app.py

import gradio as gr
import numpy as np
import librosa
from PIL import Image
import tempfile
import os
from app.preprocess import preprocess_audio
from app.model import predict
from collections import Counter, defaultdict


# Process Image Input
def process_image_input(img):
    # Classify a spectrogram image directly using model.predict
    label, confidence, probs = predict(img)
    return label, round(confidence, 3), probs


# Process Audio Input
def process_audio_input(audio_path):
    # audio_path = filepath from Gradio
    # Preprocess → mel-spectrogram → predict per chunk

    # Preprocess to mel-spectrogram chunk images
    imgs = preprocess_audio(audio_path)

    all_preds = []
    all_confs = []
    all_probs = []

    for img in imgs:
        label, conf, probs = predict(img)
        all_preds.append(label)
        all_confs.append(conf)
        all_probs.append(probs)

    # Majority Vote
    counter = Counter(all_preds)
    max_count = max(counter.values())
    candidates = [k for k, v in counter.items() if v == max_count]

    if len(candidates) == 1:
        final_label = candidates[0]
    else:
        conf_sums = defaultdict(float)
        for i, label in enumerate(all_preds):
            if label in candidates:
                conf_sums[label] += all_confs[i]
        final_label = max(conf_sums, key=conf_sums.get)

    final_conf = float(np.mean([all_confs[i] for i, l in enumerate(all_preds) if l == final_label]))

    return final_label, round(final_conf, 3), all_preds, [round(c, 3) for c in all_confs]


# Main prediction logic
def classify(audio_path, image):

    # If an image is provided → classify directly
    if image is not None:
        label, conf, probs = process_image_input(image)
        return {
            "Final Label": label,
            "Confidence": conf,
            "Details": probs
        }

    # If an audio file is provided → preprocess and classify
    if audio_path is not None:
        label, conf, all_preds, all_confs = process_audio_input(audio_path)

        return {
            "Final Label": label,
            "Confidence": conf,
            "All Chunk Labels": all_preds,
            "All Chunk Confidences": all_confs
        }

    # Neither provided
    return "Please upload an audio file OR a spectrogram image."


# GRADIO UI
interface = gr.Interface(
    fn=classify,
    inputs=[
        gr.Audio(type="filepath", label="Upload Audio (WAV/MP3)"),
        gr.Image(type="pil", label="Upload Spectrogram Image")
    ],
    outputs=gr.JSON(label="Prediction Results"),
    title="General Audio Classifier (Audio + Spectrogram Support)",
    description=(
        "Upload a raw audio file OR a spectrogram image.\n"
        "If audio → model preprocesses into mel-spectrogram chunks.\n"
        "If image → model classifies the spectrogram directly.\n"
        "Built using CNN + Mel-Spectrogram + Gradio."
    ),
)

interface.launch()