AI model service for EFISS

TorchServe API

API: https://ai.efiss.tech/predictions/image-retrieval-v1.0

Request:

{
    "top_k": 10,
    "diversity": 1,  // integer, range from 1-20
    "image": "<base64 encoded image>"
}

Response

{
    "index_database_version": "1.2.0", // version of the index database
    "relevant": [
        "data/abc.com/69_abc_com.jpg",
        "data/abc.com/42_abc_com.jpg"
    ],  // file path of relevant images, sorted by relevance from most relevant to least relevant
    "distances": [
        2,
        50
    ],
    "cropped_image": "<base64 encoded image>" // cropped image
}

Reindex the database

1. Update version by editing .env

cp .env.example .env

2. Just run

bash indexer/reindex.sh

It will gather images data, index them, dockerize them, push to GCR, and re-run the GPU service locally.

Build

1. Build CPU docker image

docker build -t efiss-ai:latest \
    -t efiss-ai:1.0.0-cpu \
    -t asia-southeast1-docker.pkg.dev/efiss-394203/efiss/efiss-ai:latest \
    -t asia-southeast1-docker.pkg.dev/efiss-394203/efiss/efiss-ai:1.0.0-cpu \
    -f searcher/Dockerfile .

docker push asia-southeast1-docker.pkg.dev/efiss-394203/efiss/efiss-ai:1.0.0-cpu

2. Build GPU docker image

docker build -t efiss-ai:latest-cuda \
    -t efiss-ai:1.0.0-cuda \
    -t asia-southeast1-docker.pkg.dev/efiss-394203/efiss/efiss-ai:latest-cuda \
    -t asia-southeast1-docker.pkg.dev/efiss-394203/efiss/efiss-ai:1.0.0-cuda \
    -f searcher/Dockerfile.cuda .