We provide a PyTorch implementation of DN4 for few-shot learning. If you use this code, please cite:

Revisiting Local Descriptor based Image-to-Class Measure for Few-shot Learning.
Wenbin Li, Lei Wang, Jinglin Xu, Jing Huo, Yang Gao and Jiebo Luo. In CVPR 2019.

• Linux
• Python 3.8
• Pytorch 1.7.0
• GPU + CUDA CuDNN
• pillow, torchvision, scipy, numpy

• Clone this repo:

git clone https://github.com/WenbinLee/DN4.git
cd DN4

• Install PyTorch 1.7.0 and other dependencies.

Caltech-UCSD Birds-200-2011, Standford Cars, Standford Dogs, miniImageNet and tieredImageNet are available at Google Drive and 百度网盘(提取码：yr1w).

• Train a 5-way 1-shot model based on Conv64:

python Train_DN4.py --dataset_dir ./path/to/miniImageNet --data_name miniImageNet --encoder_model Conv64F_Local --way_num 5 --shot_num 1

• Train a 5-way 1-shot model based on ResNet12:

python Train_DN4.py --dataset_dir ./path/to/miniImageNet --data_name miniImageNet --encoder_model ResNet12 --way_num 5 --shot_num 1

• Test the model (specify the dataset_dir, encoder_model, and data_name first):

python Test_DN4.py --resume ./results/SGD_Cosine_Lr0.05_DN4_Conv64F_Local_Epoch_30_miniImageNet_84_84_5Way_1Shot/ --encoder_model Conv64F_Local

(Compared to the originally reported results in the paper. * denotes that ResNet256F is used.)

• The results on the miniImageNet dataset reported in the orinigal paper:

If you use this code for your research, please cite our paper.

@inproceedings{DN4_CVPR_2019,
  author       = {Wenbin Li and
                  Lei Wang and
                  Jinglin Xu and
                  Jing Huo and
                  Yang Gao and
                  Jiebo Luo},
  title        = {Revisiting Local Descriptor Based Image-To-Class Measure for Few-Shot Learning},
  booktitle    = {{IEEE} Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages        = {7260--7268},
  year         = {2019}
}