Discussion: Train/val/test splitting and dataset generation

[c-salomonsen]

As mentioned in #12, we want to split the data in the different partitions, and need to decide how to.

[Seilmast]

I noticed that the SVHN dataset has a train/val split built in, which can be retrived by passing a single argument. Does the MNIST have the same?
If so, we can just make two instances of the loader class instead of doing the split mentioned in #12, point 1

[hzavadil98]

@Seilmast MNIST also comes split into training and testing datasets, but the random_split is necessary when we want to have train/val/test setup. (Splitting training part of the dataset into train and validation #53 )

[Seilmast]

@hzavadil98 Normally yes, but I don't see any reason to use a train/vali/test split for such simple tasks. We're not being tested on scientific rigour here, but rather out ability to implement the project we've been given.