ZeyadMostafa22
commited on
Commit
·
4dc47c8
1
Parent(s):
c629c7c
finall
Browse files
app.py
CHANGED
@@ -3,7 +3,6 @@ import torch
|
|
3 |
import torchaudio
|
4 |
import numpy as np
|
5 |
from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
|
6 |
-
import torch.nn.functional as F
|
7 |
import torchaudio.transforms as T
|
8 |
|
9 |
MODEL_ID = "Zeyadd-Mostaffa/wav2vec_checkpoints"
|
@@ -19,13 +18,15 @@ model.to(device)
|
|
19 |
|
20 |
label_names = ["fake", "real"] # According to your label2id = {"fake": 0, "real": 1}
|
21 |
|
|
|
22 |
def classify_audio(audio_file):
|
23 |
"""
|
24 |
audio_file: path to the uploaded file (WAV, MP3, etc.)
|
25 |
-
Returns:
|
26 |
"""
|
27 |
|
28 |
# 2) Load the audio file
|
|
|
29 |
waveform, sr = torchaudio.load(audio_file)
|
30 |
|
31 |
# If stereo, pick one channel or average
|
@@ -39,13 +40,14 @@ def classify_audio(audio_file):
|
|
39 |
waveform = resampler(waveform)
|
40 |
sr = 16000
|
41 |
|
|
|
42 |
# 3) Preprocess with feature_extractor
|
43 |
inputs = feature_extractor(
|
44 |
waveform.numpy(),
|
45 |
sampling_rate=sr,
|
46 |
return_tensors="pt",
|
47 |
truncation=True,
|
48 |
-
max_length=int(16000
|
49 |
)
|
50 |
|
51 |
# Move everything to device
|
@@ -53,24 +55,20 @@ def classify_audio(audio_file):
|
|
53 |
|
54 |
with torch.no_grad():
|
55 |
logits = model(input_values).logits
|
|
|
56 |
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
# Get predicted label and confidence
|
61 |
-
confidence, pred_id = torch.max(probabilities, dim=-1)
|
62 |
-
predicted_label = label_names[pred_id.item()]
|
63 |
|
64 |
-
# 5) Return label and confidence percentage
|
65 |
-
return f"Prediction: {predicted_label}, Confidence: {confidence.item() * 100:.2f}%"
|
66 |
|
67 |
-
#
|
68 |
demo = gr.Interface(
|
69 |
fn=classify_audio,
|
70 |
-
inputs=gr.Audio(type="filepath"),
|
71 |
outputs="text",
|
72 |
title="Wav2Vec2 Deepfake Detection",
|
73 |
-
description="Upload an audio sample to check if it is fake or real
|
74 |
)
|
75 |
|
76 |
if __name__ == "__main__":
|
|
|
3 |
import torchaudio
|
4 |
import numpy as np
|
5 |
from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
|
|
|
6 |
import torchaudio.transforms as T
|
7 |
|
8 |
MODEL_ID = "Zeyadd-Mostaffa/wav2vec_checkpoints"
|
|
|
18 |
|
19 |
label_names = ["fake", "real"] # According to your label2id = {"fake": 0, "real": 1}
|
20 |
|
21 |
+
|
22 |
def classify_audio(audio_file):
|
23 |
"""
|
24 |
audio_file: path to the uploaded file (WAV, MP3, etc.)
|
25 |
+
Returns: "fake" or "real"
|
26 |
"""
|
27 |
|
28 |
# 2) Load the audio file
|
29 |
+
# torchaudio returns (waveform, sample_rate)
|
30 |
waveform, sr = torchaudio.load(audio_file)
|
31 |
|
32 |
# If stereo, pick one channel or average
|
|
|
40 |
waveform = resampler(waveform)
|
41 |
sr = 16000
|
42 |
|
43 |
+
|
44 |
# 3) Preprocess with feature_extractor
|
45 |
inputs = feature_extractor(
|
46 |
waveform.numpy(),
|
47 |
sampling_rate=sr,
|
48 |
return_tensors="pt",
|
49 |
truncation=True,
|
50 |
+
max_length=int(16000* 6.0), # 6 second max
|
51 |
)
|
52 |
|
53 |
# Move everything to device
|
|
|
55 |
|
56 |
with torch.no_grad():
|
57 |
logits = model(input_values).logits
|
58 |
+
pred_id = torch.argmax(logits, dim=-1).item()
|
59 |
|
60 |
+
# 4) Return label text
|
61 |
+
predicted_label = label_names[pred_id]
|
62 |
+
return predicted_label
|
|
|
|
|
|
|
63 |
|
|
|
|
|
64 |
|
65 |
+
# 5) Build Gradio interface
|
66 |
demo = gr.Interface(
|
67 |
fn=classify_audio,
|
68 |
+
inputs=gr.Audio( type="filepath"),
|
69 |
outputs="text",
|
70 |
title="Wav2Vec2 Deepfake Detection",
|
71 |
+
description="Upload an audio sample to check if it is fake or real."
|
72 |
)
|
73 |
|
74 |
if __name__ == "__main__":
|