Skip to content

PIPELINE.md

Latest commit

 

History

History
executable file
·
193 lines (148 loc) · 7.77 KB

PIPELINE.md

File metadata and controls

executable file
·
193 lines (148 loc) · 7.77 KB
Raw

Continuous Improvement Pipeline

أنبوب التحسين المستمر

How corrections from Hugging Face Space flow back to improve the OCR system.

Pipeline Overview

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│ 1. GT Import    │────▶│ 2. Benchmarks    │────▶│ 3. Trainer      │
│                 │     │                  │     │                 │
│ ABBYY/ReadIRIS  │     │ CER/WER/Medical  │     │ Streamlit UI    │
│ → Ground Truth  │     │ Term Accuracy    │     │ Ensemble OCR    │
│                 │     │                  │     │ Interactive Fix │
│ Repo:           │     │ Repo:            │     │ Repo:           │
│ medical-ocr-    │     │ medical-ocr-     │     │ medical-ocr-    │
│ ground-truth    │     │ benchmarks       │     │ trainer         │
└─────────────────┘     └──────────────────┘     └────────┬────────┘
                                                         │
                                                         │ corrections
                                                         ▼
┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│ 6. Platform     │◀────│ 5. Benchmarks    │◀────│ 4. HF Space     │
│                 │     │ (Re-evaluate)    │     │                 │
│ omni-medical-   │     │                  │     │ Online Demo     │
│ suite           │     │ Baseline         │     │ User Corrections│
│                 │     │ Regression Check │     │ Quick Feedback  │
│ Full Pipeline   │     │                  │     │                 │
│ Production Use  │     │ Repo:            │     │ Repo:           │
│                 │     │ medical-ocr-     │     │ medical-ocr-    │
│                 │     │ benchmarks       │     │ trainer-hf      │
└─────────────────┘     └──────────────────┘     └─────────────────┘
        │
        │ improved models/dictionaries
        ▼
┌─────────────────┐
│ 7. Training Hub │
│                 │
│ Bridge: HF ↔ GitHub
│ Continuous Model
│ Improvement     │
│                 │
│ Repo:           │
│ medical-ocr-    │
│ training-hub    │
└─────────────────┘

Step-by-Step Flow

Step 1: Import Ground Truth

Repository: medical-ocr-ground-truth 

# Import from ABBYY FineReader
python import_ground_truth.py abbyy_output.docx --output gt_page588.txt

# Import from ReadIRIS
python import_ground_truth.py readiris_output.rtf --output gt_readiris.txt

# Compare OCR with GT and generate correction dictionary
python gt_comparison_engine.py --gt gt_page588.txt --ocr ocr_output.txt \
    --generate-dict --output report.json

Output: Ground truth files, correction dictionaries, training pairs.

Step 2: Establish Baselines

Repository: medical-ocr-benchmarks 

# Run benchmarks against current engines
medocr-bench --engines paddleocr,tesseract --check-ci

# Results saved to results/ with CER/WER/medical accuracy

Output: Baseline metrics for regression detection.

Step 3: Collect Corrections

Repository: medical-ocr-trainer 

streamlit run app.py
# 1. Upload medical documents
# 2. Run ensemble OCR (5 engines)
# 3. Review and correct errors
# 4. Export training data

Output: Corrected pairs, training datasets (JSONL/CSV/Parquet).

Step 4: Gather User Feedback (HF Space)

Hugging Face: DrAbdulmalek/medical-ocr-trainer

Users try the demo, upload documents, and make corrections. These corrections are captured and can be bridged back to GitHub.

Step 5: Training Hub Integration

Repository: medical-ocr-training-hub 

HF Space corrections → Training Hub → GitHub training data → Model improvement

The Training Hub acts as the bridge between online user feedback and offline training.

Step 6: Re-evaluate

Repository: medical-ocr-benchmarks 

# Re-run benchmarks after improvement
medocr-bench --engines paddleocr --check-ci

# Compare against previous baselines
medocr-bench --compare-baseline results/baseline_v1.json

CI automatically detects regressions (≥5% degradation).

Step 7: Deploy to Platform

Repository: omni-medical-suite

# Update postprocessor dictionary
cp improved_dict.json omni-medical-suite/data/correction_dict.json

# Deploy
cd omni-medical-suite && make deploy

Automation

CI Pipeline (Automatic)

Trigger Action Repository
Push to ground-truth data files Run GT validation → notify benchmarks ground-truth
Weekly Monday 4AM Full benchmark suite benchmarks
PR to any repo Quick benchmark check benchmarks
Push to postprocessor Publish to PyPI postprocessor
Push to suite Full CI (lint + test + build) omni-medical-suite

Manual Steps

Step Frequency Effort
Import new GT data Monthly 30 min
Review HF Space corrections Weekly 15 min
Export training data from trainer As needed 5 min
Re-baseline after changes Per release 10 min
Deploy updated dictionaries Per improvement 5 min

Quality Gates

Before any release reaches the platform:

  1. medical-ocr-benchmarks CI passes (no regression > 5%)
  2. medical-ocr-postprocessor tests pass (37+ tests) 3 [ ] New dictionary entries validated against GT
  3. No default credentials in omni-medical-suite
  4. Audit event model checksums verified

Repository Map

omni-medical-suite          ← Main Platform (production use)
├── medical-ocr-postprocessor ← Core correction library (pip install)
├── medical-handwriting-ocr   ← Production OCR (GPU/Lite modes)
├── medical-ocr-trainer       ← Data collection & correction
├── medical-ocr-trainer-hf    ← HF Space (demo only)
├── medical-ocr-benchmarks    ← Quality measurement
├── medical-ocr-ground-truth  ← Golden truth source
├── medical-ocr-training-hub  ← HF ↔ GitHub bridge
├── OmniFile_Processor        ← Legacy (read-only)
└── medical-doc-processor     ← Legacy (read-only)

This pipeline creates a closed loop: GT → Measure → Correct → Train → Re-measure → Deploy

Part of the Medical OCR Ecosystem