Splitting a dataset from tfds results in inaccurate partition sizes

[vmirly]

I am trying to split the iris dataset into train/test with 2/3 for training and 1/3 for testing. So, I used the percent as follows:

import tensorflow_datasets as tfds

first_67_percent = tfds.Split.TRAIN.subsplit(tfds.percent[:67])
last_33_percent = tfds.Split.TRAIN.subsplit(tfds.percent[-33:])

ds_train_orig = tfds.load('iris', split=first_67_percent)
ds_test = tfds.load('iris', split=last_33_percent)

but this results in 117 samples for training and 33 for test which is not correct. It should be 100 for training and 50 for test. I also tried the subsplit as follows but got similar results:

split_train, split_test = tfds.Split.TRAIN.subsplit([2, 1])

ds_train_orig = tfds.load('iris', split=split_train)
ds_test = tfds.load('iris', split=split_test)

this time, it gives me 116 train and 34 test samples. I have also posted this question on Stackoverflow: https://stackoverflow.com/questions/56553357/splitting-a-tensorflow-dataset-using-tfds-percent

[Conchylicultor]

Thanks for reporting. This is a know issue due to the implementation which round the percent if the dataset number of example % 100 !=0.
We are currently in the process of updating this with a better implementation.

[vmirly]

Ok, cool. That sounds great! 👍

For now, I figured I can load the entire data and then use ds_train = ds.take(100) and ds_test = ds.skip(100) to split it into train/test.

[Conchylicultor]

Be careful with this as the data is not guarantee to be generated in the same order for the train set unless as_dataset(shuffle_files=False). So you may end up with overlapping train/test sets

[vmirly]

Yes, thanks.
I reshuffle twice, once in the original version before splitting to train/test, and then in the train set after split. In the first shuffle, I reshuffle_each_iteration=False:

ds = tf.data.Dataset.range(15)
ds = ds.shuffle(15, reshuffle_each_iteration=False)


ds_train = ds.take(10)
ds_test = ds.skip(10)

ds_train = ds_train.shuffle(10).repeat(10)
ds_test = ds_test.shuffle(5)
ds_test = ds_test.repeat(10)

set_train = set()
for i,item in enumerate(ds_train):
    set_train.add(item.numpy())

set_test = set()
for i,item in enumerate(ds_test):
    set_test.add(item.numpy())

print(set_train, set_test)

If I don't do that, repeating the train and test dataset will result in overlapped samples.

[kaushikacharya]

this time, it gives me 116 train and 34 test samples. I have also posted this question on Stackoverflow: https://stackoverflow.com/questions/56553357/splitting-a-tensorflow-dataset-using-tfds-percent

Related issue: #292

[vmirly]

Thanks , I closed the issue then.