app.py

import json
from pathlib import Path

import joblib
import pandas as pd
import streamlit as st

# ============================================================
# Paths
# ============================================================
# Adjust MODEL_DIR if your saved_models folder lives elsewhere
# relative to wherever you run `streamlit run app.py` from.
MODEL_DIR = Path("src/saved_models")
SCALER_PATH = MODEL_DIR / "four_feature_scaler.joblib"
METRICS_PATH = MODEL_DIR / "metrics.json"

# ============================================================
# Page config
# ============================================================
st.set_page_config(
    page_title="Diamond Price Prediction",
    page_icon="💎",
    layout="centered"
)

st.title("💎 Diamond Price Prediction System")
st.write(
    """
    Pick a model, describe the diamond's characteristics, and get a live
    price estimate. All fields update the prediction instantly — no
    button needed.
    """
)

# ============================================================
# Discover available models
# ============================================================
# Any *_model.joblib file in saved_models is treated as a selectable model.
# The scaler file is explicitly excluded from this list.
if not MODEL_DIR.exists():
    st.error(f"Model directory not found: {MODEL_DIR.resolve()}")
    st.stop()

model_files = sorted(MODEL_DIR.glob("*_model.joblib"))

if not model_files:
    st.error(f"No '*_model.joblib' files found in {MODEL_DIR.resolve()}")
    st.stop()

# Friendly names for known model abbreviations
FRIENDLY_NAMES = {
    "etr": "Extra Tree Regression",
    "svr": "Support Vector Regression",
    "lr": "Linear Regression",
    "gbr": "Gradient Boosting Regression (Pro)",
}

# Build a friendly display name -> path mapping, e.g. "Extra Tree Regression" -> etr_model.joblib
MODEL_OPTIONS = {
    FRIENDLY_NAMES.get(p.stem.replace("_model", "").lower(), p.stem.replace("_model", "").upper()): p
    for p in model_files
}

# ============================================================
# Cached loaders (avoid re-reading disk on every slider move)
# ============================================================
@st.cache_resource
def load_model(path: str):
    return joblib.load(path)


@st.cache_resource
def load_scaler(path: str):
    return joblib.load(path)


@st.cache_resource
def load_encoder(path: str):
    return joblib.load(path)


@st.cache_data
def load_metrics(path: str):
    """Load precomputed R2 / MAE / RMSE per model, if available.

    Expected format (see testing.py note below for how to produce this):
    {
        "ETR": {"r2": 0.98, "mae": 250.1, "rmse": 400.3, "fit_time": 3.2},
        "GBR": {...},
        ...
    }
    """
    p = Path(path)
    if not p.exists():
        return None
    with open(p, "r") as f:
        return json.load(f)


try:
    scaler = load_scaler(str(SCALER_PATH))
except FileNotFoundError as e:
    st.error(f"Required file not found: {e.filename}")
    st.stop()

metrics = load_metrics(str(METRICS_PATH))

# ============================================================
# Model selection
# ============================================================
model_choice = st.selectbox("Select a model", list(MODEL_OPTIONS.keys()))

try:
    model = load_model(str(MODEL_OPTIONS[model_choice]))
except FileNotFoundError:
    st.error(f"Model file not found: {MODEL_OPTIONS[model_choice].resolve()}")
    st.stop()

# ============================================================
# Metrics expander
# ============================================================
with st.expander("📊 Model Performance"):
    if metrics is None:
        st.info(
            f"No metrics file found at `{METRICS_PATH}`. "
            "Run the updated `testing.py` (see notes) to generate one, "
            "then re-run this app."
        )
    elif model_choice not in metrics:
        st.warning(f"No stored metrics for '{model_choice}' in metrics.json.")
    else:
        m = metrics[model_choice]
        col1, col2, col3 = st.columns(3)
        col1.metric("R²", f"{m['r2']:.4f}")
        col2.metric("MAE", f"${m['mae']:,.2f}")
        col3.metric("RMSE", f"${m['rmse']:,.2f}")
        if "fit_time" in m:
            st.caption(f"Fit time: {m['fit_time']:.4f} s")

# ============================================================
# Training ranges (numerical features) — used for slider bounds
# Pulled from your describe() output. Update these if your split differs.
# ============================================================
CARAT_MIN, CARAT_MAX = 0.20, 2.00
X_MIN, X_MAX = 3.70, 8.30
Y_MIN, Y_MAX = 3.68, 8.30
Z_MIN, Z_MAX = 1.40, 5.30


# ============================================================
# User inputs — all 9 features
# ============================================================
st.subheader("Diamond Characteristics")


carat = st.slider("Carat", CARAT_MIN, CARAT_MAX, 1.00, 0.01)
x = st.slider("Length (x) mm", X_MIN, X_MAX, 5.50, 0.01)
y = st.slider("Width (y) mm", Y_MIN, Y_MAX, 5.50, 0.01)
z = st.slider("Depth (z) mm", Z_MIN, Z_MAX, 3.50, 0.01)


# ============================================================
# Logical (cross-field) sanity checks
# ============================================================
warnings = []
if z > x:
    warnings.append("Depth (z) is currently greater than Length (x) — that's unusual for a real diamond.")
if z > y:
    warnings.append("Depth (z) is currently greater than Width (y) — that's unusual for a real diamond.")

for msg in warnings:
    st.warning(msg)

# ============================================================
# Live prediction
# ============================================================
# Column order MUST match the order used during training in preprocessing.py.
# This order follows the feature table you provided (carat, cut, color,
# clarity, depth, table, x, y, z). Adjust if your training pipeline differs.
input_data = pd.DataFrame({
    "carat": [carat],
    "x": [x],
    "y": [y],
    "z": [z],
})


try:
    input_scaled = scaler.transform(input_data)
    prediction = model.predict(input_scaled)
    price = prediction[0]
    st.success(f"Estimated Diamond Price ({model_choice}): ${price:,.2f}")
except Exception as e:
    st.error(f"Prediction failed: {e}")
    st.caption(
        "This usually means the feature columns/order, the encoder's "
        "expected structure, or the category label spelling doesn't match "
        "what preprocessing.py used for training."
    )

# py -m streamlit run app.py