app.py

"""
Sinama Audio Classifier – Hugging Face Spaces API
--------------------------------------------------
FastAPI app that loads the trained CNN model, accepts audio uploads,
and returns top-5 predicted Cebuano/Sinama words with confidence scores.

Deploy this as a Hugging Face Space (Docker or Gradio SDK).
"""

import json
import os
import tempfile
from contextlib import asynccontextmanager

import librosa
import numpy as np
import tensorflow as tf
from fastapi import FastAPI, File, HTTPException, UploadFile
from fastapi.middleware.cors import CORSMiddleware

# ---------------------------------------------------------------------------
# Config – must match training preprocessing
# ---------------------------------------------------------------------------
SAMPLE_RATE = 22050
DURATION = 4.0
N_MELS = 128
N_FFT = 2048
HOP_LENGTH = 512
TARGET_LEN = int(SAMPLE_RATE * DURATION)

# Model files (uploaded to the Space repo)
MODEL_PATH = "best_model.keras"
LABEL_MAP_PATH = "label_map.json"

# ---------------------------------------------------------------------------
# Global state
# ---------------------------------------------------------------------------
model = None
label_map = None


@asynccontextmanager
async def lifespan(app: FastAPI):
    """Load model and label map on startup."""
    global model, label_map

    if not os.path.exists(MODEL_PATH):
        raise RuntimeError(f"Model file not found: {MODEL_PATH}")

    model = tf.keras.models.load_model(MODEL_PATH)
    print(f"[app] Model loaded from {MODEL_PATH}")

    with open(LABEL_MAP_PATH, "r", encoding="utf-8") as f:
        raw = json.load(f)
    label_map = {int(k): v for k, v in raw.items()}
    print(f"[app] Loaded {len(label_map)} classes")

    yield

    # Cleanup
    model = None
    label_map = None


app = FastAPI(
    title="Sinama Audio Classifier",
    description="Classify spoken Cebuano/Sinama words from audio clips",
    version="1.0.0",
    lifespan=lifespan,
)

# Allow Flutter app / any frontend to call this API
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)


# ---------------------------------------------------------------------------
# Audio preprocessing
# ---------------------------------------------------------------------------
def preprocess_audio(audio_bytes: bytes) -> np.ndarray:
    """Convert raw audio bytes → Mel spectrogram feature array."""
    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
        tmp.write(audio_bytes)
        tmp_path = tmp.name

    try:
        waveform, _ = librosa.load(tmp_path, sr=SAMPLE_RATE, mono=True)
    finally:
        os.unlink(tmp_path)

    # Pad or trim to fixed duration
    if len(waveform) < TARGET_LEN:
        waveform = np.pad(waveform, (0, TARGET_LEN - len(waveform)))
    else:
        waveform = waveform[:TARGET_LEN]

    # Mel spectrogram
    mel = librosa.feature.melspectrogram(
        y=waveform, sr=SAMPLE_RATE,
        n_mels=N_MELS, n_fft=N_FFT, hop_length=HOP_LENGTH,
    )
    mel_db = librosa.power_to_db(mel, ref=np.max)

    # Normalise
    mean, std = mel_db.mean(), mel_db.std()
    mel_db = (mel_db - mean) / (std + 1e-9)

    # Shape: (1, freq_bins, time_steps, 1)
    return mel_db[np.newaxis, ..., np.newaxis]


# ---------------------------------------------------------------------------
# Endpoints
# ---------------------------------------------------------------------------
@app.get("/health")
async def health():
    return {"status": "ok", "classes": len(label_map) if label_map else 0}


@app.post("/predict")
async def predict(file: UploadFile = File(...)):
    """
    Accept an audio file and return top-5 predictions.

    Returns:
        [{"label": "word", "score": 0.95}, ...]
    """
    if model is None or label_map is None:
        raise HTTPException(status_code=503, detail="Model not loaded")

    audio_bytes = await file.read()
    if len(audio_bytes) == 0:
        raise HTTPException(status_code=400, detail="Empty audio file")

    try:
        features = preprocess_audio(audio_bytes)
    except Exception as e:
        raise HTTPException(status_code=400, detail=f"Audio processing failed: {e}")

    preds = model.predict(features, verbose=0)[0]

    top_indices = np.argsort(preds)[::-1][:5]
    results = [
        {"label": label_map[int(i)], "score": round(float(preds[i]), 4)}
        for i in top_indices
    ]
    return results


@app.post("/predict/raw")
async def predict_raw(file: UploadFile = File(...)):
    """
    Same as /predict but returns ALL class probabilities.
    Useful for debugging or custom logic in the app.
    """
    if model is None or label_map is None:
        raise HTTPException(status_code=503, detail="Model not loaded")

    audio_bytes = await file.read()
    if len(audio_bytes) == 0:
        raise HTTPException(status_code=400, detail="Empty audio file")

    try:
        features = preprocess_audio(audio_bytes)
    except Exception as e:
        raise HTTPException(status_code=400, detail=f"Audio processing failed: {e}")

    preds = model.predict(features, verbose=0)[0]

    return {
        "predictions": {
            label_map[i]: round(float(preds[i]), 4)
            for i in range(len(preds))
        }
    }