Fix multiprocessing with spawn in iterable datasets

[bruno-hays]

The "Spawn" method is preferred when multiprocessing on macOS or Windows systems, instead of the "Fork" method on linux systems.

This causes some methods of Iterable Datasets to break when using a dataloader with more than 0 workers.

I fixed the issue by replacing lambda and local methods which are not pickle-able.

See the example below:

from datasets import load_dataset
from torch.utils.data import DataLoader


if __name__ == "__main__":
    dataset = load_dataset("lhoestq/demo1", split="train")
    dataset = dataset.to_iterable_dataset(num_shards=3)

    dataset = dataset.remove_columns(["package_name"])
    dataset = dataset.rename_columns({
        "review": "review1"
    })
    dataset = dataset.rename_column("date", "date1")
    for sample in DataLoader(dataset, batch_size=None, num_workers=3):
        print(sample)

To notice the fix on a linux system, adding these lines should do the trick:

import multiprocessing
multiprocessing.set_start_method('spawn')

I also removed what looks like code duplication between rename_colums and rename_column

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[bruno-hays]

@lhoestq
A test is failing, but I don't think it is due to my changes

[lhoestq]

Good catch ! Could you add a test to make sure transformed IterableDataset objects are still picklable ?

Something like test_pickle_after_many_transforms in in test_iterable_dataset.py that does a bunch or rename, map, take on a dataset and checks that the dataset can be pickled at the end and the reloaded dataset returns the same elements

[bruno-hays]

@lhoestq
I added the test and fixed one last method

[lhoestq]

Awesome thanks !

[github-actions]

Show benchmarks

PyArrow==8.0.0

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

metric	read_batch_formatted_as_numpy after write_array2d	read_batch_formatted_as_numpy after write_flattened_sequence	read_batch_formatted_as_numpy after write_nested_sequence	read_batch_unformated after write_array2d	read_batch_unformated after write_flattened_sequence	read_batch_unformated after write_nested_sequence	read_col_formatted_as_numpy after write_array2d	read_col_formatted_as_numpy after write_flattened_sequence	read_col_formatted_as_numpy after write_nested_sequence	read_col_unformated after write_array2d	read_col_unformated after write_flattened_sequence	read_col_unformated after write_nested_sequence	read_formatted_as_numpy after write_array2d	read_formatted_as_numpy after write_flattened_sequence	read_formatted_as_numpy after write_nested_sequence	read_unformated after write_array2d	read_unformated after write_flattened_sequence	read_unformated after write_nested_sequence	write_array2d	write_flattened_sequence	write_nested_sequence
new / old (diff)	0.006537 / 0.011353 (-0.004816)	0.003960 / 0.011008 (-0.007048)	0.085135 / 0.038508 (0.046627)	0.079271 / 0.023109 (0.056162)	0.383743 / 0.275898 (0.107845)	0.414622 / 0.323480 (0.091143)	0.004202 / 0.007986 (-0.003784)	0.003537 / 0.004328 (-0.000791)	0.065758 / 0.004250 (0.061508)	0.054225 / 0.037052 (0.017173)	0.395715 / 0.258489 (0.137226)	0.438985 / 0.293841 (0.145144)	0.030590 / 0.128546 (-0.097956)	0.008754 / 0.075646 (-0.066892)	0.288415 / 0.419271 (-0.130857)	0.051863 / 0.043533 (0.008330)	0.382501 / 0.255139 (0.127363)	0.414428 / 0.283200 (0.131228)	0.024084 / 0.141683 (-0.117599)	1.478726 / 1.452155 (0.026572)	1.544763 / 1.492716 (0.052047)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.285143 / 0.018006 (0.267136)	0.603859 / 0.000490 (0.603369)	0.004330 / 0.000200 (0.004131)	0.000108 / 0.000054 (0.000054)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.027856 / 0.037411 (-0.009555)	0.081963 / 0.014526 (0.067437)	0.104106 / 0.176557 (-0.072451)	0.151378 / 0.737135 (-0.585757)	0.096476 / 0.296338 (-0.199862)

Benchmark: benchmark_iterating.json

metric	read 5000	read 50000	read_batch 50000 10	read_batch 50000 100	read_batch 50000 1000	read_formatted numpy 5000	read_formatted pandas 5000	read_formatted tensorflow 5000	read_formatted torch 5000	read_formatted_batch numpy 5000 10	read_formatted_batch numpy 5000 1000	shuffled read 5000	shuffled read 50000	shuffled read_batch 50000 10	shuffled read_batch 50000 100	shuffled read_batch 50000 1000	shuffled read_formatted numpy 5000	shuffled read_formatted_batch numpy 5000 10	shuffled read_formatted_batch numpy 5000 1000
new / old (diff)	0.402938 / 0.215209 (0.187729)	4.042312 / 2.077655 (1.964657)	2.068421 / 1.504120 (0.564301)	1.877870 / 1.541195 (0.336675)	1.947643 / 1.468490 (0.479153)	0.482031 / 4.584777 (-4.102746)	3.554747 / 3.745712 (-0.190965)	3.307811 / 5.269862 (-1.962050)	2.082886 / 4.565676 (-2.482791)	0.056853 / 0.424275 (-0.367422)	0.007535 / 0.007607 (-0.000072)	0.483694 / 0.226044 (0.257649)	4.827906 / 2.268929 (2.558978)	2.567572 / 55.444624 (-52.877052)	2.167206 / 6.876477 (-4.709271)	2.414442 / 2.142072 (0.272369)	0.579472 / 4.805227 (-4.225755)	0.132976 / 6.500664 (-6.367688)	0.059315 / 0.075469 (-0.016154)

