Appendix.pdf [download]
We listed the details for reproducibility and additional experiments analysis here. The directory is as follows:
-
Details for reproducibility:
- Experimental environments.
- The hyper-parameters used in LTD/FT.
- The settings for teacher/student models and other distillation frameworks.
- Some brief comments on data preparation.
-
Additional experiments analysis:
- Balance hyper-parameter analysis.
- Performance under different training ratios.
- Generalization gap analysis.
- Original results of baselines.
- dgl==0.6.0
- Keras==2.4.3
- numpy==1.19.2
- optuna==2.6.0
- pandas==1.2.4
- python==3.8.8
- scikit-learn==0.19.2
- scipy==1.6.2
- sklearn==0.0
- tabulate==0.8.9
- torch==1.8.1
- torchvision==0.9.1
Because the A-Computers dataset is too large, we did not put it in the project file. The dataset can be downloaded from the following link and placed in the directory data/npz:
https://www.dropbox.com/s/26zd460xn4u6gmn/amazon_electronics_computers.npz?dl=0
Run student model:
python spawn_worker.py --dataset XXXX --teacher XXX --student XXX
For example, if you want train GCN on cora dataset:
python spawn_worker.py --dataset cora --teacher GCN --student GCN
We fix hyper-parameters setting as reported in our paper, and you also can use --automl
to search hyper-parameters with the help of Optuna.