Skip to content

Roboflow AMD Training - Historical

Status: Historical local experiment track Last verified: 2026-07-16 Purpose: Preserve the earlier Roboflow and AMD Windows runtime work without presenting it as the current shared DraftLint architecture.

Why This Subsection Exists

This subsection exists because the current drawing-analysis training work is not just about labeling classes. It also depends on the runtime and tooling path used to make detector experiments practical on local hardware.

The historical tracked workspace is:

D:\02_Code\41_Training_drawing_roboflow_windowscode

That repo combines several concerns in one place:

  • Roboflow-hosted layout detection experiments
  • OCR runtime and backend-selection work
  • YOLO-based symbol-detection experiments
  • Windows AMD DirectML inference enablement

This subsection is the archive home for that combined training/runtime story. Current system identity lives in DraftLint System Family, and current training operations live in DraftLint Model Training.

What Lives Here

Current Operational Truth

The historical workspace remains useful, but it should be treated as a parts donor and experiment bed rather than a product-ready runtime. The strongest reusable work there is the OCR and inference runtime split, especially for Windows AMD local inference.

The weakest area is still layout detection, because it depends on a hosted Roboflow model and falls back to simple heuristics when that hosted path fails.

Sources

  • D:\02_Code\41_Training_drawing_roboflow_windowscode
  • C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260411_draftlint-repo-assessment-and-doc-map.md