How to split a dataset in v2?

Part 1 (2020)
1

I am trying to run a simple mnist classification with 50% of the training samples.

I created the datasets as this:

splitter = GrandparentSplitter(train_name='training', valid_name='testing')
mnist = DataBlock(blocks=(ImageBlock(PILImageBW), CategoryBlock),
                  get_items=get_image_files, splitter=splitter, get_y=parent_label)
data = mnist.dataloaders(untar_data(URLs.MNIST), bs=64)

I tried to split the train_ds this way:

train_ds_50 = data.train_ds[0:len(data.train_ds)//2]
data.train_ds = train_ds_50

which leads to error, as one cannot set data.train_ds.

I guess there is a way to do this with transforms, but I cannot find any example in the documentation or the forum.

2

You’ve got a bunch of splitters here that you can use to do this.

3

It seems to me that those splitters help you take files in the filesystem and organize in your Datablock. But I already did that. I do have a Datablock with the images. My problem is how to take the train_ds of a Datablock and change it. The code prevents setting a new value directly.