app.py

import streamlit as st
import librosa
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import firebase_admin
from firebase_admin import credentials, firestore
from datetime import datetime

# Optional: for microphone recording
try:
    from streamlit_audiorecorder import audiorecorder
    RECORDING_ENABLED = True
except ImportError:
    RECORDING_ENABLED = False

# --- Firebase Initialization ---
# On Hugging Face Spaces, add your service account JSON to Secrets.toml as "firebase_key"
@st.cache_resource
def init_firestore():
    # Load service account from st.secrets
    cred_dict = st.secrets["firebase_key"]
    cred = credentials.Certificate(cred_dict)
    firebase_admin.initialize_app(cred)
    return firestore.client()

db = init_firestore()

# --- Load ML Model ---
@st.cache_resource
def load_model():
    return tf.keras.models.load_model('Heart_ResNet.h5')

model = load_model()

# --- Audio Processing & Classification ---
SAMPLE_RATE = 22050
DURATION = 10
INPUT_LEN = SAMPLE_RATE * DURATION
CLASS_NAMES = ["artifact", "murmur", "normal"]

@st.cache

def process_audio(raw_bytes):
    # Save raw bytes to temp file
    with open("temp.wav", "wb") as f:
        f.write(raw_bytes)
    X, sr = librosa.load("temp.wav", sr=SAMPLE_RATE, duration=DURATION)
    if len(X) < INPUT_LEN:
        X = np.pad(X, (0, INPUT_LEN - len(X)), mode='constant')
    mfccs = np.mean(librosa.feature.mfcc(y=X, sr=sr, n_mfcc=52, n_fft=512, hop_length=256).T, axis=0)
    return mfccs, X, sr


def classify(mfccs):
    feats = mfccs.reshape(1, 52, 1)
    preds = model.predict(feats)
    return {name: float(preds[0][i]) for i, name in enumerate(CLASS_NAMES)}

# --- User Authentication Helpers ---

def register_user(email, password):
    user_ref = db.collection("users").document(email)
    if user_ref.get().exists:
        return False, "User already exists"
    user_ref.set({"email": email, "password": password, "history": []})
    return True, "Registration successful"


def login_user(email, password):
    user_ref = db.collection("users").document(email)
    user = user_ref.get()
    if user.exists and user.to_dict().get("password") == password:
        return True, user.to_dict()
    return False, None

# --- Firestore History Functions ---

def save_history(email, result):
    record = {"timestamp": datetime.utcnow().isoformat(), "result": result}
    db.collection("users").document(email).update({
        "history": firestore.ArrayUnion([record])
    })


def load_history(email):
    user = db.collection("users").document(email).get().to_dict()
    return user.get("history", [])

# --- Streamlit App Layout ---
st.title("🩺 Heartbeat Sound Classifier with Firestore")

# Session state for user
if "user" not in st.session_state:
    st.session_state.user = None

# Sidebar: Auth or Logout
with st.sidebar:
    if st.session_state.user:
        st.markdown(f"**Logged in as:** {st.session_state.user['email']}")
        if st.button("Logout"):
            st.session_state.user = None
            st.experimental_rerun()
    else:
        tab = st.radio("Account", ["Login", "Register"])
        email = st.text_input("Email")
        password = st.text_input("Password", type="password")
        if tab == "Register":
            if st.button("Sign Up"):
                success, msg = register_user(email, password)
                st.success(msg) if success else st.error(msg)
        else:
            if st.button("Login"):
                success, user = login_user(email, password)
                if success:
                    st.session_state.user = user
                    st.experimental_rerun()
                else:
                    st.error("Invalid credentials")

# Main: show after login
if st.session_state.user:
    st.header("Upload or Record Your Heartbeat")
    mode = st.radio("Input Mode:", ["Upload File", "Record (mic)" if RECORDING_ENABLED else "Upload File"])

    raw_audio = None
    if mode == "Upload File":
        up = st.file_uploader("Select WAV/MP3 file", type=["wav", "mp3"])
        if up:
            raw_audio = up.read()
            st.audio(raw_audio, format='audio/wav')
    else:
        audio_data = audiorecorder()
        if audio_data is not None:
            raw_audio = audio_data.tobytes()
            st.audio(raw_audio, format='audio/wav')

    if raw_audio:
        if st.button("Classify Heartbeat"):
            with st.spinner("Analyzing..."):
                mfccs, waveform, sr = process_audio(raw_audio)
                results = classify(mfccs)
                save_history(st.session_state.user["email"], results)

                # Display metrics
                st.subheader("Results")
                cols = st.columns(len(CLASS_NAMES))
                for col, name in zip(cols, CLASS_NAMES):
                    col.metric(name.title(), f"{results[name]*100:.2f}%")

                # Waveform plot
                fig, ax = plt.subplots(figsize=(8, 3))
                librosa.display.waveshow(waveform, sr=sr, ax=ax)
                ax.set(title="Heartbeat Waveform", xlabel="Time (s)", ylabel="Amplitude")
                st.pyplot(fig)

    # Show history
    st.header("Your Classification History")
    history = load_history(st.session_state.user["email"])
    if history:
        for rec in sorted(history, key=lambda x: x['timestamp'], reverse=True):
            st.write(f"**{rec['timestamp']}**")
            st.json(rec['result'])
    else:
        st.info("No history yet. Classify your first heartbeat!")
else:
    st.info("Please login or register to continue.")

# Footer
st.markdown("---")
st.markdown("Built with ❤️ and deployed on Hugging Face Spaces")