FixedPoints: sampling with a single point also resamples y

🐛 Bug

FixedPoints resamples also y if the pos-matrix contains only a single point.

It might be that this is intended behaviour - then don’t worry just close this issue. I can (and probably should) take care of samples with only one point anyways before learning

To Reproduce

import torch
from torch_geometric.data.data import Data
from torch_geometric.transforms.fixed_points import FixedPoints

pointcloud = Data(pos=torch.randn(100, 3), y=torch.tensor([1]))
single_point = Data(pos=torch.randn(1, 3), y=torch.tensor([1]))

random_select_n_points = FixedPoints(num=10)

print(random_select_n_points(pointcloud))
    # Data(pos=[10, 3], y=[1])
print(random_select_n_points(single_point))
    # Data(pos=[10, 3], y=[10])

Expected behavior

The second example should leave y=[1] and should not duplicate the values in y to the amount of fixed points.

print(random_select_n_points(single_point))

>>> Data(pos=[10, 3], y=[1])

Why does this happen and possible solution:

In fixed_points.py @ master, the following loop checks if the item is a tensor and if the item.size() is equal to the number of nodes.

After the number of nodes is in the given scenario equal to one, both conditions hold and y is subset by choice

for key, item in data:
    if bool(re.search('edge', key)):
        continue

    if torch.is_tensor(item) and item.size(0) == num_nodes:
        data[key] = item[choice]

An easy solution could be:

for key, item in data:
    if bool(re.search('edge', key)):
        continue

    if torch.is_tensor(item) and item.size(0) == num_nodes and item.size(0) != 1:
        data[key] = item[choice]

Which solves the described problem, but I do not know if there are any side-effects 😦

Environment

• OS: Win10
• Python version: 3.7.0
• PyTorch version: 1.4.0

[rusty1s]

You can install the latest PyG to test that as follows:

pip install git+https://github.com/rusty1s/pytorch_geometric.git

[rusty1s]

Ok, I see. I do not think adding item.size(0) != 1 does any harm, so I added that functionality into master.