The implementation of the paper "Feature pyramid-based graph convolutional neural network for graph classification"
-
OS: Ubuntu18.04
- torch==1.4
according to the version of CUDA
replace following ${CUDA} with cpu，cu92，cu101，cu102 or cu110，according to your cuda version
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.4.0+cu101.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.4.0+cu101.html
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-1.4.0+cu101.html
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.4.0+cu101.html
pip install torch-geometric
pip install -r ./requirements.txt
./data
the datasets are formatted as .pt file generated by PYG
For example：
-DD
-raw the raw DD dataset，https://chrsmrrs.github.io/datasets/
-processed Proprietary data pt file for PYG framework generated from raw
-raw_cleaned cleaned dataset of DD，containing only non-isomorphic graphs
-processed_cleaned generated from raw_cleaned
./examples
comes from https://github.com/rusty1s/pytorch_geometric
Some well implemented graph neural networks
./GraphFPN
the different version of our model, the log file record the optimized and improved parts
you can change the dataset in config.py and then train the model
v7-GraphFPN.py is our complete model
./GraphFPN_likepaper
-batch_program.py
Batch training of our models under different random numbers.
-new-v1-GraphFPN
Our complete model
you can change the dataset in config.py
-new-v2-GraphFPN
different pooling method
Ablation based on new-v1-GraphFPN
no atten
Masking Code Lines 127 - 131
no zero
Masking Code Lines 108 - 115
./plotpng
- plt-v1-GraphFPN : visualization model by T-SNE
./result
-extract
save result
for example: v1_new_record_log_210308,txt -> graphfpn.csv
./run
save the model weights
./util
some tool functions
-MeanunPooling.py: contain the unpooling methods
config.py