Transform Data Attribute in a Dataset

[kyriakosschwarz]

I created a CustomDataset based on InMemoryDataset. When I get the first Data object from the dataset it looks like this:

data0 = dataset[0]  # Get the first graph object.

data0

PairData(edge_index_a=[2, 26], x_a=[13, 9], edge_attr_a=[26, 3],
    edge_index_b=[2, 76], x_b=[35, 9], edge_attr_b=[76, 3],
    id=[1], y=[1], exp=[1, 908])

Lets say this dataset contains 1000 samples, then the exp attribute of this dataset would originate from a [1000, 908] matrix.

So, two questions:

1. Is there a way to easily reconstruct the [1000, 908] exp matrix from all the exp [1,908] vectors of all the 1000 samples?
2. How can I apply any function (e.g. rescaling function) to the exp attribute of all the PairData objects and replace their current exp vectors?

[rusty1s]

1. The concatenation of features will always be in order. If you want to do it by hand you can easily do

exp = torch.cat([data.exp for data in dataset], dim=0)

2. This is best done via transforms.