Few-Shot Learning with Graph Neural Networks on CIFAR-100

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.

Dataset

CIFAR-100 (Get the dataset from torchvision)

Requirement

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

OS == Linux system (Ubuntu 16.04LTS)

Execution

training

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

testing

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]

Experiment Result

• 5 way (validation / test)
  

	1-shot	5-shot	10-shot
end-to-end	37.68 % / 35.60 %	61.42 % / 63.20 %	70.05 % / 68.60 %
pretrain and fine-tune	49.66 % / 49.60 %	63.84 % / 59.00 %	69.69 % / 67.20 %
pretrain and fixed	48.30 % / 42.80 %	63.74 % / 65.60 %	68.01 % / 67.20 %

• 20 way (validation / test)
  

	1-shot	5-shot	10-shot
end-to-end	19.85 % / 16.55 %	36.58 % / 35.85 %	38.34 % / 44.05 %
pretrain and fine-tune	20.85 % / 22.95 %	35.50 % / 37.15 %	42.61 % / 41.10 %
pretrain and fixed	22.68 % / 19.25 %	35.66 % / 30.75 %	41.67 % / 42.25 %

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.
   

Contact

Yi-Lin Sung, r06942076@ntu.edu.tw