Explanable Graph Neural Architecture Search via Monte-Carlo tree search

This is an implementation of a graph NAS method with Monte-Carlo tree search.

Setup

• python=3.9
• CUDA11.1
• pytorch=1.9.0
• torchvision=0.10.0

Run

python3 main.py -train_size $train_size -val_size $val_size -test_size $test_size -epoch $epoch -dataset_name $dataset_name

• $train_size set train data size
• $val_size set validation data size
• $test_size set test data size
• $epoch set the number of epochs
• $dataset_name set the detaset name
• $num_models set the number of architectures
• $search threshold set the visit times to generate child nodes
• $mcts_score_sqrt set c of ucs
• $eval_type set the mode of search algorithm (max or avg)
• $eval_mode set the mode of evaluation (acc or auc)

ex) python3 main.py -train_size 0.6 -val_size 0.2 -test_size 0.2 -epoch 500 -trial 5 -num_models 1000 -search_threshold 10 -mcts_score_sqrt 2 -eval_type avg -eval_mode acc -dataset_name cora

CSV files are generated at results directory.