Knowledge Graph Transfer Network for Few-Shot Recognition

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}
}

Preparation

• Python==2.7
• Pytorch==1.0.1
• Torchvision==0.2.1

Running the code

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.

1.Train feature extractor

./scripts/TrainFeatureExtractor.sh 

2. Save feature

./scripts/SaveFeature.sh

3. Train the Knowledge Graph Transfer Network (KGTN)

For semantic similarity knowledge graph:

./scripts/KGTN_wv_InnerProduct.sh [GPU_ID]

or

For category hierarchy knowledge graph:

./scripts/KGTN_wordnet.sh [GPU_ID]

Performance

*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

Contributing

Feel free to contact us sysucrq@gmail.com & tianshuichen@gmail.com & wuhefeng@gmail.com if you have any question.