This is a fork of the original VLMEvalKit: repo link.
Custom interop & eval code are in ./vlmeval/vlm_custom.
Building & restarting the environment. Make sure to check your path to kornia is correct in vlmeval/vlm_custom/kornia_vlm_pyo3/Cargo.toml.
# create environment
source .venv/bin/activate # for example, I'm using venv
# VLMEvalKit setup
pip install -e .
# maturin setup
pip install maturin
# build your rust-python interop
maturin develop --release -m vlmeval/vlm_custom/kornia_vlm_pyo3/Cargo.tomlRunning the validation commands.
# Initialize the environment if haven't yet
source .venv/bin/activate # for example, I'm using venv
# for running the Kornia's version
python run.py --data <dataset> --model SmolVLM-Kornia --reuse
# for running the Python's version with the same feature
python run.py --data <dataset> --model SmolVLM-Py-Comparison --reuse
# for running the Python's default features (rarely used)
python run.py --data <dataset> --model SmolVLM --reuse
# e.g., <dataset>=MMBench_Video_8frame_nopack
python run.py --data <dataset> --model SmolVLM2-Kornia --reuse
python run.py --data <dataset> --model SmolVLM2 --reuse
python run.py --data <dataset> --model SmolVLM2-Py-Comparison --reuseLink to their supported dataset name.