szymonrucinski/graph-convolutional-networks

Train your own GCN model using the latest pytorch-geometric 📉 to solve image classification problem 🧠

Spline graph networks

All experiments performed on a computer with the following configuration.

CPU	GPU	RAM
AMD Ryzen 5 2600	GTX 1660 Super OC 6Gb	32 Gb DDR4

Category 1 experiment

The task is to prepare a program and conduct an experiment to test the effectiveness of a selected graph neural network model in an inductive approach using the graph equivalent of the MNIST set.

Experiment 1.1

Network architecture
Net( (conv1): SplineConv(1, 32, dim=2) (conv2): SplineConv(32, 64, dim=2) (fc1): Linear(in_features=64, out_features=128, bias=True) (fc2): Linear(in_features=128, out_features=10, bias=True) )
Optimizer
Adam

Graph 1.1 - Visualization of the analyzed graph and image from the MNIST set.

Figure 1.2 - Percentage effectiveness for the training set

Graph 1.3 - Graph showing changes in error (objective function) during learning

The final accuracy achieved for the test set is 94.2%.

Experiment 1.2

Network architecture
GCN2( (conv1): SplineConv(1, 64, dim=2) (conv2): SplineConv(64, 128, dim=2) (fc1): Linear(in_features=128, out_features=256, bias=True) (fc2): Linear(in_features=256, out_features=10, bias=True) )
Optimizer
Adam

Figure 1.4 - Percentage effectiveness for the training set

Graph 1.5 - Graph showing changes in error (objective function) during learning

The final accuracy achieved for the test set is 94.6%.

Reference results of obtained % classification error for a given set published in the literature for other network architectures:

GCGP	PNCNN	Dynamic Reduction Network
4.2	1.24	0.95

Category 2 experiment

The task is to prepare a program and conduct an experiment to test the effectiveness of the selected graph neural network model in the transduction case approach using an example from one of the elements in the Planetoid set.

Experiment 2.1

Network architecture
GCN( (conv1): GCNConv(1433, 32) (conv2): GCNConv(32, 7) )
Optimizer
Adam

graphic 2.1 - Visualization of elements in a set

Graph 2.2 - Percentage performance for the training set

Graph 2.3 - Graph showing changes in error (objective function) during learning

The final accuracy achieved is 78.9%.

Experiment 2.2

Network architecture
GCN( (conv1): GCNConv(1433, 512) (conv2): GCNConv(512, 7) )
Optimizer
Adam

Figure 2.4 - Percentage effectiveness for the training set

Graph 2.5 - Graph showing changes in error (objective function) during learning

The final accuracy achieved is 79.6%.

Reference % accuracy results obtained for a given set published in the literature for other network architectures:

GAT	SPLINECNN	SSP
83,00 %	89,48 %	90,160%