This is the PyTorch-0.4.0 implementation of few-shot learning on CIFAR-100 with graph neural networks (GNN). And the codes is on the basis of following paper/github/course.

• FEW-SHOT LEARNING WITH GRAPH NEURAL NET-WORKS
• https://github.com/vgsatorras/few-shot-gnn
• Course: Deep Learning for Computer Vision, Spring 2018, National Taiwan University

Besides directly end-to-end training both CNN-embedding layers and graph convolution layers, I also tried different training methods and got better results sometimes.

1. Pretrain the CNN-embedding layers by do classification on CIFAR-100 (excluding few-shot data), and fine-tune them while training graph convolution layer.
2. Pretrain the CNN-embedding layers but fixing them while training graph convolution layer.

CIFAR-100 (Get the dataset from torchvision)

python == 3.6
pytorch == 0.4.0
torchvision == 0.2.1
scikit-image == 0.14.0

OS == Linux system (Ubuntu 16.04LTS)

Train for N way M shots

python3 main.py --data_root 'data' --nway N --shots M

Pretrain CNN-embedding layer and train for N way M shots

python3 main.py --data_root 'data' --nway N --shots M --pretrain

Pretrain CNN-embedding layer and train for N way M shots while fixing the CNN-embedding layer

python3 main.py --data_root 'data' --nway N --shots M --pretrain --freeze_cnn

Test for N way M shot:
python main.py --todo 'test' --data_root 'data' --nway N --shots M --load --load_dir [path for folder saving model.pth]

• 5 way (validation / test)
  

• 20 way (validation / test)
  

1. The seed for choosing few-shot class is 1.
   
2. Only run each experiment for 1 time. Running it for multiple times can get more convincing results.
   
3. Note that if increasing the layers of the model too much, end-to-end training might fail. Pretraining CNN-embedding layers can cure this problem.
   

Yi-Lin Sung, [email protected]