train_tensors in BaseDataset is mis-interpreted

[nabenabe0928]

The continuation of issue#352.

if I understand the code correctly, we are assuming that train_tensors[0] is always feature tensors and train_tensors[1] is always labels.
However, Dataset in torch assumes the following structure:

In [1]: import torchvision.datasets

In [2]: train_tensors = torchvision.datasets.CIFAR10('cifar10/')

In [3]: train_tensors[0]
Out[3]: (<PIL.Image.Image image mode=RGB size=32x32 at 0x7F8B6DAD3370>, 6)

In [4]: train_tensors[1]
Out[4]: (<PIL.Image.Image image mode=RGB size=32x32 at 0x7F8B6DAD33A0>, 9)

It indicates that the __getitem__ of Dataset (actually it applies to other Dataset child classes as well) returns:

train_tensors[i] := the i-th instance in the given dataset

It does not matter unless we use torch stuff, but those stuffs will cause many issues when we try to merge image tasks

[ravinkohli]

Hey, actually we do follow the structure of a Dataset defined in torch see here. Specifically for image dataset, for FilePathDataset also it follows the structure here. I don't think we will have issues when we try to merge image tasks. Moreover, if we were to have issues, we would have had them also for tabular tasks as we use a torch DataLoader and it works fine. However, we may have issues with other parts of the ImageDataset code but we can look at it when we start working with images.