We use the open source library Guild AI to simplify experiments reproducibility.
You can run every models defined in the yaml file as follows:
$ guild run model:train
or
$ python model/train.py
To reproduce our experimental results
$ guild run planetoid:evaluate100
$ guild run chebnet:evaluate100
$ guild run gcn:evaluate100
$ guild run gat:evaluate100
$ guild compare -o planetoid --csv -> results/planetoid.csv
$ python3 results.py -p results/planetoid.csv
$ guild compare -o chebnet --csv -> results/chebnet.csv
$ python3 results.py -p results/chebnet.csv
$ guild compare -o gcn --csv -> results/gcn.csv
$ python3 results.py -p results/gcn.csv
$ guild compare -o gat --csv -> results/gat.csv
$ python3 results.py -p results/gat.csv
In the test scripts you can test the models and create the t-SNE plot of the learned hidden space.
- train and save a model
$ python3 GAT/train.py --dataset cora --checkpoint-path GAT_ckpt
- test the model and make t-SNE plot
$ python3 GAT/test.py --dataset cora --checkpoint-path GAT_ckpt --tsne
T.B.N.:
- --checkpoint-path specifies the path wherethe model will be saved/loaded;
- --tsne specifies to create the t-SNE plot.