Image Captioning System — Dev Scaffold (v1.0.0)

InceptionV3 + Transformer image captioning architecture.

This release contains a deployment scaffold used for end-to-end system validation and infrastructure testing. It is intentionally published before the production training run so the full serving stack (FastAPI backend, Hugging Face Spaces container, Vercel frontend, GitHub Actions CI/CD) can be exercised end-to-end.

Purpose

FastAPI inference serving
Hugging Face Hub snapshot_download integration
Frontend / backend deployment validation
CI/CD pipeline validation
Production ML system architecture demonstration

Architecture

Encoder: frozen InceptionV3 (ImageNet weights, 2048-dim features)
Decoder: single Transformer decoder layer, d_model=512, 8 heads
Vocab size: 52 tokens (scaffold) — production target is 15,000 (COCO)
Max caption length: 40 tokens

⚠️ Current limitations

The decoder weights are bootstrap development artefacts generated by a synthetic 10-sentence corpus, not trained on the full COCO dataset. Caption outputs will be incoherent and limited to the 52-token scaffold vocabulary. The encoder is fully functional (real ImageNet weights); only the decoder is untrained.

Future revisions will replace these weights with a model trained on MS COCO 2017 via scripts/train.py and configs/train/stabilized.yaml.

Files

File	Size	SHA-256
`model.h5`	158 MB	`bfe020d920aa2f3d019bf7b5b33904384057372e7c304a9e101a2a59fe110084`
`vocab.json`	566 B	`45ec1704d73046303cbd5292590b2e204b194a2d8345dfb84de81370b4ab4eef`
`vocab.pkl`	3,013 B	`c6700d2bbcd8dc705d6b0ca53e0f8848baa6225e9b3e836036d94ab5accd306c`

Usage

This repo is consumed by the backend via huggingface_hub.snapshot_download:

BACKEND_WEIGHTS_HUB_REPO=apoorvrajdev/captioning-inceptionv3-transformer
BACKEND_WEIGHTS_HUB_REVISION=v1.0.0
BACKEND_WEIGHTS_HUB_FILENAME=model.h5

Downloads last month: -