GNN using PyTorch

Spektral is an easy library used to create graph neural network for graph classification. A graph is a mathematical object that represents relations between entities. We call the entities "nodes" and the relations "edges".

Colab Code

Run this code in colab to install spektral library.

!pip install spektral

Import all the neccessary packages needed

import numpy as np
import tensorflow as tf
import spektral

Datasets:

We will be using Cora dataset. It contains 2708 scientific publications. It has seven different classes. The dataset contains 5429 links referencing to the different citations used by the scientific publication.

The raw files can be downloaded from here.

data= spektral.datasets.citation.Citation(name='Cora', random_split=False, normalize_x=False, dtype=np.float32)

Split the dataset into train, validation and test.

train_mask, val_mask, test_mask = data.mask_te, data.mask_va, data.mask_te

Check the graph info.

data.graphs

Output

[Graph(n_nodes=2708, n_node_features=1433, n_edge_features=None, n_labels=7)]

Get Adjacent nodes present in the graph.

adj = data.graphs[0].a
print(adj)

Output:

(0, 633)  1.0   
(0, 1862) 1.0   
(0, 2582) 1.0   
(1, 2)	  1.0   
(1, 652)  1.0
...

Get the features of each nodes present in the graph.

features = data.graphs[0].x
features

Output:

array([[0., 0., 0., ..., 0., 0., 0.],        
       [0., 0., 0., ..., 0., 0., 0.],        
       [0., 0., 0., ..., 0., 0., 0.],        
       ...,        
       [0., 0., 0., ..., 0., 0., 0.],        
       [0., 0., 0., ..., 0., 0., 0.],        
       [0., 0., 0., ..., 0., 0., 0.]], dtype=float32)

Checking the shape of the feature vector.

features.shape

Output:

(2708, 1433)

Get all the labels from the graph.

labels = data.graphs[0].y
labels

Output:

array([[0., 0., 0., ..., 0., 0., 0.],        
       [0., 0., 0., ..., 1., 0., 0.],       
       [0., 0., 0., ..., 1., 0., 0.],        
       ...,        
       [0., 0., 0., ..., 0., 0., 0.],        
       [0., 0., 0., ..., 0., 0., 0.],        
       [0., 0., 0., ..., 0., 0., 0.]], dtype=float32)

Check the label shape.

labels.shape

Output:

(2708, 7)

The adjacent matrix need to be converted from sparse matrix to dense so that it can be passed to the GNN model for training.

adj = adj.todense() + np.eye(adj.shape[0])
adj = adj.astype('float32')

Let's check number of samples present in training samples, validation samples and test samples.

np.sum(train_mak), np.sum(val_mask), np.sum(test_mask)

Output:

(1000, 500, 1000)

Next section will be on Graph Neural Network.