Source code for our ICLR 2020 paper: FEW-SHOT LEARNING ON GRAPHS VIA SUPER-CLASSES BASED ON GRAPH SPECTRAL MEASURES. Associated Blogpost can be found here.
Please create a virtual environment for smoother functioning and to avoid any dependency issues. Please refer to Managing Virtual Environments for details on creating virtual environment. After activating the virtual environment, run the following command to install all the required libraries and modules(Note: the requirements.txt file contains some additional libraries which you may or may not need) -
`pip install -r requirements.txt`
The datasets directory contains the graphs for the datasets - Letter-High and TRIANGLES used for this paper. Inside each dataset sub-directory, there is a file named class_prototype_numbers.json which contains the precomputed class prototypes as given in the paper. The file make_class_prototypes.py contains the code to create the class prototype graphs. The train_test_classes.json contains the training-testing class splits used. The sub-directory json_format contains the graphs in json format.
The src directory contains the source code for this paper. The main.py file contains the initiation code, starting with the argument parser. The important arguments are -
'--dataset_name' : Name of the dataset
'--num_layers' : Number of layers in GIN model
'--hidden_dim' : Number of dimensions in GIN's MLP layers
'--graph_pooling_type' : Pooling over nodes to get graph embedding: sum or average
'--neighbor_pooling_type' : Pooling over neighboring nodes: sum, average or max
'--num_gat_layers' : Number of GAT layers
'--gat_heads' : Number of heads per GAT layer
'--batch_size' : size of mini-batch
'--n_shot' : The shot scenario to evaluate upon
'--knn_value' : The number of neighbors per graph representation node in super-graph
'--train_clusters' : The number of super-classes
Rest all the arguments are self-explanatory.
To evaluate the model for 20-shot on TRIANGLES datasets run -
python3 main.py --dataset_name TRIANGLES --batch_size 64 --knn_value 2
An example usage for the above has been provided in the jupyter notebook: src/example.ipynb .
The src/checkpoints directory stores the trained weights for the model. Run the bash file clear_checkpoints.sh to clear existing checkpoints. The dataloader.py file contains the class for loading data and creating splits for fine-tuning as well as testing. The base code for the files graphcnn.py, mlp.py and util.py have been taken from the original implementation of the GIN paper and further modified for our purpose.
Please cite the following paper if you use this code in your work.
@inproceedings{chauhan2020fewshot,
title=FEW-SHOT LEARNING ON GRAPHS VIA SUPER-CLASSES BASED ON GRAPH SPECTRAL MEASURES,
author={Jatin Chauhan and Deepak Nathani and Manohar Kaul},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=Bkeeca4Kvr}
}