PyTorch implementation for Kipf & Welling (2017). Semi-Supervised Classification with Graph Convolutional Networks. ICLR. The data comes from the first author's repo here.
networkx, numpy, pickle, torch (0.3.1), scipy.
To train the model, run run_train.sh with correctly chosen parameters as per your need.
Example output on the Cora dataset:
Iteration 50:
Train loss = 1.9893128 (at step 50)
Train/Valid/Test accuracy: 0.2929 | 0.1800 | 0.1790
Iteration 100:
Train loss = 1.8589772 (at step 100)
Train/Valid/Test accuracy: 0.8143 | 0.5220 | 0.5170
Iteration 150:
Train loss = 1.7107995 (at step 150)
Train/Valid/Test accuracy: 0.9429 | 0.6540 | 0.6930
Iteration 200:
Train loss = 1.5520712 (at step 200)
Train/Valid/Test accuracy: 0.9714 | 0.7040 | 0.7420
Iteration 250:
Train loss = 1.3880322 (at step 250)
Train/Valid/Test accuracy: 0.9714 | 0.7200 | 0.7550
Iteration 300:
Train loss = 1.2250416 (at step 300)
Train/Valid/Test accuracy: 0.9786 | 0.7480 | 0.7670
Iteration 350:
Train loss = 1.0695306 (at step 350)
Train/Valid/Test accuracy: 0.9857 | 0.7520 | 0.7780
Iteration 400:
Train loss = 0.9264522 (at step 400)
Train/Valid/Test accuracy: 0.9857 | 0.7660 | 0.7850
Iteration 450:
Train loss = 0.79867464 (at step 450)
Train/Valid/Test accuracy: 0.9857 | 0.7660 | 0.7850
Iteration 500:
Train loss = 0.68706673 (at step 500)
Train/Valid/Test accuracy: 0.9857 | 0.7700 | 0.7920
Iteration 550:
Train loss = 0.5910516 (at step 550)
Train/Valid/Test accuracy: 0.9857 | 0.7760 | 0.7980
Iteration 600:
Train loss = 0.5092095 (at step 600)
Train/Valid/Test accuracy: 0.9857 | 0.7760 | 0.7990
Iteration 650:
Train loss = 0.43980482 (at step 650)
Train/Valid/Test accuracy: 1.0000 | 0.7760 | 0.7990
Training accuracy reaches 1.0, training done!
Accuracy within 2% of that is reported in the original work. The model is sensitive to hyperparameter tuning.