Visualization

ck.visualize_circuit() has three modes: an interactive Plotly graph, a comparison dashboard, and a Streamlit dashboard. The graph and comparison modes export to standalone HTML.

Quick start

import circuitkit as ck

circuit = ck.load_scores("./circuit.pt")
model = ck.load_model("gpt2")

# Interactive graph view
fig = ck.visualize_circuit(circuit, mode="graph")
fig.show()

# Export to HTML
ck.visualize_circuit(circuit, mode="graph", output="./circuit_graph.html")

Via Pipeline:

pipe.visualize(mode="graph", output="./circuit.html")

Mode 1: Graph view

Renders the circuit as a directed graph: nodes are attention heads and MLP layers, edges represent information flow weighted by attribution score.

fig = ck.visualize_circuit(circuit, mode="graph")

• Nodes coloured by type (blue = attention head, amber = MLP)
• Node size proportional to importance score
• Edges only for components at or above sparsity threshold
• Hover tooltips with layer, head index, and importance score

Graph mode delegates to Circuit.plot, which takes only an optional output path. Pass output to write an HTML export, or omit it for the inline widget:

ck.visualize_circuit(circuit, mode="graph", output="./circuit_graph.html")

Mode 2: Comparison dashboard

Compare two circuits side-by-side: algorithm comparisons, cross-task transfer, or before/after pruning.

circuit_a = ck.load_scores("./circuit_eap_ig.pt")
circuit_b = ck.load_scores("./circuit_acdc.pt")

dashboard = ck.visualize_circuit(
    circuit_a,
    mode="comparison",
    second_circuit=circuit_b,
    comparison_type="stability",   # "stability", "robustness", or "generalization"
    labels=["EAP-IG", "ACDC"],
)

Both circuits must carry node scores. Passing output="./compare.html" writes the dashboard to HTML and returns None; without output you get the ComparisonDashboard instance back.

comparison_type accepts "stability" (default), "robustness", or "generalization". It labels the comparison in the exported metadata; the HTML export itself renders all available views (stability heatmap, correlation matrix, robustness, transfer matrix, and score distribution).

Mode 3: Streamlit dashboard

Launch the interactive Streamlit dashboard as a background subprocess.

ck.visualize_circuit(circuit, mode="dashboard")

This runs streamlit run on the bundled circuitkit/visualize/streamlit_app.py and returns None. Streamlit must be installed, and the app opens in your browser rather than rendering inline.

CLI

There is no circuitkit visualize CLI command. Visualization is currently available only through the Python API shown above (ck.visualize_circuit() / Pipeline.visualize()).

Jupyter widgets

In a Jupyter notebook, the visualizations render inline automatically. For a bundled multi-tab analysis view (graph, saliency, feature attribution, editor, comparison), use display_circuit_analysis from circuitkit.visualize:

from circuitkit.visualize import display_circuit_analysis

display_circuit_analysis(
    circuit={"nodes": [...], "edges": [...]},
    node_scores=circuit.scores,
)

This requires ipywidgets. See the JupyterWidgetSuite class in circuitkit.visualize for the full set of tabs and options.

Next steps

• Applications
• Selectors