architecture.md

Architectures

GNN architectures can be imported from a list of implemented ones, but new ones can also be defined by instantiating.

1. Literature Implementations
2. Defining New Architectures
3. Quick Customization
4. Wrapping Keras Layers

Literature Implementations

The following architectures are currently implemented: GCN, GCNII, APPNP. You can import them from the package's top level. For example, instantiating the APPNP architecture is as simple as calling:

import gnntf
G, features, num_classes = ...
gnn = gnntf.APPNP(gnntf.graph2adj(G), features, num_classes=num_classes)

Contrary to general-purpose machine learning frameworks, you need to provide the graph (e.g. starting from a networkx graph G) and node features during architecture definition. This means that the architecture is defined for specific input data.

Defining New Architectures

Custom GNNs can be defined by extending the GNN class and adding layers during the constructor method. Typical Neural Network layers can be found in the module core.gnn.nn.layers. For example, a traditional perceptron with two dense layers and dropout to be used for classification can be defined per the following code.

import gnntf
import tensorflow as tf

class CustomGNN(gnntf.GNN):
    def __init__(self, 
                 G: tf.Tensor,
                 features: tf.Tensor, 
                 hidden_layer: int = 64, 
                 num_classes: int = 3, 
                 **kwargs):
        super().__init__(G, features, **kwargs)
        self.add(gnntf.Dropout(0.5))
        self.add(gnntf.Dense(hidden_layer, activation=tf.nn.relu))
        self.add(gnntf.Dropout(0.5))
        self.add(gnntf.Dense(num_classes,  regularize=False))

⚠️ The dropout argument is applied for the time being on layer outputs.

In addition to typical functionalities provided by neural network libraries, we also provide flow control functionality on the layer level that removes the need of understanding tensorflow primitives at all by using Branch and Concatenate layers.

Quick Customization

You can simplify architecture definition by instantiating it via gnntf.GNN(...) and adding layers afterwards. In fact, you can add layers after training just fine or even remove top layers by calling

Wrapping Keras Layers

You can turn Keras layers by wrapping them with the gnntf interface by calling layer = gnntf.Wrap(layer). This can be added to architectures.