Faiss + Pytorch + FastAPI + LMDB
Uses CLIP ViT B/16, aQE
Supported operations: add new image, delete image,find similar images by text query, find similar images by image file, find similar images by image id
You should install torch yourself https://pytorch.org/get-started/locally/.
pip3 install -r requirements.txt
generate_image_text_features.py ./path_to_img_folder
-> generates features
--use_int_filenames_as_id=0
- images get sequential ids
--use_int_filenames_as_id=1
- image id is parsed from filename ("123.jpg" -> 123)
train.py
-> trains index
add_to_index.py
-> adds features from lmdb to index (if ./trained.index is not found, defaults to Flat Index)
image_text_features_web.py
-> web microservice
GET_FILENAMES=1 image_text_features_web.py
-> when searching, include filename in search results
Docker (onnx, web inference only):
cd to ./onnx_web:
download clip_textual.onnx and clip_visual.onnx from here
build image -
docker build -t qwertyforce/onnx_image_text_features_web:1.0.0 --network host -t qwertyforce/onnx_image_text_features_web:latest ./
cd to main directory
run interactively -
docker run -ti --rm -p 127.0.0.1:33338:33338 --network=ambience_net --mount type=bind,source="$(pwd)"/data,target=/app/data --name onnx_image_text_features_web qwertyforce/onnx_image_text_features_web:1.0.0
run as daemon
docker run -d --rm -p 127.0.0.1:33338:33338 --network=ambience_net --mount type=bind,source="$(pwd)"/data,target=/app/data --name onnx_image_text_features_web qwertyforce/onnx_image_text_features_web:1.0.0
Docker (w pytorch):
in main directory:
build image -
docker build -t qwertyforce/image_text_features_web:1.0.0 --network host -t qwertyforce/image_text_features_web:latest ./
run interactively -
docker run -ti --rm -p 127.0.0.1:33338:33338 --network=ambience_net --mount type=bind,source="$(pwd)"/data,target=/app/data --name image_text_features_web qwertyforce/image_text_features_web:1.0.0
run as daemon
docker run -d --rm -p 127.0.0.1:33338:33338 --network=ambience_net --mount type=bind,source="$(pwd)"/data,target=/app/data --name image_text_features_web qwertyforce/image_text_features_web:1.0.0