This is the implementation for the KGTN-ens paper, published in Applied Intelligence.
We propose KGTN-ens, a framework to incorporate multiple knowledge graph embeddings at a small cost. There are many real-world scenarios in which the amount of data is severely limited. Prior knowledge can be used to tackle this task. The purpose of this study is to investigate the possibility of combining multiple knowledge sources. We evaluate it with different embeddings in a few-shot image classification task. We also construct a new knowledge source – Wikidata embeddings – and evaluate it with KGTN and KGTN-ens. With ResNet50, our approach outperforms KGTN in terms of the top-5 accuracy on the ImageNet-FS dataset for the majority of tested settings. For
$k \in {1, 2, 5, 10}$ respectively, we obtained +0.63/+0.58/+0.43/+0.26 pp. (novel classes) and +0.26/+0.25/+0.32/–0.04 pp. (all classes).
If you use this work, please cite it accordingly:
Filipiak, D., Fensel, A. & Filipowska, A. KGTN-ens: Few-Shot Image Classification with Knowledge Graph Ensembles. Applied Intelligence (2024). https://doi.org/10.1007/s10489-023-05129-8
or, in BibTeX:
@article{filipiak2024kgtnens,
title={KGTN-ens: Few-Shot Image Classification with Knowledge Graph Ensembles},
author={Filipiak, Dominik and Fensel, Anna and Filipowska, Agata},
journal={Applied Intelligence},
year={2024},
doi = {10.1007/s10489-023-05129-8}
}The code is based on the KGTN implementation, so running it is almost the same as in described in KGTN readme.
The only difference is the script location and parameters (see e.g. KGTN2/scripts/ResNet50_sgm/ for ResNet50 experiments).
Usually, the first arguments denotes the GPU number to run the script on, but make sure you read the script before running it.
The embeddings are in the repository in KnowledgeGraphMatrix, but if you want to check the code for their creation (and see the addtional experiments), please check the notebooks/ directory.
The easiest way will be through using this docker image.
The experiments were run on Python 3.6.9.
root@2e35de100a8f:/tf/KGTN2# nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2018 NVIDIA Corporation
Built on Sat_Aug_25_21:08:01_CDT_2018
Cuda compilation tools, release 10.0, V10.0.130
root@2e35de100a8f:/tf/KGTN2# nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.89.02 Driver Version: 525.89.02 CUDA Version: 12.0 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... On | 00000000:19:00.0 Off | N/A |
| 27% 25C P8 20W / 260W | 1MiB / 11264MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 1 NVIDIA GeForce ... On | 00000000:1A:00.0 Off | N/A |
| 27% 26C P8 20W / 260W | 1MiB / 11264MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 2 NVIDIA GeForce ... On | 00000000:67:00.0 Off | N/A |
| 27% 28C P8 7W / 260W | 1MiB / 11264MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 3 NVIDIA GeForce ... On | 00000000:68:00.0 Off | N/A |
| 27% 32C P8 21W / 260W | 1MiB / 11264MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
root@2e35de100a8f:/tf/KGTN2# uname-a
Linux 2e35de100a8f 5.10.0-21-amd64 #1 SMP Debian 5.10.162-1 (2023-01-21) x86_64 x86_64 x86_64 GNU/Linux