Skip to content

Latest commit

 

History

History
74 lines (55 loc) · 2.4 KB

README.md

File metadata and controls

74 lines (55 loc) · 2.4 KB

Image Captioning Production

Production-grade image captioning system with attention-based encoder-decoder architecture (TensorFlow).

Overview

  • Training: Attention-based caption model with EfficientNetB3/InceptionV3 feature extraction, experiment tracking, TensorBoard logging, and model versioning
  • Inference: FastAPI REST API with lazy model loading, health/metrics endpoints, beam search & greedy decoding
  • Testing: Unit, integration, and performance tests with pytest markers
  • MLOps: Automated experiment tracking, model versioning with SHA256 integrity checks, deployment manifest generation

Quick Start

# Training (outputs to shared/artifacts/)
python -m training.train --config training/config.yaml

# Inference API
docker-compose up inference
# OR: cd inference && uvicorn main:app --host 0.0.0.0 --port 8000

# Generate caption
curl -X POST "http://localhost:8000/predict" \
  -F "file=@image.jpg" \
  -F "algorithm=beam"

Key Features

Training Pipeline

  • Dataclass-based config with validation
  • Automated experiment tracking (metrics, configs, git info, TensorBoard)
  • Feature extractor registry (EfficientNet/Inception)
  • Mixed precision training, early stopping, LR scheduling
  • Kaggle notebook compatible

Inference Service

  • Thread-safe model registry with lazy loading
  • /predict, /health, /ready, /metrics endpoints
  • Greedy and beam search decoding
  • Request logging with timing

MLOps

  • Experiment Tracking: Automated logging of configs, metrics, BLEU scores, git state, environment info → run_summary.json, metrics.csv, TensorBoard logs
  • Testing: Pytest with markers (slow, integration, performance), coverage reports

Project Structure

image-captioning-production/
├── training/          # Training pipeline (configs, data, models, trainers, utils)
├── inference/         # FastAPI service (main, registry, caption_generator, preprocessing)
├── tests/             # Test suite (unit, integration, performance)
└── shared/artifacts/  # Model artifacts (models, tokenizers, logs, metrics)

Documentation

For developers & contributors: See DOCS.md for:

  • Architecture deep-dive
  • MLOps runbook (training, tracking, testing)
  • Configuration options
  • Extending the pipeline
  • Kaggle setup

Requirements

  • Python ≥ 3.9
  • TensorFlow ≥ 2.10
  • FastAPI, NumPy, PyYAML, tqdm, NLTK

License

MIT