Vision-to-Infrastructure: Turning Network Diagrams into Real Deployments

The Problem

Drawing a network diagram is one thing. Actually deploying it is another — and the gap between the two is full of manual, error-prone work. I wanted to eliminate that gap entirely.

The Approach

The pipeline has three stages:

1. Computer Vision Detection

I trained a YOLOv8 model on a custom dataset of network diagrams, labeling:

Routers (R)
Switches (S)
PCs / Endpoints (PC)
Links (edges between nodes)

After augmentation and tuning, the model reaches ~98% mAP on the validation set.

from ultralytics import YOLO
 
model = YOLO("yolov8n.pt")
results = model.train(
    data="network_diagrams.yaml",
    epochs=100,
    imgsz=640,
    batch=16,
)

2. Graph Extraction

Once nodes and links are detected, I build a graph from bounding-box centroids and proximity relationships:

import networkx as nx
 
def build_topology_graph(detections):
    G = nx.Graph()
    nodes = [d for d in detections if d.label in ("Router", "Switch", "PC")]
    links = [d for d in detections if d.label == "Link"]
 
    for node in nodes:
        G.add_node(node.id, type=node.label, bbox=node.bbox)
 
    for link in links:
        src, dst = find_endpoints(link, nodes)
        if src and dst:
            G.add_edge(src.id, dst.id)
 
    return G

3. Terraform + Containerlab Generation

Using Jinja2 templates, the graph is rendered into deployment files:

from jinja2 import Template
 
CLAB_TEMPLATE = Template("""
name: {{ topology_name }}
topology:
  nodes:
  {% for node in nodes %}
    {{ node.id }}:
      kind: {{ node.kind }}
  {% endfor %}
  links:
  {% for link in links %}
    - endpoints: ["{{ link.src }}:eth1", "{{ link.dst }}:eth1"]
  {% endfor %}
""")

Results

End-to-end: one image in, one running network out — no manual YAML writing, no topology files to maintain by hand.

The full code is on GitHub.