RX12-Housing-Competition / visualize.py
GBDB02's picture
Update visualize.py
a255ca2 verified
"""
visualize.py — Visual diagnostics and statistics for the House Price Predictor.
Adds a "📊 Analytics" tab to the Gradio UI that shows:
1. Feature Importance — XGBoost gain-based + Lasso coefficient bar charts
2. Prediction Distribution — histogram + KDE of predicted prices
3. Residual Analysis — residual vs predicted scatter + Q-Q plot
4. Training Data Stats — target distribution, correlation heatmap, numeric summary
5. Model Comparison — CV RMSE bar chart across the three base learners
"""
import os
import io
import base64
import warnings
import numpy as np
import pandas as pd
import joblib
import matplotlib
matplotlib.use("Agg") # non-interactive backend for Gradio
import matplotlib.pyplot as plt
import matplotlib.gridspec as gridspec
from matplotlib.ticker import FuncFormatter
import scipy.stats as stats
warnings.filterwarnings("ignore")
# ── shared style ──────────────────────────────────────────────────────────────
PALETTE = ["#2D6A4F", "#40916C", "#74C69D", "#B7E4C7", "#D8F3DC"]
ACCENT = "#1B4332"
WARN = "#E76F51"
BG = "#F8F9FA"
GRID_CLR = "#DEE2E6"
def _style_ax(ax, title="", xlabel="", ylabel=""):
ax.set_facecolor(BG)
ax.grid(axis="y", color=GRID_CLR, linewidth=0.7, linestyle="--", zorder=0)
ax.spines[["top", "right"]].set_visible(False)
ax.spines[["left", "bottom"]].set_color(GRID_CLR)
if title: ax.set_title(title, fontsize=12, fontweight="bold", pad=10, color=ACCENT)
if xlabel: ax.set_xlabel(xlabel, fontsize=9, color="#495057")
if ylabel: ax.set_ylabel(ylabel, fontsize=9, color="#495057")
ax.tick_params(colors="#495057", labelsize=8)
def _fig_to_image(fig):
"""Convert a matplotlib figure → PIL Image (Gradio gr.Image compatible)."""
buf = io.BytesIO()
fig.savefig(buf, format="png", dpi=130, bbox_inches="tight", facecolor=fig.get_facecolor())
buf.seek(0)
from PIL import Image
img = Image.open(buf)
plt.close(fig)
return img
# ── helpers ───────────────────────────────────────────────────────────────────
def _load_artifacts():
from config import MODEL_PATH, PREPROCESSOR_PATH, META_PATH
for p in (MODEL_PATH, PREPROCESSOR_PATH, META_PATH):
if not os.path.exists(p):
raise FileNotFoundError("No trained model found. Train the model first.")
return joblib.load(MODEL_PATH), joblib.load(PREPROCESSOR_PATH), joblib.load(META_PATH)
def _feature_names(preprocessor, meta):
"""Reconstruct feature names after ColumnTransformer."""
num_feats = meta["numerical_features"]
try:
cat_enc = preprocessor.named_transformers_["cat"].named_steps["onehot"]
cat_feats = cat_enc.get_feature_names_out(meta["categorical_features"]).tolist()
except Exception:
cat_feats = []
return num_feats + cat_feats
# ══════════════════════════════════════════════════════════════════════════════
# PLOT 1 — Feature Importance
# ══════════════════════════════════════════════════════════════════════════════
def plot_feature_importance():
try:
ensemble, preprocessor, meta = _load_artifacts()
feature_names = _feature_names(preprocessor, meta)
n = 20 # top-N to show
estimators = dict(ensemble.named_estimators_)
fig, axes = plt.subplots(1, 2, figsize=(14, 6), facecolor="white")
fig.suptitle("Feature Importance", fontsize=15, fontweight="bold", color=ACCENT, y=1.01)
# ── XGBoost gain importance ──
ax = axes[0]
xgb_model = estimators.get("xgb")
if xgb_model is not None:
raw_imp = xgb_model.feature_importances_
n_feat = min(len(raw_imp), len(feature_names))
imp = pd.Series(raw_imp[:n_feat], index=feature_names[:n_feat])
