app.py

import os
import torch
import librosa
import numpy as np
import gradio as gr
from sonics import HFAudioClassifier

# Model configurations
MODEL_IDS = {
    "SpecTTTra-α (5s)": "awsaf49/sonics-spectttra-alpha-5s",
    "SpecTTTra-β (5s)": "awsaf49/sonics-spectttra-beta-5s",
    "SpecTTTra-γ (5s)": "awsaf49/sonics-spectttra-gamma-5s",
    "SpecTTTra-α (120s)": "awsaf49/sonics-spectttra-alpha-120s",
    "SpecTTTra-β (120s)": "awsaf49/sonics-spectttra-beta-120s",
    "SpecTTTra-γ (120s)": "awsaf49/sonics-spectttra-gamma-120s",
}

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_cache = {}

def load_model(model_name):
    """Load model if not already cached"""
    if model_name not in model_cache:
        model_id = MODEL_IDS[model_name]
        model = HFAudioClassifier.from_pretrained(model_id)
        model = model.to(device)
        model.eval()
        model_cache[model_name] = model
    return model_cache[model_name]

def process_audio(audio_path, model_name):
    """Process audio file and return prediction"""
    try:
        model = load_model(model_name)
        max_time = model.config.audio.max_time
        
        # Load and process audio
        audio, sr = librosa.load(audio_path, sr=16000)
        chunk_samples = int(max_time * sr)
        total_chunks = len(audio) // chunk_samples
        middle_chunk_idx = total_chunks // 2
        
        # Extract middle chunk
        start = middle_chunk_idx * chunk_samples
        end = start + chunk_samples
        chunk = audio[start:end]
        
        if len(chunk) < chunk_samples:
            chunk = np.pad(chunk, (0, chunk_samples - len(chunk)))
            
        # Get prediction
        with torch.no_grad():
            chunk = torch.from_numpy(chunk).float().to(device)
            pred = model(chunk.unsqueeze(0))
            prob = torch.sigmoid(pred).cpu().numpy()[0]
            
        return {"Real": 1 - prob, "Fake": prob}

    except Exception as e:
        return {"Error": str(e)}

def predict(audio_file, model_name):
    """Gradio interface function"""
    if audio_file is None:
        return {"Message": "Please upload an audio file"}
    return process_audio(audio_file, model_name)

# Create Gradio interface
with gr.Blocks() as demo:
    gr.HTML(
        """
        <div style="text-align: center; margin-bottom: 1rem;">
            <img src="https://i.postimg.cc/3Jx3yZ5b/real-vs-fake-sonics-w-logo.jpg" 
                 style="max-width: 300px; margin: 0 auto;">
            <h1>SONICS: Synthetic Or Not - Identifying Counterfeit Songs</h1>
            <h3>ICLR 2025 [Poster]</h3>
        </div>
        """
    )
    
    with gr.Row():
        with gr.Column():
            audio_input = gr.Audio(
                label="Upload Audio File",
                type="filepath"
            )
            model_dropdown = gr.Dropdown(
                choices=list(MODEL_IDS.keys()),
                value="SpecTTTra-γ (5s)",
                label="Select Model"
            )
            submit_btn = gr.Button("Analyze Audio")
        
        with gr.Column():
            output = gr.Label(
                label="Analysis Result",
                num_top_classes=2
            )
    
    gr.Markdown(
        """
        ### Resources
        - [📄 Paper](https://openreview.net/forum?id=PY7KSh29Z8)
        - [🎵 Dataset](https://huggingface.co/datasets/awsaf49/sonics)
        - [🔬 ArXiv](https://arxiv.org/abs/2408.14080)
        - [💻 GitHub](https://github.com/awsaf49/sonics)
        """
    )
    
    submit_btn.click(
        fn=predict,
        inputs=[audio_input, model_dropdown],
        outputs=[output]
    )

if __name__ == "__main__":
    demo.launch()