app.py

import torch
import torchaudio
import gradio as gr
from pathlib import Path
from preprocess import PreprocessingPipeline

from model_arch import ResNetModel
 


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

# Load model
model = ResNetModel(ch_in=1, num_classes=5)  
model.load_state_dict(torch.load("resnet3.pth", map_location=device))
model.to(device)
model.eval()

pipeline = PreprocessingPipeline(save_dir="tmp", max_duration=4, sr_out=22050)


BIRD_CLASSES = ['short_chirp', 'sweep_down', 'sweep_up', 'warble', 'whistle_click']  # adjust

def predict_bird(audio_file):
    # Preprocess
    signal, sr = pipeline.loader.load(audio_file)
    signal = pipeline.stereo2mono.stereo_to_mono(signal)
    signal, sr = pipeline.resampler.resample(signal, sr, pipeline.sr_out, debug=False)
    signal = pipeline.truncate_pad.truncate_or_pad(signal, debug=False)
    
    # Extract log-mel spectrogram
    log_mel = pipeline.fe.log_mel_scale(signal)       # [1, n_mels, T]
    log_mel = log_mel.unsqueeze(0).to(device)        # add batch dim

    # Forward pass
    with torch.no_grad():
        output = model(log_mel)
        probs = torch.softmax(output, dim=1).cpu().numpy()[0]
    
    # Build result dictionary
    result = {cls: float(prob) for cls, prob in zip(BIRD_CLASSES, probs)}
    predicted_class = BIRD_CLASSES[probs.argmax()]
    
    return predicted_class, result

iface = gr.Interface(
    fn=predict_bird,
    inputs=gr.Audio(type="filepath"),
    outputs=[gr.Textbox(label="Predicted Bird"), gr.Label(num_top_classes=5)],
    title="Bird Call Classifier",
    description="Upload a bird call .wav file and get the predicted bird species."
)

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