Pytorch implementation for "Diversity Transfer Network for Few-Shot Learning" (deep backbone, on miniImageNet).

We also provide our trainded models.

If you find DTN useful in your research, please consider citing:

@inproceedings{chen2020diversity,
  title={Diversity Transfer Network for Few-Shot Learning.},
  author={Chen, Mengting and Fang, Yuxin and Wang, Xinggang and Luo, Heng and Geng, Yifeng and Zhang, Xinyu and Huang, Chang and Liu, Wenyu and Wang, Bo},
  booktitle={AAAI},
  pages={10559--10566},
  year={2020}
}

Few-shot learning is a challenging task that aims at training a classifier for unseen classes with only a few training examples. The main difficulty of few-shot learning lies in the lack of intra-class diversity within insufficient training samples.

To alleviate this problem, we propose a novel generative framework, Diversity Transfer Network (DTN), that learns to transfer latent diversities from known categories and composite them with support features to generate diverse samples for novel categories in feature space. The learning problem of the sample generation (i.e., diversity transfer) is solved via minimizing an effective meta-classification loss in a single-stage network, instead of the generative loss in previous works.

Besides, an organized auxiliary task co-training(OAT) over known categories is proposed to stabilize the meta-training process of DTN.

Download the dataset from this link, put the images folder in ./miniImageNet/.

Run:

bash make.sh

We provide our trainded models for practitioners to reproduce our results listed in Table 3 of the DTN paper.

python main_DTN.py --checkpoint 'your_checkpoint'

# 5-way 5-shot
python main_DTN.py --N-way 5 --N-shot 5 --evaluate 1 --resume 'your_checkpoint/checkpoint.pth.tar'
# 5-way 1-shot
python main_DTN.py --N-way 5 --N-shot 1 --evaluate 1 --resume 'your_checkpoint/checkpoint.pth.tar'

DTN is released under the MIT license. See LICENSE for additional details.

Horizon Robotics

Pytorch