rokeller/photo-search-embedding repository overview

⁠Photo Search by flrx39.net

This image is directly related to the image photo-search⁠. It provides an efficient web server that follows the interface required by the photo-search image to create embeddings for textual searches on photos.

This image therefore replaces the old inefficient embedding server created with Python. Initial tests show that memory consumption of the embedding server contained within this image is about 40% of the memory consumption of the Python-based embedding server.

⁠Arguments

⁠Path to the embedding model

Running the embedding server in the image requires the --model-path parameter. The model path should be set to the root path where the clip-ViT-B-32-multilingual-v1 Model⁠ is made available. A read-only copy of the model is sufficient, and in fact the only files required to be present are:

• ./config.json
• ./tokenizer.json
• ./model.safetensors
• ./2_Dense/model.safetensors

All other files can be ignored / skipped / deleted. This means that the overall size of the model and related files is about 540 - 550 MB.

⁠Host binding

By default, the embedding server binds to 127.0.0.1:8082, but this can easily be changed by specifying the --binding argument. For example, running with --binding 0.0.0.0:8082 will listen on all IPv4 interfaces on port 8082.

⁠Help

Add the --help argument when running the container to get more help.

⁠More Information

More information can be found in the git repo⁠