Dataset splits do not have exactly the requested weights

[ageron]

Short description
When I split the tf_flowers dataset into subsplits with weights 10, 15 and 75, I actually get datasets of size 400, 600, and 2670. This translates to 10.9%, 16.3%, 72.8%, which is pretty different from what I requested.
Moreover, apart from iterating through the whole datasets, there does not seem to be a way to know the size of the splits.

Environment information

• Operating System: MacOSX 10.13.6
• Python version: 3.6.8
• tfds-nightly version: tfds-nightly-1.0.1.dev201903180105
• tf-nightly-2.0-preview version: tf-nightly-2.0-preview-2.0.0.dev20190319

Reproduction instructions

import tensorflow_datasets as tfds

test_split, valid_split, train_split = tfds.Split.TRAIN.subsplit([10, 15, 75])

test_set = tfds.load("tf_flowers", split=test_split, as_supervised=True)
valid_set = tfds.load("tf_flowers", split=valid_split, as_supervised=True)
train_set = tfds.load("tf_flowers", split=train_split, as_supervised=True)

def dataset_length(dataset):
    count = 0
    for image in dataset:
        count += 1
    return count

print(dataset_length(test_set)) # 400
print(dataset_length(valid_set)) # 600
print(dataset_length(train_set)) # 2670

Expected behavior
I expected split sizes with the requested ratios (rounded up or down to the nearest integer): in this example, the correct sizes should have been 367, 550 and 2753 (or 551 and 2752).
I also expect to be able to know the subsplit sizes without iterating through the datasets.

Additional context
TFDS is cool.

[Conchylicultor]

Yes, this is due to the current implementation. Because each shard do not contains exactly x*100 examples, the remaining examples of each shard are arbitrary assign to one of the split, as this error accumulate among shards, it may lead to an incorrect number of sample per subsplit.

We're going to try to improve the subsplit implementation to reduce the issue. Note that because the total number of sample 2752 % 100 == 52 != 0, the last 52 example of the dataset will still be arbitrary assigned.

[Conchylicultor]

With the new split reading, independent of the number of shards, this should be fixed.
#458

For the tf_flowers, because the number of examples % 100 isn't null, the last examples will create some small unbalance.

For your use case, this should gives:

test_split, valid_split, train_split = tfds.Split.TRAIN.subsplit([10, 15, 75])

total_num_examples = 3670 = 3600 + 70
test = 3600 * 0.1 + 10 = 370
valid =  3600 * 0.15 + 15 = 555
train =  3600 * 0.75 + 45 = 2745

Thanks again for reporting this. I'm closing this but do not hesitate to re-open if this doesn't solve your issue.