Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
57 lines (38 loc) · 3.53 KB

ARCHITECTURE.md

File metadata and controls

57 lines (38 loc) · 3.53 KB
Raw

Project Architecture

This repo is the application layer in the vision-stack cluster. It owns CLI parsing, app configuration, runtime wiring, visualization, and end-to-end execution flow.

Canonical Sources

Repo Boundaries

  • neuriplo-infer: CLI, configuration, runtime wiring, visualization, end-to-end app flow
  • neuriplo-tasks: task contracts, preprocessing, postprocessing, result types, model-specific task logic
  • neuriplo: backend abstractions, backend adapters, runtime compatibility, backend dependency versions
  • videocapture: source semantics, file/stream/camera handling, video backend priority and behavior

What This Repo Intentionally Does Not Own

  • Tensor shapes, dtype semantics, and result-schema meaning
  • Backend implementation details or backend package versions
  • Video backend selection policy beyond wiring through videocapture
  • Upstream model-type inventory beyond the subset this app routes and validates end to end

Those contracts live in sibling repos and should not be redefined here in hand-maintained prose.

Build and Dependency Notes

  • The build currently requires CMake 3.24 and C++20. Read these from CMakeLists.txt, not from copied snippets in docs.
  • This repo derives a shared dependency ref for neuriplo-tasks, neuriplo, and videocapture in cmake/versions.cmake.
  • Backend package versions such as ONNX Runtime, TensorRT, LibTorch, OpenVINO, TensorFlow, and CUDA are owned by neuriplo and consumed through setup/build wiring here.

Model Types

The generated list in docs/generated/supported-model-types.md reflects the upstream neuriplo-tasks TaskFactory inventory.

That list is broader than the guarantees made by this application repo. End-to-end behavior still depends on the app's own CLI validation, task routing, rendering, and test coverage.

Runtime Flow

The app runtime is split by responsibility:

  • CommandLineParser: parses CLI flags and validates app-owned input paths and task-specific assets.
  • NeuriploInfer: sets up logging, builds an InferencePipeline, and selects a CLI command.
  • InferencePipelineBuilder: wires labels, backend engine, model metadata, neuriplo-tasks task config, task instance, and result renderer.
  • CLICommands: owns executable workflows such as normal inference, warmup, benchmarking, image-understanding dispatch, and metadata export.
  • ResultRenderer: renders neuriplo-tasks::Result variants for each routed TaskType.
  • TaskRouting: maps app model-type strings to neuriplo_tasks::TaskType and must stay aligned with neuriplo_tasks::TaskFactory.

--export_metadata follows the same backend and task setup path as inference, then prints model type, routed task type, input layers, and output layers before exiting. It intentionally does not require a source.

Local Quality Checks

Repo-local validation remains the canonical CMake test flow from ops metadata. For deeper local checks, scripts/check_code_quality.sh runs format checking, cppcheck, an AddressSanitizer/UndefinedBehaviorSanitizer test build, and a ThreadSanitizer test build when those tools are installed.