Benchmark: benchmark_map_filter.json

metric	filter	map fast-tokenizer batched	map identity	map identity batched	map no-op batched	map no-op batched numpy	map no-op batched pandas	map no-op batched pytorch	map no-op batched tensorflow
new / old (diff)	1.260086 / 1.841788 (-0.581702)	19.438297 / 8.074308 (11.363989)	14.188161 / 10.191392 (3.996769)	0.168534 / 0.680424 (-0.511890)	0.018070 / 0.534201 (-0.516131)	0.394241 / 0.579283 (-0.185043)	0.411057 / 0.434364 (-0.023307)	0.461123 / 0.540337 (-0.079215)	0.626844 / 1.386936 (-0.760092)

PyArrow==latest

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

metric	read_batch_formatted_as_numpy after write_array2d	read_batch_formatted_as_numpy after write_flattened_sequence	read_batch_formatted_as_numpy after write_nested_sequence	read_batch_unformated after write_array2d	read_batch_unformated after write_flattened_sequence	read_batch_unformated after write_nested_sequence	read_col_formatted_as_numpy after write_array2d	read_col_formatted_as_numpy after write_flattened_sequence	read_col_formatted_as_numpy after write_nested_sequence	read_col_unformated after write_array2d	read_col_unformated after write_flattened_sequence	read_col_unformated after write_nested_sequence	read_formatted_as_numpy after write_array2d	read_formatted_as_numpy after write_flattened_sequence	read_formatted_as_numpy after write_nested_sequence	read_unformated after write_array2d	read_unformated after write_flattened_sequence	read_unformated after write_nested_sequence	write_array2d	write_flattened_sequence	write_nested_sequence
new / old (diff)	0.006896 / 0.011353 (-0.004457)	0.004207 / 0.011008 (-0.006801)	0.064981 / 0.038508 (0.026473)	0.080261 / 0.023109 (0.057152)	0.399403 / 0.275898 (0.123505)	0.433099 / 0.323480 (0.109619)	0.005697 / 0.007986 (-0.002288)	0.003601 / 0.004328 (-0.000728)	0.065924 / 0.004250 (0.061673)	0.058868 / 0.037052 (0.021815)	0.403705 / 0.258489 (0.145216)	0.439218 / 0.293841 (0.145377)	0.032789 / 0.128546 (-0.095757)	0.008675 / 0.075646 (-0.066971)	0.071217 / 0.419271 (-0.348055)	0.048487 / 0.043533 (0.004954)	0.399878 / 0.255139 (0.144739)	0.412816 / 0.283200 (0.129616)	0.023905 / 0.141683 (-0.117778)	1.541402 / 1.452155 (0.089247)	1.588080 / 1.492716 (0.095364)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.322863 / 0.018006 (0.304856)	0.530291 / 0.000490 (0.529802)	0.004862 / 0.000200 (0.004662)	0.000097 / 0.000054 (0.000042)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.032697 / 0.037411 (-0.004715)	0.092416 / 0.014526 (0.077891)	0.107355 / 0.176557 (-0.069201)	0.160217 / 0.737135 (-0.576918)	0.109286 / 0.296338 (-0.187052)

Benchmark: benchmark_iterating.json

metric	read 5000	read 50000	read_batch 50000 10	read_batch 50000 100	read_batch 50000 1000	read_formatted numpy 5000	read_formatted pandas 5000	read_formatted tensorflow 5000	read_formatted torch 5000	read_formatted_batch numpy 5000 10	read_formatted_batch numpy 5000 1000	shuffled read 5000	shuffled read 50000	shuffled read_batch 50000 10	shuffled read_batch 50000 100	shuffled read_batch 50000 1000	shuffled read_formatted numpy 5000	shuffled read_formatted_batch numpy 5000 10	shuffled read_formatted_batch numpy 5000 1000
new / old (diff)	0.437375 / 0.215209 (0.222166)	4.362644 / 2.077655 (2.284990)	2.335404 / 1.504120 (0.831284)	2.173215 / 1.541195 (0.632020)	2.254061 / 1.468490 (0.785571)	0.493906 / 4.584777 (-4.090871)	3.609025 / 3.745712 (-0.136687)	3.352380 / 5.269862 (-1.917481)	2.074185 / 4.565676 (-2.491492)	0.057863 / 0.424275 (-0.366412)	0.007297 / 0.007607 (-0.000310)	0.512464 / 0.226044 (0.286420)	5.135921 / 2.268929 (2.866993)	2.788889 / 55.444624 (-52.655736)	2.479097 / 6.876477 (-4.397379)	2.717848 / 2.142072 (0.575776)	0.590442 / 4.805227 (-4.214785)	0.133721 / 6.500664 (-6.366943)	0.061491 / 0.075469 (-0.013978)

Benchmark: benchmark_map_filter.json

metric	filter	map fast-tokenizer batched	map identity	map identity batched	map no-op batched	map no-op batched numpy	map no-op batched pandas	map no-op batched pytorch	map no-op batched tensorflow
new / old (diff)	1.429564 / 1.841788 (-0.412224)	20.628733 / 8.074308 (12.554425)	15.299571 / 10.191392 (5.108179)	0.171032 / 0.680424 (-0.509392)	0.019995 / 0.534201 (-0.514206)	0.401283 / 0.579283 (-0.178000)	0.416504 / 0.434364 (-0.017860)	0.471219 / 0.540337 (-0.069118)	0.641299 / 1.386936 (-0.745637)