# %% Per-model drilldown — render one block per released model # Per-version trend (for the per-model drilldown Performance Trend chart # and the version-over-version Regression PASS/FAIL tile). trend_df = pd.read_csv("dashboard_trend_export.csv") trend_df["released_at"] = pd.to_datetime(trend_df["released_at"], errors="coerce") # %% Per-model drilldown — render one block per released model for _, m in df[df["status"] == "released"].iterrows(): color = PALETTE.get(m["status"], PALETTE["muted"]) # --- Section header --- badge = _badge(m["status"], m["status"]) display(HTML(f"""

        {m['model_name']} — {m['task']}
      
        {m['sensor_modality']} · {m['version'] or '(no version)'}
          {badge}

""")) # --- 6 KPI tiles --- miou_val = m["miou"] if pd.notna(m["miou"]) else None f1_val = m["macro_f1"] if pd.notna(m["macro_f1"]) else None p95 = m["p95_latency_ms"] if pd.notna(m["p95_latency_ms"]) else None p99 = m["p99_latency_ms"] if pd.notna(m["p99_latency_ms"]) else None n_dep = int(m["field_deployments_count"]) if pd.notna(m["field_deployments_count"]) else 0 # Regression — derived from the trend (latest version's regression_vs_prev) model_trend = (trend_df[trend_df["model_id"] == m["model_id"]] .sort_values("released_at")) if len(model_trend) and pd.notna(model_trend.iloc[-1].get("regression_vs_prev")): regression_text = model_trend.iloc[-1]["regression_vs_prev"] regression_color = (PALETTE["released"] if regression_text == "PASS" else PALETTE["issue"]) else: regression_text = "—" regression_color = PALETTE["muted"] # 6 KPI tiles in one subplot row — including REGRESSION as the 6th. # The 6th uses plotly Indicator's title slot to render text (PASS/FAIL) # at the same visual position as the numbers in the other 5 tiles; the # numeric "value" area is hidden so nothing renders there. fig = make_subplots(rows=1, cols=6, specs=[[{"type": "indicator"}] * 6]) numeric_tiles = [ (m["headline_metric_name"].upper(), miou_val if m["sensor_modality"] == "Camera" else f1_val, PALETTE["released"], ".3f"), ("F1 SCORE", f1_val, PALETTE["released"], ".3f"), ("P95 LATENCY (ms)", p95, PALETTE["deploy"], ".1f"), ("P99 LATENCY (ms)", p99, PALETTE["deploy"], ".1f"), ("FIELD DEPLOYMENTS", n_dep, PALETTE["deploy"], ",d"), ] for i, (label, value, color, vfmt) in enumerate(numeric_tiles): fig.add_trace( go.Indicator( mode="number", value=value if value is not None else None, title={"text": label, "font": {"size": 10, "color": PALETTE["muted"]}}, number={"font": {"size": 38, "color": color}, "valueformat": vfmt or ""}, ), row=1, col=i + 1, ) # 6th tile: REGRESSION text. The Indicator itself is fully invisible # (title + number both hidden) and only serves to reserve the 6th # subplot's space. Two paper-coord annotations render the visible # content at fixed y positions to match the other 5 tiles' title/number # baselines. fig.add_trace( go.Indicator( mode="number", value=0, title={"text": "", "font": {"size": 1, "color": "rgba(0,0,0,0)"}}, number={"font": {"size": 1, "color": "rgba(0,0,0,0)"}, "valueformat": ""}, ), row=1, col=6, ) # x center of subplot 6 of 6: (5/6 + 1) / 2 = 11/12. # y values are empirically tuned to match the other tiles' baselines: # - REGRESSION label sits where the other tiles' titles render # - PASS/FAIL big text sits where the other tiles' numbers render REGRESSION_LABEL_Y = 0.81 # adjust if label is too high / too low REGRESSION_VALUE_Y = 0.50 # adjust if value is too high / too low fig.add_annotation( x=11.0 / 12.0, y=REGRESSION_LABEL_Y, xref="paper", yref="paper", xanchor="center", yanchor="middle", text="REGRESSION", showarrow=False, font=dict(size=10, color=PALETTE["muted"], family="monospace"), align="center", ) fig.add_annotation( x=11.0 / 12.0, y=REGRESSION_VALUE_Y, xref="paper", yref="paper", xanchor="center", yanchor="middle", text=f'{regression_text}', showarrow=False, font=dict(size=38, color=regression_color, family="monospace"), align="center", ) fig.update_layout(**_plotly_layout( title=f"{m['model_name']} — KPIs", subtitle="Headline tiles for the per-model engineering view", height=220, )) fig.show() # --- Performance Trend --- # Reads vw_model_version_trend (one row per historical version with # the metric pulled from the materialized aggregates). m_trend = (trend_df[trend_df["model_id"] == m["model_id"]] .sort_values("released_at")) fig = go.Figure() if len(m_trend): fig.add_trace(go.Scatter( x=m_trend["version"], y=m_trend["headline_metric_value"], mode="lines+markers", line=dict(color=PALETTE["released"], width=3), marker=dict(size=12, color=PALETTE["released"]), text=[f"{v:.3f}" if pd.notna(v) else "—" for v in m_trend["headline_metric_value"]], textposition="top center", name=m["headline_metric_name"], )) fig.update_layout(**_plotly_layout( title="Performance Trend", subtitle=f"{m['headline_metric_name']} across versions", height=300, xaxis=dict(title="Version", showgrid=False), yaxis=dict(title=m["headline_metric_name"], showgrid=True, gridcolor="#1e293b"), )) fig.show() # --- Readiness Radar --- # 5 composite axes. Where the source data exists, normalize to [0,1]. # Where not, fall back to None (rendered as 0 with the placeholder note). axes = ["Accuracy", "Latency", "Field Maturity", "Regression", "Stability"] radar_values = [ m["headline_metric_value"] if pd.notna(m["headline_metric_value"]) else 0, # Latency: lower is better. Normalize against a 0–50 ms band, invert. (1.0 - _normalize(p95, 0, 50)) if p95 is not None else 0, # Field Maturity: normalize deployment count against 0–10 band. _normalize(n_dep, 0, 10) or 0, # Regression: 1.0 if latest version PASSed, 0 if FAILed, 0 if unknown. (1.0 if regression_text == "PASS" else 0.0), # Stability: placeholder — needs a source metric we don't track yet. 0, ] # Close the polygon radar_values_closed = radar_values + [radar_values[0]] axes_closed = axes + [axes[0]] fig = go.Figure() fig.add_trace(go.Scatterpolar( r=radar_values_closed, theta=axes_closed, fill="toself", line=dict(color=PALETTE["released"], width=2), fillcolor="rgba(34, 197, 94, 0.25)", name=m["model_name"], )) fig.update_layout(**_plotly_layout( title="Readiness Radar", subtitle="Composite readiness — Accuracy / Latency / Field Maturity " "(Regression & Stability placeholders)", height=400, polar=dict( bgcolor=PALETTE["card"], radialaxis=dict(visible=True, range=[0, 1], gridcolor="#1e293b", linecolor="#1e293b"), angularaxis=dict(gridcolor="#1e293b", linecolor="#1e293b"), ), showlegend=False, )) fig.show()