top = imp.nlargest(n).sort_values()
bars = ax.barh(top.index, top.values, color=PALETTE[1], edgecolor="white", height=0.65)
for bar, val in zip(bars, top.values):
ax.text(val + top.values.max() * 0.01, bar.get_y() + bar.get_height() / 2,
f"{val:.4f}", va="center", fontsize=7, color=ACCENT)
_style_ax(ax, f"XGBoost — Top {n} Features (Gain)", "Importance", "")
else:
ax.text(0.5, 0.5, "XGBoost not available", ha="center", va="center")
# ── Lasso coefficients ──
ax = axes[1]
lasso_model = estimators.get("lasso")
if lasso_model is not None:
n_coef = min(len(lasso_model.coef_), len(feature_names))
coef = pd.Series(np.abs(lasso_model.coef_[:n_coef]), index=feature_names[:n_coef])
top = coef.nlargest(n).sort_values()
colors = [PALETTE[0] if v > 0 else WARN for v in top.values]
bars = ax.barh(top.index, top.values, color=colors, edgecolor="white", height=0.65)
for bar, val in zip(bars, top.values):
ax.text(val + top.values.max() * 0.01, bar.get_y() + bar.get_height() / 2,
f"{val:.4f}", va="center", fontsize=7, color=ACCENT)
_style_ax(ax, f"Lasso — Top {n} |Coefficients|", "|Coefficient|", "")
else:
ax.text(0.5, 0.5, "Lasso not available", ha="center", va="center")
fig.tight_layout()
return _fig_to_image(fig), "✅ Feature importance loaded."
except Exception as e:
return None, f"❌ {e}"
# ══════════════════════════════════════════════════════════════════════════════
# PLOT 2 — Prediction Distribution (requires test CSV)
# ══════════════════════════════════════════════════════════════════════════════
def plot_prediction_distribution(test_file):
try:
if test_file is None:
return None, "Please upload a test.csv file."
ensemble, preprocessor, meta = _load_artifacts()
from predict import _prepare
test_path = test_file.name if hasattr(test_file, "name") else test_file
test_df = pd.read_csv(test_path)
X_test = _prepare(test_df, meta)
preds = np.expm1(ensemble.predict(preprocessor.transform(X_test)))
fig, axes = plt.subplots(1, 2, figsize=(13, 5), facecolor="white")
fig.suptitle("Predicted Sale Price Distribution", fontsize=15, fontweight="bold", color=ACCENT)
# Histogram
ax = axes[0]
ax.hist(preds, bins=40, color=PALETTE[1], edgecolor="white", alpha=0.85)
ax.axvline(np.median(preds), color=WARN, linewidth=1.8, linestyle="--", label=f"Median: ${np.median(preds):,.0f}")
ax.axvline(np.mean(preds), color=ACCENT, linewidth=1.8, linestyle="-", label=f"Mean: ${np.mean(preds):,.0f}")
ax.xaxis.set_major_formatter(FuncFormatter(lambda x, _: f"${x/1e3:.0f}k"))
ax.legend(fontsize=8)
_style_ax(ax, "Histogram", "Predicted Price", "Count")
# Box + strip
ax = axes[1]
bp = ax.boxplot(preds, vert=True, patch_artist=True, widths=0.4,
boxprops=dict(facecolor=PALETTE[2], color=ACCENT),
medianprops=dict(color=WARN, linewidth=2),
whiskerprops=dict(color=ACCENT),
capprops=dict(color=ACCENT),
flierprops=dict(marker="o", color=PALETTE[0], alpha=0.3, markersize=3))
jitter = np.random.uniform(-0.15, 0.15, size=len(preds))
ax.scatter(1 + jitter, preds, alpha=0.12, s=6, color=PALETTE[0], zorder=3)
ax.yaxis.set_major_formatter(FuncFormatter(lambda y, _: f"${y/1e3:.0f}k"))
_style_ax(ax, "Box Plot + Jitter", "", "Predicted Price")
ax.set_xticks([])
# Stats table below
stats_txt = (f"n={len(preds):,} min=${preds.min():,.0f} "
f"Q1=${np.percentile(preds,25):,.0f} median=${np.median(preds):,.0f} "
f"Q3=${np.percentile(preds,75):,.0f} max=${preds.max():,.0f}")
fig.text(0.5, -0.02, stats_txt, ha="center", fontsize=8, color="#6C757D")
fig.tight_layout()
return _fig_to_image(fig), f"✅ Predictions generated for {len(preds):,} houses."
