scoring-pipeline

Sterne0

Forks0

Aktualisiert1. April 2026 um 03:21

Python scoring pipeline architecture — models, workers, engine, and how to extend the system.

Installation

Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.

Quelle

synthet

synthet/image-scoring-gallery

GitHub-Repository öffnen Creator-Repositorys ansehen

Download

Verwandte BerufeSOC

Basierend auf der SOC-Berufsklassifikation

SoftwareentwicklerInformatik- und Mathematikberufe·SOC 15-1252

SKILL.md

readonly

name	scoring-pipeline
description	Python scoring pipeline architecture — models, workers, engine, and how to extend the system.

Scoring Pipeline

The image-scoring project uses a multi-stage pipeline to assess image quality using neural network models. This skill covers the architecture, key modules, and how to safely make changes.

Architecture Overview

Input (folder/file)
  → PrepWorker (RAW conversion, thumbnails, skip checks)
    → ScoringWorker (GPU inference: MUSIQ, LIQE, TOPIQ)
      → ResultWorker (DB upsert, XMP metadata write, cleanup)

The pipeline uses producer-consumer queues with threading.Thread workers.

Key Modules

Module	Purpose
`modules/pipeline.py`	Worker classes: `PrepWorker`, `ScoringWorker`, `ResultWorker`, `ImageJob` dataclass
`modules/scoring.py`	`ScoringRunner` — orchestrates batch/single scoring, Fix DB, and metadata repair
`modules/engine.py`	`BatchImageProcessor` — low-level batch processing engine
`modules/config.py`	`load_config()`, `get_config_value()` — reads `config.json`
`modules/utils.py`	Path conversion (WSL ↔ Windows), hashing, thumbnail utilities

Model Wrappers

File	Model	Framework	Score Range
`musiq/run_all_musiq_models.py`	MUSIQ (SPAQ, AVA, KonIQ, PaQ2PiQ)	TensorFlow	0–100 (normalized to 0–1)
`modules/liqe.py`	LIQE (language-image quality evaluator)	PyTorch (pyiqa)	1–5 (normalized to 0–1)
`modules/topiq.py`	TOPIQ-IAA	PyTorch (pyiqa)	0–1
`modules/qalign.py`	Q-Align	PyTorch	1–5

Scoring Formulas (Current)

General  = 0.50 × LIQE + 0.30 × AVA + 0.20 × SPAQ
Technical = LIQE (primary)
Aesthetic = 0.60 × AVA + 0.40 × SPAQ

All composite scores are stored as floats 0.0–1.0 in the database.

Hybrid Environment

Windows: Runs the WebUI, PostgreSQL Docker, and file management.
WSL 2 (Linux): Runs GPU inference (TensorFlow + PyTorch with CUDA).
CRITICAL: Database access from WSL must use TCP (port 5432). PostgreSQL runs in Docker on the Windows host.

How to Add a New Model

Create a wrapper in modules/ (e.g., modules/newmodel.py) that exposes a score(image_path) → float method.
Import and call it in ScoringWorker.process() inside modules/pipeline.py.
Add the raw score column to the DB schema in modules/db.py → _init_db_impl().
Update the composite formulas in modules/scoring.py → ScoringRunner.fix_image_metadata().
Add normalization logic if the model's native range isn't 0–1.

Configuration

All config lives in config.json at the project root. Key sections:

database — filename, user, password
scoring — force_rescore_default, model weights

Mehr aus diesem Repository

gleiches Repository

changelog-commit-push

synthet/image-scoring-gallery

Update CHANGELOG.md and package.json via compiled release_bump harness, then commit/push only when the user asks. Use for /release, changelog updates, or shipping gallery versions.

2026-07-280

gallery-electron-ts

synthet/image-scoring-gallery

Works in image-scoring-gallery on electron/db.ts contract, IPC, apiService, React/Vite, and alignment with sibling image-scoring-backend schema. Use for desktop app bugs, database.engine modes, API URL/port from webui.lock, or TypeScript changes under electron/ or src/. Triggers include Electron gallery, db.ts, IPC, apiService, config.api, PostgreSQL vs api engine, and Vite renderer.

