Data manipultation after set creation

[anthonysirico]

Is there a way to manipulate the data after I have already created the InMemoryDataset? After it was created I tried running dataset[0].y = 1 and nothing happened. Note, that I purposefully left out the y to try to manipulate it after creation.

BLUF: Can I add anything to the dataset using a loop?

for i in dataset.indeces():
    if x <= 0.5:
        dataset[i].y = 1
    else:
        dataset[i].y = 0

[EdisonLeeeee]

You cannot directly manipulate the data via indexing on dataset, as dataset[i] creates a copy of the data object and the operation is not inplaced. Alternatively, you can use the argument transform inherited from InMemoryDataset to manipulate the dataset:

def my_transform(data):
    if x <= 0.5:
        data.y = 1
    else:
        data.y = 0
    return data

dataset = Your_dataset(..., transform=my_transform)

[anthonysirico]

@EdisonLeeeee Also, the technique you suggested is not working unless I did something wrong:

def add_graph_label(data):
    for i in known.indices():
        if i <= len(known)*0.25:
            data.y = torch.tensor(np.asarray([1]), dtype=torch.int64)
        else:
            data.y = torch.tensor(np.asarray([0]), dtype=torch.int64)
    return data
known_ = ImbCircuitDataset('test_imb', split=0.25, transform=add_graph_label)

For some reason, all the y values are 0.

And of course I want to sort the set based on known.data.performance then based on index value give it a y of 0 or 1.

[EdisonLeeeee]

Thanks, you wouldnt happen to have a suggestion on how to sort the data? For example, if I have a set with an attribute called performance, how can I sort the data via dataset.data.performance?

Did you mean that?

performance = torch.tensor([data.performance for data in dataset]) # Suppose data.performance is a scalar
idx = torch.argsort(performance)
dataset = dataset[idx]

Also, the technique you suggested is not working unless I did something wrong:

In your add_graph_label function, data.y is always 0 as there is always a case i> len(known)*0.25 in the loop. Could you clarify what is known.indices() in your dataset? Or you can give a minimal example to make it clear.

[anthonysirico]

known is the name of the dataset representing the training set. I used known.indices() for i since it is range(0,10812) The overall goal is to create the known dataset, sort it based on known.data.performance, then split it 25/75, where the top 25 is 1 and the rest is 0. Then I have the unknown set, which is the test set.

[anthonysirico]

I tried your code for the sorting and it didn't work. Performance is a single value float tensor for each graph.

known[0].performance = tensor([22.7112])

performance = torch.tensor([data.performance for data in known])
idx = torch.argsort(performance)
known_ = known[idx]
known_.data.performance

This is what I ran and did not sort it.

[rusty1s]

Not that dataset.data is not modified, but it instead creates a view on the underlying data with shuffled indices.

print(known[0])
known_ = known[idx]
print(known_[0])

will yield different outputs. If you do not want to create a view but a copy instead, try

known_ = known[idx].copy()