except Exception as e:
return None, f"❌ {e}"
# ══════════════════════════════════════════════════════════════════════════════
# PLOT 3 — Residual Analysis (requires train CSV to compute in-sample)
# ══════════════════════════════════════════════════════════════════════════════
def plot_residuals(train_file):
try:
if train_file is None:
return None, "Please upload train.csv to compute residuals."
ensemble, preprocessor, meta = _load_artifacts()
from predict import _prepare
train_path = train_file.name if hasattr(train_file, "name") else train_file
train_df = pd.read_csv(train_path)
if "SalePrice" not in train_df.columns:
return None, "train.csv must contain a SalePrice column."
y_true = train_df["SalePrice"].copy()
train_df = train_df.drop(columns=["SalePrice"], errors="ignore")
X = _prepare(train_df, meta)
y_pred = np.expm1(ensemble.predict(preprocessor.transform(X)))
residuals = y_true.values - y_pred
fig, axes = plt.subplots(1, 3, figsize=(16, 5), facecolor="white")
fig.suptitle("Residual Analysis (In-Sample)", fontsize=15, fontweight="bold", color=ACCENT)
# Residuals vs Predicted
ax = axes[0]
ax.scatter(y_pred, residuals, alpha=0.25, s=12, color=PALETTE[1])
ax.axhline(0, color=WARN, linewidth=1.5, linestyle="--")
ax.xaxis.set_major_formatter(FuncFormatter(lambda x, _: f"${x/1e3:.0f}k"))
ax.yaxis.set_major_formatter(FuncFormatter(lambda y, _: f"${y/1e3:.0f}k"))
_style_ax(ax, "Residuals vs Predicted", "Predicted Price", "Residual")
# Residual histogram
ax = axes[1]
ax.hist(residuals, bins=50, color=PALETTE[1], edgecolor="white", alpha=0.85)
ax.axvline(0, color=WARN, linewidth=1.5, linestyle="--")
ax.xaxis.set_major_formatter(FuncFormatter(lambda x, _: f"${x/1e3:.0f}k"))
_style_ax(ax, "Residual Distribution", "Residual", "Count")
# Q-Q plot
ax = axes[2]
(osm, osr), (slope, intercept, r) = stats.probplot(residuals, dist="norm")
ax.scatter(osm, osr, alpha=0.3, s=12, color=PALETTE[1])
line_x = np.array([osm[0], osm[-1]])
ax.plot(line_x, slope * line_x + intercept, color=WARN, linewidth=1.8)
_style_ax(ax, f"Q-Q Plot (R²={r**2:.3f})", "Theoretical Quantiles", "Sample Quantiles")
rmse = np.sqrt(np.mean(residuals**2))
mae = np.mean(np.abs(residuals))
fig.text(0.5, -0.02,
f"In-sample RMSE: ${rmse:,.0f} | MAE: ${mae:,.0f}",
ha="center", fontsize=9, color="#6C757D")
fig.tight_layout()
return _fig_to_image(fig), f"✅ Residuals computed. RMSE=${rmse:,.0f} MAE=${mae:,.0f}"
except Exception as e:
return None, f"❌ {e}"
# ══════════════════════════════════════════════════════════════════════════════
# PLOT 4 — Training Data Statistics (requires train CSV)
# ══════════════════════════════════════════════════════════════════════════════
def plot_data_stats(train_file):
try:
if train_file is None:
return None, "Please upload train.csv."
