SpatialCLAP/simple_similarity.py

#!/usr/bin/env python3
import sys
import os
sys.path.append(os.path.dirname(os.path.abspath(__file__)))

from model import CLAPEncoder
import torch
import torchaudio

def main():
    # Load model
    model = CLAPEncoder()
    model.load_pretrained()
    model.eval()
    
    # Audio file path
    audio_path = "/home/karan/sda_link/GitHub/EMMA2/EMMA2_contextual/experiments/emma2_txt_cond_128_2targets/single_source_extraction/test_outputs/000/gt.wav"
    
    # Text descriptions
    texts = ["car horn", "baby crying", "child speaking", "gunshot", "car horn on the left", "car horn on the right", "car horn in front", "car horn in front-left", "car horn in front-right"]
    
    # Load audio
    waveform, sr = torchaudio.load(audio_path)
    if sr != 16000:
        resampler = torchaudio.transforms.Resample(orig_freq=sr, new_freq=16000)
        waveform = resampler(waveform)
    
    # Ensure stereo
    if waveform.shape[0] == 1:
        waveform = waveform.repeat(2, 1)
    elif waveform.shape[0] > 2:
        waveform = waveform[:2, :]
    
    # Add batch dimension
    audio = waveform.unsqueeze(0)
    
    # Get audio embedding
    with torch.no_grad():
        audio_embedding = model.embed_audio(audio)
    
    # Calculate similarities
    print("Similarity scores:")
    for text in texts:
        with torch.no_grad():
            text_embedding = model.embed_text([text])
        
        # Calculate cosine similarity
        audio_norm = torch.nn.functional.normalize(audio_embedding, dim=-1)
        text_norm = torch.nn.functional.normalize(text_embedding, dim=-1)
        similarity = torch.cosine_similarity(audio_norm, text_norm, dim=-1).item()
        
        print(f"'{text}': {similarity:.6f}")

if __name__ == "__main__":
    main()