 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
2
Interpreter duration: 1.5 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
3
Interpreter duration: 1.4 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
4
Interpreter duration: 1.3 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
5
Interpreter duration: 1.3 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
6
Interpreter duration: 1.4 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
7
Interpreter duration: 1.4 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
8
Interpreter duration: 1.4 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
9
Interpreter duration: 1.4 ms
Top guess:  dog_bark  ( 50 %)
2nd guess:  gun_shot  ( 21 %)
10
^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^C^CTerminated

Am I missing something obvious or a library to make it work ? Or maybe it is something in the containers themselves ?

Thank you very much in advance!