This repository is the official PyTorch implementation of the experiments in the following paper:
Junran Wu, Jianhao Li, Yicheng Pan*, Ke Xu*. Structural Optimization Makes Graph Classification Simpler and Better.
Install PyTorch following the instuctions on the official website. The code has been tested over Python 3.7.4, PyTorch 1.3.1 and CUDA 10.1.
Experiment datasets are adopted from GIN-repo.
- Unfold data
tar -xvzf datasets.tar.gz
- Load graphs from original dataset file for better code.
python load_graph.py
- Structural optimization (Transform graphs to encoding trees).
python structural_optimization.py
Default 10-fold corss-validation is conducted with cross_val_score in scikit-learn.
usage: treeKernel.py [-h] [-d DATASET] [-k {2,3,4,5}]
Tree kernel with SVM for whole-tree classification
optional arguments:
-h, --help show this help message and exit
-d DATASET, --dataset DATASET
name of dataset (default: None, search all datasets)
-k {2,3,4,5}, --tree_depth {2,3,4,5}
the depth of coding tree (default: None, search all depthes)
The default parameters are not the best performing-hyper-parameters used to reproduce our results in the paper. Hyper-parameters need to be specified through the commandline arguments. Please refer to our paper for the details of how we set the hyper-parameters.
To learn hyper-parameters to be specified, please see
usage: encodingTreeLearning.py [-h] [-d DATASET] [-k TREE_DEPTH]
[-b BATCH_SIZE] [-e EPOCHS] [-lr LEARNING_RATE]
[--iters_per_epoch ITERS_PER_EPOCH] [-s SEED]
[-fi FOLD_IDX] [-lm NUM_MLP_LAYERS]
[-hd HIDDEN_DIM] [-fd FINAL_DROPOUT]
[-tp {root,sum,average}] [-gd]
[--device DEVICE]
PyTorch tree convolutional neural net for whole-tree classification
optional arguments:
-h, --help show this help message and exit
-d DATASET, --dataset DATASET
name of dataset (default: MUTAG)
-k TREE_DEPTH, --tree_depth TREE_DEPTH
the depth of coding tree (default: 2)
-b BATCH_SIZE, --batch_size BATCH_SIZE
input batch size for training (default: 32)
-e EPOCHS, --epochs EPOCHS
number of epochs to train (default: 350)
-lr LEARNING_RATE, --learning_rate LEARNING_RATE
learning rate (default: 0.01)
--iters_per_epoch ITERS_PER_EPOCH
number of iterations per each epoch (default: 50)
-s SEED, --seed SEED random seed for splitting the dataset into 10
(default: 0)
-fi FOLD_IDX, --fold_idx FOLD_IDX
the index of fold in 10-fold validation. Should be less then 10.
-lm NUM_MLP_LAYERS, --num_mlp_layers NUM_MLP_LAYERS
number of layers for MLP EXCLUDING the input one
(default: 2). 1 means linear model.
-hd HIDDEN_DIM, --hidden_dim HIDDEN_DIM
number of hidden units (default: 64)
-fd FINAL_DROPOUT, --final_dropout FINAL_DROPOUT
final layer dropout (default: 0.5)
-tp {root,sum,average}, --tree_pooling_type {root,sum,average}
Pooling for over nodes in a tree: root, sum or average
-gd, --global_degree add the degree of nodes to features
--device DEVICE which gpu to use if any (default: 0)
An example of training process is as follows:
python encodingTreeLearning.py -d PTC -k 2 -b 128 -e 100 -hd 32 -tp sum -gd
The cross-validation in our paper only uses training and validation sets (no test set) due to small dataset size. Specifically, after obtaining 10 validation curves corresponding to 10 folds, we first took average of validation curves across the 10 folds (thus, we obtain an averaged validation curve), and then selected a single epoch that achieved the maximum averaged validation accuracy. Finally, the standard devision over the 10 folds was computed at the selected epoch.