| from __future__ import annotations |
|
|
| import json |
| from pathlib import Path |
|
|
| import gradio as gr |
| import numpy as np |
| import plotly.graph_objects as go |
| import torch |
| from model import MeshGraphGAT |
| from safetensors.torch import load_file |
|
|
| PROJECT_DIR = Path(__file__).resolve().parent |
| ARTIFACT_DIR = PROJECT_DIR / "artifacts" / "meshgraph-gat" |
| GRAPH = np.load(PROJECT_DIR / "data" / "meshgraph.npz") |
| PREPROCESSING = np.load(ARTIFACT_DIR / "preprocessing.npz") |
| FEATURES = torch.from_numpy( |
| ( |
| GRAPH["features"].astype(np.float32) - PREPROCESSING["mean"] |
| ) |
| / PREPROCESSING["scale"] |
| ) |
| ADJACENCY = torch.from_numpy(GRAPH["adjacency"].astype(np.float32)) |
| MODEL = MeshGraphGAT(FEATURES.shape[1]) |
| MODEL.load_state_dict(load_file(ARTIFACT_DIR / "model.safetensors")) |
| MODEL.eval() |
| REPORT = json.loads((ARTIFACT_DIR / "evaluation.json").read_text(encoding="utf-8")) |
|
|
|
|
| @torch.inference_mode() |
| def inspect_node(node: int) -> tuple[go.Figure, dict]: |
| node = int(node) |
| logits, attention = MODEL(FEATURES, ADJACENCY, return_attention=True) |
| risk = torch.softmax(logits, dim=1)[node, 1] |
| neighbors = torch.nonzero(ADJACENCY[node] > 0).flatten() |
| weights = attention[:, node, neighbors].mean(0) |
| order = torch.argsort(weights, descending=True) |
| neighbor_ids = neighbors[order].tolist() |
| sorted_weights = weights[order].tolist() |
| figure = go.Figure(go.Bar(x=[str(value) for value in neighbor_ids], y=sorted_weights)) |
| figure.update_layout( |
| template="plotly_dark", |
| title=f"Mean four-head attention from node {node}", |
| xaxis_title="Neighbor node", |
| yaxis_title="Attention", |
| ) |
| metrics = { |
| "node": node, |
| "compromise_probability": float(risk), |
| "predicted_compromised": bool( |
| float(risk) >= float(PREPROCESSING["threshold"]) |
| ), |
| "true_label": int(GRAPH["labels"][node]), |
| "degree": len(neighbor_ids), |
| "verified_test_roc_auc": REPORT["gat_test"]["roc_auc"], |
| } |
| return figure, metrics |
|
|
|
|
| with gr.Blocks(title="MeshGraph GAT") as demo: |
| gr.Markdown( |
| "# MeshGraph GAT\n" |
| "Inspect which enterprise communication edges a four-head graph-attention " |
| "network uses when estimating compromise risk." |
| ) |
| node = gr.Slider(0, len(FEATURES) - 1, value=25, step=1, label="Node") |
| initial = inspect_node(25) |
| chart = gr.Plot(value=initial[0]) |
| metrics = gr.JSON(value=initial[1]) |
| button = gr.Button("Inspect attention", variant="primary") |
| button.click(inspect_node, inputs=node, outputs=[chart, metrics]) |
|
|
|
|
| if __name__ == "__main__": |
| demo.launch() |
|
|
