This is the implementation code of KGTN paper:
@inproceedings{chen2020knowledge,
title={Knowledge Graph Transfer Network for Few-Shot Recognition.},
author={Chen, Riquan and Chen, Tianshui and Hui, Xiaolu and Wu, Hefeng and Li, Guanbin and Lin, Liang},
booktitle={AAAI},
pages={10575--10582},
year={2020}
}
- Python==2.7
- Pytorch==1.0.1
- Torchvision==0.2.1
Our trained feature extractor is available at onedrive. Download the feature extractor and place it into directory checkpoints/ResNet50_sgm/
. If do so, go to Step-2.
./scripts/TrainFeatureExtractor.sh
./scripts/SaveFeature.sh
For semantic similarity knowledge graph:
./scripts/KGTN_wv_InnerProduct.sh [GPU_ID]
or
For category hierarchy knowledge graph:
./scripts/KGTN_wordnet.sh [GPU_ID]
*We take the model with a plain fully-connected layer as the baseline model, which is used in Low-shot Visual Recognition by Shrinking and Hallucinating Features.
Top5 Novel
Knowledge Graph | (n-shot) 1 | 2 | 5 | 10 |
---|---|---|---|---|
None(baseline) | 53.7 | 67.5 | 77.8 | 82.2 |
category hierarchy | 60.3 | 69.9 | 78.3 | 82.3 |
semantic similarity | 62.1 | 71.0 | 78.5 | 82.4 |
Top5 All
Knowledge Graph | (n-shot) 1 | 2 | 5 | 10 |
---|---|---|---|---|
None(baseline) | 61.7 | 72.1 | 80.2 | 83.5 |
category hierarchy | 67.0 | 74.5 | 80.8 | 83.3 |
semantic similarity | 68.4 | 75.2 | 80.9 | 83.4 |
Feel free to contact us sysucrq@gmail.com & tianshuichen@gmail.com & wuhefeng@gmail.com if you have any question.