Add an ipython notebook tutorial to introduce the new torch_geometric.nn.aggr package

[lightaime]

🚀 The feature, motivation and pitch

We would like to add an ipython notebook to demonstrate the new flexible interface of the new torch_geometric.nn.aggr package with MessagePassing. In the new integration, the following functionality is applicable:

# Original interface with string type as aggregation argument
class MyConv(MessagePassing):
    def __init__(self):
        super().__init__(aggr="mean")

# Use a single aggregation module as aggregation argument
class MyConv(MessagePassing):
    def __init__(self):
        super().__init__(aggr=MeanAggregation())

# Use a list of aggregation strings as aggregation argument
class MyConv(MessagePassing):
    def __init__(self):
        super().__init__(aggr=['mean', 'max', 'sum', 'std', 'var'])

# Use a list of aggregation modules as aggregation argument
class MyConv(MessagePassing):
    def __init__(self):
        super().__init__(aggr=[
          MeanAggregation(),
          MaxAggregation(),
          SumAggregation(),
          StdAggregation(),
          VarAggregation(),
          ])

# Use a list of mixed modules and strings as aggregation argument
class MyConv(MessagePassing):
    def __init__(self):
        super().__init__(aggr=[
          'mean',
          MaxAggregation(),
          'sum',
          StdAggregation(),
          'var',
          ])

# Define multiple learnable aggregations with keyword arguments
class MyConv(MessagePassing):
    def __init__(self):
        super().__init__(aggr=['softmax', 'softmax', 'softmax'],
        aggr_kwargs = dict(aggrs_kwargs=[
                            dict(t=0.1, learn=True),
                            dict(t=1, learn=True),
                            dict(t=10, learn=True)]))

# Define multiple aggregations with `MultiAggregation` module
class MyConv(MessagePassing):
    def __init__(self):
        super().__init__(aggr=MultiAggregation([
          SoftmaxAggregation(t=0.1, learn=True),
          SoftmaxAggregation(t=1, learn=True),
          SoftmaxAggregation(t=10, learn=True)]))

[rusty1s]

Thank you @lightaime. I will go over the notebook. In the meantime, can you add a link to it to our documentation Colab page?

[lightaime]

Sure, added a pull request here: #4927.

[rusty1s]

Thanks!