Update app.py
Browse files
app.py
CHANGED
|
@@ -2,41 +2,61 @@ import gradio as gr
|
|
| 2 |
import tensorflow as tf
|
| 3 |
import librosa
|
| 4 |
import numpy as np
|
|
|
|
| 5 |
|
| 6 |
# Diccionario de etiquetas
|
| 7 |
labels = ['down', 'go', 'left', 'no', 'off', 'on', 'right', 'stop', 'up', 'yes']
|
| 8 |
|
| 9 |
def extract_features(file_name):
|
| 10 |
try:
|
| 11 |
-
audio
|
| 12 |
-
|
| 13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
except Exception as e:
|
| 16 |
print(f"Error encountered while parsing file: {file_name}")
|
| 17 |
-
print(e)
|
| 18 |
return None
|
| 19 |
-
|
| 20 |
-
return
|
| 21 |
|
| 22 |
def classify_audio(audio_file):
|
| 23 |
-
print(f"Tipo de audio_file: {type(audio_file)}")
|
| 24 |
|
| 25 |
-
#
|
| 26 |
-
|
|
|
|
|
|
|
|
|
|
| 27 |
|
| 28 |
if features is None:
|
| 29 |
-
return "Error al procesar el audio"
|
| 30 |
-
|
| 31 |
-
|
|
|
|
| 32 |
|
| 33 |
-
# Carga el modelo
|
| 34 |
model = tf.keras.models.load_model('my_model.h5', compile=False)
|
| 35 |
|
| 36 |
with tf.device('/CPU:0'):
|
| 37 |
prediction = model.predict(features)
|
| 38 |
predicted_label_index = np.argmax(prediction)
|
| 39 |
-
|
| 40 |
predicted_label = labels[predicted_label_index]
|
| 41 |
return predicted_label
|
| 42 |
|
|
@@ -48,4 +68,4 @@ iface = gr.Interface(
|
|
| 48 |
description="Sube un archivo de audio para clasificarlo."
|
| 49 |
)
|
| 50 |
|
| 51 |
-
iface.launch()
|
|
|
|
| 2 |
import tensorflow as tf
|
| 3 |
import librosa
|
| 4 |
import numpy as np
|
| 5 |
+
import tempfile
|
| 6 |
|
| 7 |
# Diccionario de etiquetas
|
| 8 |
labels = ['down', 'go', 'left', 'no', 'off', 'on', 'right', 'stop', 'up', 'yes']
|
| 9 |
|
| 10 |
def extract_features(file_name):
|
| 11 |
try:
|
| 12 |
+
# Carga el audio sin cambiar el sample rate
|
| 13 |
+
audio, sample_rate = librosa.load(file_name, sr=None)
|
| 14 |
+
|
| 15 |
+
# Saca el espectrograma de magnitud
|
| 16 |
+
spectrogram = np.abs(librosa.stft(audio, n_fft=512, hop_length=256))
|
| 17 |
+
|
| 18 |
+
# Convierte a escala logarítmica (como normalmente esperan los modelos de audio)
|
| 19 |
+
log_spectrogram = librosa.amplitude_to_db(spectrogram)
|
| 20 |
+
|
| 21 |
+
# Ajusta tamaño exacto
|
| 22 |
+
log_spectrogram = librosa.util.fix_length(log_spectrogram, size=257, axis=0)
|
| 23 |
+
log_spectrogram = librosa.util.fix_length(log_spectrogram, size=97, axis=1)
|
| 24 |
+
|
| 25 |
+
# Normaliza
|
| 26 |
+
log_spectrogram = (log_spectrogram - np.mean(log_spectrogram)) / np.std(log_spectrogram)
|
| 27 |
+
|
| 28 |
+
# Añade canal para la red convolucional
|
| 29 |
+
log_spectrogram = log_spectrogram[..., np.newaxis]
|
| 30 |
|
| 31 |
except Exception as e:
|
| 32 |
print(f"Error encountered while parsing file: {file_name}")
|
| 33 |
+
print(e)
|
| 34 |
return None
|
| 35 |
+
|
| 36 |
+
return log_spectrogram
|
| 37 |
|
| 38 |
def classify_audio(audio_file):
|
| 39 |
+
print(f"Tipo de audio_file: {type(audio_file)}")
|
| 40 |
|
| 41 |
+
# El tipo es string (ruta), no hace falta leer ni escribir en temp files
|
| 42 |
+
file_path = audio_file
|
| 43 |
+
|
| 44 |
+
# Extrae características
|
| 45 |
+
features = extract_features(file_path)
|
| 46 |
|
| 47 |
if features is None:
|
| 48 |
+
return "Error al procesar el audio"
|
| 49 |
+
|
| 50 |
+
# Añade batch dimension
|
| 51 |
+
features = features[np.newaxis, ...] # (1, 97, 257, 1)
|
| 52 |
|
| 53 |
+
# Carga el modelo en CPU
|
| 54 |
model = tf.keras.models.load_model('my_model.h5', compile=False)
|
| 55 |
|
| 56 |
with tf.device('/CPU:0'):
|
| 57 |
prediction = model.predict(features)
|
| 58 |
predicted_label_index = np.argmax(prediction)
|
| 59 |
+
|
| 60 |
predicted_label = labels[predicted_label_index]
|
| 61 |
return predicted_label
|
| 62 |
|
|
|
|
| 68 |
description="Sube un archivo de audio para clasificarlo."
|
| 69 |
)
|
| 70 |
|
| 71 |
+
iface.launch()
|