Large Model-Driven Balanced Multimodal Knowledge Graph Completion

Data preparation

We use the MMKG datasets proposed in MMRNS.

We use the SDXL and CLIP provided by Hugging Face.

You need to download the embeddings and place them in the "./embeddings".

Dependencies

• torch==2.2.2
• numpy==2.1.2
• scikit-learn==1.6.1
• tqdm==4.64.1
• Our code is based on OpenKE, an open-source KGC project. You can refer to the OpenKE repo to build the environment.

Train and Evaluation

You should run the script mmkgc/make.sh to ensure that the release/Base.so file is compatible with your environment.

You can use the shell scripts to conduct the experiments.

python Main_LBMKGC.py -dataset MKG-W -margin 16 -epoch 1000 -save MKG-W-checkpoint -learning_rate 1e-5

python Main_LBMKGC.py -dataset MKG-Y -margin 24 -epoch 1250 -save MKG-Y-checkpoint -learning_rate 1e-5

python Main_LBMKGC.py -dataset DB15K -margin 12 -epoch 1250 -save DB15K-checkpoint -learning_rate 2e-5