train_path = train_file.name if hasattr(train_file, "name") else train_file
df = pd.read_csv(train_path)
fig = plt.figure(figsize=(16, 10), facecolor="white")
fig.suptitle("Training Data Statistics", fontsize=15, fontweight="bold", color=ACCENT, y=1.01)
gs = gridspec.GridSpec(2, 3, figure=fig, hspace=0.45, wspace=0.35)
# ── SalePrice distribution ──
ax = fig.add_subplot(gs[0, 0])
ax.hist(df["SalePrice"], bins=50, color=PALETTE[1], edgecolor="white", alpha=0.85)
ax.axvline(df["SalePrice"].median(), color=WARN, linewidth=1.5, linestyle="--",
label=f"Median ${df['SalePrice'].median()/1e3:.0f}k")
ax.xaxis.set_major_formatter(FuncFormatter(lambda x, _: f"${x/1e3:.0f}k"))
ax.legend(fontsize=7)
_style_ax(ax, "SalePrice Distribution", "Sale Price", "Count")
# ── Log SalePrice ──
ax = fig.add_subplot(gs[0, 1])
log_price = np.log1p(df["SalePrice"])
ax.hist(log_price, bins=50, color=PALETTE[0], edgecolor="white", alpha=0.85)
_style_ax(ax, "log(SalePrice) Distribution", "log(1 + SalePrice)", "Count")
# ── Missing values (top 15) ──
ax = fig.add_subplot(gs[0, 2])
missing = (df.isnull().sum() / len(df) * 100).sort_values(ascending=False).head(15)
missing = missing[missing > 0]
if len(missing):
bars = ax.barh(missing.index[::-1], missing.values[::-1],
color=WARN, edgecolor="white", height=0.6)
for bar, val in zip(bars, missing.values[::-1]):
ax.text(val + 0.3, bar.get_y() + bar.get_height() / 2,
f"{val:.1f}%", va="center", fontsize=7, color=ACCENT)
_style_ax(ax, "Missing Values (top 15)", "Missing %", "")
# ── Overall Quality vs Price ──
ax = fig.add_subplot(gs[1, 0])
if "OverallQual" in df.columns:
groups = [df[df["OverallQual"] == q]["SalePrice"].values
for q in sorted(df["OverallQual"].unique())]
labels = sorted(df["OverallQual"].unique())
bp = ax.boxplot(groups, labels=labels, patch_artist=True,
boxprops=dict(facecolor=PALETTE[2], color=ACCENT),
medianprops=dict(color=WARN, linewidth=1.8),
whiskerprops=dict(color=ACCENT), capprops=dict(color=ACCENT),
flierprops=dict(marker=".", color=PALETTE[0], alpha=0.3, markersize=4))
ax.yaxis.set_major_formatter(FuncFormatter(lambda y, _: f"${y/1e3:.0f}k"))
_style_ax(ax, "Price by Overall Quality", "Quality Score", "Sale Price")
# ── Correlation with SalePrice (top 12 numerics) ──
ax = fig.add_subplot(gs[1, 1])
num_df = df.select_dtypes(include=[np.number]).drop(columns=["Id"], errors="ignore")
corr = num_df.corr()["SalePrice"].drop("SalePrice").abs().sort_values(ascending=False).head(12)
corr_signed = num_df.corr()["SalePrice"].drop("SalePrice").loc[corr.index]
colors = [PALETTE[0] if v > 0 else WARN for v in corr_signed.values]
ax.barh(corr.index[::-1], corr.values[::-1], color=colors[::-1], edgecolor="white", height=0.65)
_style_ax(ax, "Top Correlations with SalePrice", "|Pearson r|", "")
# ── Scatter GrLivArea vs SalePrice ──
ax = fig.add_subplot(gs[1, 2])
if "GrLivArea" in df.columns:
sc = ax.scatter(df["GrLivArea"], df["SalePrice"],
alpha=0.25, s=10, c=df.get("OverallQual", pd.Series(5, index=df.index)),
cmap="YlGn", edgecolors="none")
plt.colorbar(sc, ax=ax, label="Overall Quality", shrink=0.8)
ax.yaxis.set_major_formatter(FuncFormatter(lambda y, _: f"${y/1e3:.0f}k"))
_style_ax(ax, "GrLivArea vs SalePrice", "Above-Grade Living Area (sqft)", "Sale Price")
return _fig_to_image(fig), f"✅ Stats for {len(df):,} training samples loaded."
except Exception as e:
return None, f"❌ {e}"
# ══════════════════════════════════════════════════════════════════════════════
# PLOT 5 — Model CV Comparison (reads saved meta)
# ══════════════════════════════════════════════════════════════════════════════
def plot_model_comparison():
try:
_, _, meta = _load_artifacts()
cv_scores = meta.get("cv_scores", None)
if cv_scores is None:
return None, ("ℹ️ CV score details not stored in this model version.\n"
"Re-train to enable this chart.")
