Spaces:
Sleeping
Sleeping
Commit
路
76e83c1
1
Parent(s):
18b8d3c
zmiana
Browse files- app.py +19 -21
- requirements.txt +2 -1
app.py
CHANGED
|
@@ -4,8 +4,10 @@ import librosa
|
|
| 4 |
import librosa.display
|
| 5 |
import tensorflow as tf
|
| 6 |
import matplotlib.pyplot as plt
|
|
|
|
|
|
|
| 7 |
|
| 8 |
-
# Parametry
|
| 9 |
SR = 22050
|
| 10 |
N_MELS = 128
|
| 11 |
TARGET_FRAMES = 216
|
|
@@ -16,13 +18,13 @@ model = tf.keras.models.load_model("model.h5")
|
|
| 16 |
|
| 17 |
def compute_melspectrogram(y, sr=SR, n_mels=N_MELS):
|
| 18 |
S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=n_mels)
|
| 19 |
-
|
| 20 |
-
return S_DB
|
| 21 |
|
| 22 |
def resize_spectrogram(S, target_frames=TARGET_FRAMES):
|
| 23 |
if S.shape[1] < target_frames:
|
| 24 |
pad = target_frames - S.shape[1]
|
| 25 |
-
left = pad // 2
|
|
|
|
| 26 |
S = np.pad(S, ((0, 0), (left, right)), mode='constant')
|
| 27 |
elif S.shape[1] > target_frames:
|
| 28 |
start = (S.shape[1] - target_frames) // 2
|
|
@@ -30,46 +32,42 @@ def resize_spectrogram(S, target_frames=TARGET_FRAMES):
|
|
| 30 |
return S
|
| 31 |
|
| 32 |
def predict_and_plot(audio_path):
|
| 33 |
-
# Wczytaj
|
| 34 |
y, _ = librosa.load(audio_path, sr=SR)
|
| 35 |
|
| 36 |
-
# Oblicz spektrogram
|
| 37 |
S_full = compute_melspectrogram(y)
|
| 38 |
S = resize_spectrogram(S_full)
|
| 39 |
|
| 40 |
# Przygotuj do predykcji
|
| 41 |
x = S[np.newaxis, ..., np.newaxis]
|
| 42 |
-
|
| 43 |
|
| 44 |
-
#
|
| 45 |
fig, ax = plt.subplots(figsize=(8, 4))
|
| 46 |
librosa.display.specshow(S_full, sr=SR, x_axis='time', y_axis='mel', cmap='magma', ax=ax)
|
| 47 |
ax.set_title("Mel-spektrogram")
|
| 48 |
-
ax.set_xlabel("Czas")
|
| 49 |
-
ax.set_ylabel("Cz臋stotliwo艣膰 (Mel)")
|
| 50 |
plt.tight_layout()
|
| 51 |
-
|
| 52 |
-
# Zapisz obrazek do zmiennej
|
| 53 |
-
import io
|
| 54 |
buf = io.BytesIO()
|
| 55 |
-
|
| 56 |
plt.close(fig)
|
| 57 |
buf.seek(0)
|
|
|
|
| 58 |
|
| 59 |
-
#
|
| 60 |
-
|
| 61 |
-
return
|
| 62 |
|
| 63 |
-
# Gradio UI
|
| 64 |
demo = gr.Interface(
|
| 65 |
fn=predict_and_plot,
|
| 66 |
inputs=gr.Audio(type="filepath", label="Wgraj plik WAV"),
|
| 67 |
outputs=[
|
| 68 |
-
gr.Label(num_top_classes=5, label="
|
| 69 |
gr.Image(label="Spektrogram")
|
| 70 |
],
|
| 71 |
-
title="Rozpoznawanie instrument贸w
|
| 72 |
-
description="Model
|
| 73 |
)
|
| 74 |
|
| 75 |
if __name__ == "__main__":
|
|
|
|
| 4 |
import librosa.display
|
| 5 |
import tensorflow as tf
|
| 6 |
import matplotlib.pyplot as plt
