Simcoder

A utility for encoding images using the SimCLR2 pretrained model weigths.

Installation

The tool is designed to run inside a docker container and requires a build step to use. There are a set of make targets that automate these steps. They are in the Makefile. You may need to modify them for your machine.

Here is an example of installation:

git clone https://github.com/davemor/simcoder.git
cd simcoder
make docker_image

Usage

To run a container and execute the program, there is a handy make target called run_batch.

make run_batch

This assumes that a host directory containing images is mapped to /input inside the container and that the directory to save the embeddings is mapped to /output. You will need to remap these directories in order to use it on your specific machine. This can be done inside the run_batch target in the Makefile.

The program accepts two options:

• --format - specifies the output format which can be csv, npy, or mat.
• --chunksize - the number of image embeddings per output file.