models = list(cv_scores.keys())
rmses = [cv_scores[m]["rmse"] for m in models]
stds = [cv_scores[m].get("std", 0) for m in models]
fig, ax = plt.subplots(figsize=(7, 4), facecolor="white")
x = np.arange(len(models))
bars = ax.bar(x, rmses, yerr=stds, color=PALETTE[:len(models)],
edgecolor="white", width=0.45, capsize=6,
error_kw=dict(ecolor=ACCENT, elinewidth=1.5))
for bar, val in zip(bars, rmses):
ax.text(bar.get_x() + bar.get_width() / 2, val + max(stds) * 0.05,
f"{val:.4f}", ha="center", va="bottom", fontsize=9, fontweight="bold", color=ACCENT)
ax.set_xticks(x)
ax.set_xticklabels(models, fontsize=10)
_style_ax(ax, "Cross-Validation RMSE (log scale)", "Model", "CV RMSE (log)")
fig.tight_layout()
return _fig_to_image(fig), "✅ Model comparison loaded."
except Exception as e:
return None, f"❌ {e}"
# ══════════════════════════════════════════════════════════════════════════════
# Gradio Tab builder — call this from app.py
# ══════════════════════════════════════════════════════════════════════════════
def build_analytics_tab():
"""
Returns a gr.Tab block. Import and embed it inside the gr.Tabs() block in app.py.
Usage in app.py:
from visualize import build_analytics_tab
with gr.Tabs():
...existing tabs...
build_analytics_tab()
"""
import gradio as gr
with gr.Tab("📊 Analytics") as tab:
gr.Markdown(
"### Visual Diagnostics\n"
"Explore model internals, data statistics, predictions and residuals.\n"
"> **Tip:** Train the model first; some charts also need a CSV upload."
)
with gr.Tabs():
# ── Feature Importance ──────────────────────────────────────────
with gr.Tab("Feature Importance"):
gr.Markdown("XGBoost gain-based importance **and** Lasso |coefficients|.")
btn_fi = gr.Button("Load Feature Importance", variant="primary")
img_fi = gr.Image(label="Feature Importance", type="pil")
msg_fi = gr.Markdown()
btn_fi.click(fn=plot_feature_importance, inputs=[], outputs=[img_fi, msg_fi])
# ── Prediction Distribution ─────────────────────────────────────
with gr.Tab("Prediction Distribution"):
gr.Markdown("Upload **test.csv** to visualise the distribution of predicted prices.")
f_pred = gr.File(label="Upload test.csv", file_types=[".csv"])
btn_pd = gr.Button("Generate Distribution", variant="primary")
img_pd = gr.Image(label="Prediction Distribution", type="pil")
msg_pd = gr.Markdown()
btn_pd.click(fn=plot_prediction_distribution, inputs=[f_pred], outputs=[img_pd, msg_pd])
# ── Residual Analysis ───────────────────────────────────────────
with gr.Tab("Residual Analysis"):
gr.Markdown("Upload **train.csv** to compute in-sample residuals.")
f_res = gr.File(label="Upload train.csv", file_types=[".csv"])
btn_res = gr.Button("Analyse Residuals", variant="primary")
img_res = gr.Image(label="Residual Analysis", type="pil")
msg_res = gr.Markdown()
btn_res.click(fn=plot_residuals, inputs=[f_res], outputs=[img_res, msg_res])
# ── Training Data Stats ─────────────────────────────────────────
with gr.Tab("Data Statistics"):
gr.Markdown("Upload **train.csv** to explore raw data distributions and correlations.")
f_stat = gr.File(label="Upload train.csv", file_types=[".csv"])
btn_st = gr.Button("Show Data Stats", variant="primary")
img_st = gr.Image(label="Data Statistics", type="pil")
msg_st = gr.Markdown()
btn_st.click(fn=plot_data_stats, inputs=[f_stat], outputs=[img_st, msg_st])
# ── Model Comparison ────────────────────────────────────────────
with gr.Tab("Model Comparison"):
gr.Markdown("CV RMSE across base learners.")
btn_mc = gr.Button("Load Model Comparison", variant="primary")
img_mc = gr.Image(label="Model Comparison", type="pil")
msg_mc = gr.Markdown()
btn_mc.click(fn=plot_model_comparison, inputs=[], outputs=[img_mc, msg_mc])
return tab