|
| 7 |
+
from PIL import Image
|
| 8 |
+
import io
|
| 9 |
|
| 10 |
+
# Parametry
|
| 11 |
SR = 22050
|
| 12 |
N_MELS = 128
|
| 13 |
TARGET_FRAMES = 216
|
|
|
|
| 18 |
|
| 19 |
def compute_melspectrogram(y, sr=SR, n_mels=N_MELS):
|
| 20 |
S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=n_mels)
|
| 21 |
+
return librosa.power_to_db(S, ref=np.max)
|
|
|
|
| 22 |
|
| 23 |
def resize_spectrogram(S, target_frames=TARGET_FRAMES):
|
| 24 |
if S.shape[1] < target_frames:
|
| 25 |
pad = target_frames - S.shape[1]
|
| 26 |
+
left = pad // 2
|
| 27 |
+
right = pad - left
|
| 28 |
S = np.pad(S, ((0, 0), (left, right)), mode='constant')
|
| 29 |
elif S.shape[1] > target_frames:
|
| 30 |
start = (S.shape[1] - target_frames) // 2
|
|
|
|
| 32 |
return S
|
| 33 |
|
| 34 |
def predict_and_plot(audio_path):
|
| 35 |
+
# Wczytaj d藕wi臋k
|
| 36 |
y, _ = librosa.load(audio_path, sr=SR)
|
| 37 |
|
| 38 |
+
# Oblicz i przeskaluj spektrogram
|
| 39 |
S_full = compute_melspectrogram(y)
|
| 40 |
S = resize_spectrogram(S_full)
|
| 41 |
|
| 42 |
# Przygotuj do predykcji
|
| 43 |
x = S[np.newaxis, ..., np.newaxis]
|
| 44 |
+
preds = model.predict(x, verbose=0)[0]
|
| 45 |
|
| 46 |
+
# Rysuj spektrogram i zapisz do obrazu
|
| 47 |
fig, ax = plt.subplots(figsize=(8, 4))
|
| 48 |
librosa.display.specshow(S_full, sr=SR, x_axis='time', y_axis='mel', cmap='magma', ax=ax)
|
| 49 |
ax.set_title("Mel-spektrogram")
|
|
|
|
|
|
|
| 50 |
plt.tight_layout()
|
| 51 |
+
|
|
|
|
|
|
|
| 52 |
buf = io.BytesIO()
|
| 53 |
+
fig.savefig(buf, format='png')
|
| 54 |
plt.close(fig)
|
| 55 |
buf.seek(0)
|
| 56 |
+
image = Image.open(buf)
|
| 57 |
|
| 58 |
+
# Predykcje jako s艂ownik
|
| 59 |
+
pred_dict = {label: float(p) for label, p in zip(LABELS, preds)}
|
| 60 |
+
return pred_dict, image
|
| 61 |
|
|
|
|
| 62 |
demo = gr.Interface(
|
| 63 |
fn=predict_and_plot,
|
| 64 |
inputs=gr.Audio(type="filepath", label="Wgraj plik WAV"),
|
| 65 |
outputs=[
|
| 66 |
+
gr.Label(num_top_classes=5, label="Predykcja"),
|
| 67 |
gr.Image(label="Spektrogram")
|
| 68 |
],
|
| 69 |
+
title="Rozpoznawanie instrument贸w",
|
| 70 |
+
description="Model klasyfikuje d藕wi臋ki do jednej z klas instrument贸w."
|
| 71 |
)
|
| 72 |
|
| 73 |
if __name__ == "__main__":
|
requirements.txt
CHANGED
|
@@ -2,4 +2,5 @@ tensorflow
|
|
| 2 |
librosa
|
| 3 |
gradio
|
| 4 |
numpy
|
| 5 |
-
matplotlib
|
|
|
|
|
|
| 2 |
librosa
|
| 3 |
gradio
|
| 4 |
numpy
|
| 5 |
+
matplotlib
|
| 6 |
+
Pillow
|