2026-07-280

gallery-ui

synthet/image-scoring-gallery

React component patterns, CSS Modules styling, design tokens, virtualization, and UX constitution for the Driftara Gallery renderer. Use for src/ UI work, styling, FilterPanel/GalleryGrid/ImageViewer, or design:check—not electron/db unless coordinated with gallery-electron-ts.

2026-07-280

validate-implementation

synthet/image-scoring-gallery

Verify implementation against spec AC-n criteria via compiled harness. Use after /implement or /test-and-fix, before /pr-ready. Parses ACs and emits the report; LLM assigns Verified/Failed/Unknown when evidence is not a clean command exit.

2026-07-280

verification-before-completion

synthet/image-scoring-gallery

Use before claiming work is complete, fixed, passing, ready to commit, or ready for PR. Runs scripts/agent_skills/verification_before_completion.py for the claim→proof catalog; LLM interprets output. Never upgrade incomplete verification.

2026-07-280

image-scoring-mcp

synthet/image-scoring-gallery

Driftara Gallery MCP — is-ui-* router-first; backend triage via sibling is-be-mcp search+dispatch.

2026-07-250

Scoring Pipeline

The image-scoring project uses a multi-stage pipeline to assess image quality using neural network models. This skill covers the architecture, key modules, and how to safely make changes.

Architecture Overview

Input (folder/file) → PrepWorker (RAW conversion, thumbnails, skip checks) → ScoringWorker (GPU inference: MUSIQ, LIQE, TOPIQ) → ResultWorker (DB upsert, XMP metadata write, cleanup)

The pipeline uses producer-consumer queues with threading.Thread workers.

Key Modules

Module

Purpose

modules/pipeline.py

Worker classes: PrepWorker, ScoringWorker, ResultWorker, ImageJob dataclass

modules/scoring.py

ScoringRunner — orchestrates batch/single scoring, Fix DB, and metadata repair

modules/engine.py

BatchImageProcessor — low-level batch processing engine

modules/config.py

load_config(), get_config_value() — reads config.json

modules/utils.py

Path conversion (WSL ↔ Windows), hashing, thumbnail utilities

Model Wrappers

File

Model

Framework

Score Range

musiq/run_all_musiq_models.py

MUSIQ (SPAQ, AVA, KonIQ, PaQ2PiQ)

TensorFlow

0–100 (normalized to 0–1)

modules/liqe.py

LIQE (language-image quality evaluator)

PyTorch (pyiqa)

1–5 (normalized to 0–1)

modules/topiq.py

TOPIQ-IAA

PyTorch (pyiqa)

0–1

modules/qalign.py

Q-Align

PyTorch

1–5

Scoring Formulas (Current)

General = 0.50 × LIQE + 0.30 × AVA + 0.20 × SPAQ Technical = LIQE (primary) Aesthetic = 0.60 × AVA + 0.40 × SPAQ

All composite scores are stored as floats 0.0–1.0 in the database.

Hybrid Environment

Windows: Runs the WebUI, PostgreSQL Docker, and file management.

WSL 2 (Linux): Runs GPU inference (TensorFlow + PyTorch with CUDA).

CRITICAL: Database access from WSL must use TCP (port 5432). PostgreSQL runs in Docker on the Windows host.

How to Add a New Model

Create a wrapper in modules/ (e.g., modules/newmodel.py) that exposes a score(image_path) → float method.

Import and call it in ScoringWorker.process() inside modules/pipeline.py.

Add the raw score column to the DB schema in modules/db.py → _init_db_impl().

Update the composite formulas in modules/scoring.py → ScoringRunner.fix_image_metadata().

Add normalization logic if the model's native range isn't 0–1.

Configuration

All config lives in config.json at the project root. Key sections:

database — filename, user, password

scoring — force_rescore_default, model weights