Fix parallel downloads for datasets without scripts

[lhoestq]

Enable parallel downloads using multiprocessing when num_proc is passed to load_dataset.

It was enabled for datasets with scripts already (if they passed lists to dl_manager.download) but not for no-script datasets (we pass dicts {split: [list of files]} to dl_manager.download for those ones).

I fixed this by parallelising on the lists contained in the data files dicts when possible.

I also added a context manager stack_multiprocessing_download_progress_bars in DownloadManager to stack the progress bard of the downloads (from cached_path(...) calls). Otherwise the progress bars overlap each other with an annoying flickering effect.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[albertvillanova]

Thanks for the useful enhancement.

[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)	0.005002 / 0.011353 (-0.006350)	0.003300 / 0.011008 (-0.007708)	0.062509 / 0.038508 (0.024001)	0.029807 / 0.023109 (0.006698)	0.249935 / 0.275898 (-0.025963)	0.264320 / 0.323480 (-0.059160)	0.003790 / 0.007986 (-0.004195)	0.002554 / 0.004328 (-0.001774)	0.048207 / 0.004250 (0.043956)	0.042033 / 0.037052 (0.004981)	0.245725 / 0.258489 (-0.012764)	0.276695 / 0.293841 (-0.017146)	0.026502 / 0.128546 (-0.102044)	0.010379 / 0.075646 (-0.065268)	0.207002 / 0.419271 (-0.212269)	0.034648 / 0.043533 (-0.008885)	0.247957 / 0.255139 (-0.007182)	0.263921 / 0.283200 (-0.019278)	0.017710 / 0.141683 (-0.123973)	1.105851 / 1.452155 (-0.346304)	1.163315 / 1.492716 (-0.329401)

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)	0.089842 / 0.018006 (0.071836)	0.352499 / 0.000490 (0.352009)	0.000201 / 0.000200 (0.000001)	0.000054 / 0.000054 (-0.000000)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.018094 / 0.037411 (-0.019317)	0.060463 / 0.014526 (0.045937)	0.073257 / 0.176557 (-0.103300)	0.119771 / 0.737135 (-0.617364)	0.075210 / 0.296338 (-0.221128)

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)	0.288365 / 0.215209 (0.073156)	2.825377 / 2.077655 (0.747722)	1.532436 / 1.504120 (0.028316)	1.393475 / 1.541195 (-0.147719)	1.381859 / 1.468490 (-0.086632)	0.564155 / 4.584777 (-4.020622)	2.398177 / 3.745712 (-1.347535)	2.730271 / 5.269862 (-2.539590)	1.713779 / 4.565676 (-2.851898)	0.062789 / 0.424275 (-0.361486)	0.004991 / 0.007607 (-0.002616)	0.340789 / 0.226044 (0.114744)	3.323543 / 2.268929 (1.054615)	1.861925 / 55.444624 (-53.582700)	1.555181 / 6.876477 (-5.321296)	1.559512 / 2.142072 (-0.582560)	0.634565 / 4.805227 (-4.170663)	0.116529 / 6.500664 (-6.384135)	0.041312 / 0.075469 (-0.034157)

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)	0.945739 / 1.841788 (-0.896049)	11.376130 / 8.074308 (3.301822)	10.007752 / 10.191392 (-0.183640)	0.126815 / 0.680424 (-0.553609)	0.013898 / 0.534201 (-0.520303)	0.287438 / 0.579283 (-0.291845)	0.261532 / 0.434364 (-0.172832)	0.320197 / 0.540337 (-0.220140)	0.414444 / 1.386936 (-0.972492)

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)	0.004994 / 0.011353 (-0.006359)	0.003407 / 0.011008 (-0.007601)	0.049281 / 0.038508 (0.010773)	0.042815 / 0.023109 (0.019706)	0.268291 / 0.275898 (-0.007607)	0.285877 / 0.323480 (-0.037603)	0.004006 / 0.007986 (-0.003980)	0.002607 / 0.004328 (-0.001721)	0.047682 / 0.004250 (0.043431)	0.044281 / 0.037052 (0.007228)	0.268287 / 0.258489 (0.009798)	0.298649 / 0.293841 (0.004808)	0.028607 / 0.128546 (-0.099939)	0.010367 / 0.075646 (-0.065279)	0.057114 / 0.419271 (-0.362158)	0.053753 / 0.043533 (0.010220)	0.269010 / 0.255139 (0.013871)	0.285057 / 0.283200 (0.001858)	0.017693 / 0.141683 (-0.123990)	1.134718 / 1.452155 (-0.317436)	1.186609 / 1.492716 (-0.306107)

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)	0.091109 / 0.018006 (0.073103)	0.298603 / 0.000490 (0.298113)	0.000216 / 0.000200 (0.000016)	0.000050 / 0.000054 (-0.000004)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.022125 / 0.037411 (-0.015286)	0.076570 / 0.014526 (0.062044)	0.088903 / 0.176557 (-0.087654)	0.126427 / 0.737135 (-0.610708)	0.091001 / 0.296338 (-0.205338)

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)	0.300332 / 0.215209 (0.085123)	2.971106 / 2.077655 (0.893452)	1.617886 / 1.504120 (0.113766)	1.476679 / 1.541195 (-0.064516)	1.483750 / 1.468490 (0.015260)	0.582569 / 4.584777 (-4.002208)	2.441804 / 3.745712 (-1.303908)	2.753927 / 5.269862 (-2.515935)	1.733546 / 4.565676 (-2.832130)	0.062653 / 0.424275 (-0.361622)	0.005019 / 0.007607 (-0.002588)	0.355556 / 0.226044 (0.129512)	3.497431 / 2.268929 (1.228503)	1.951711 / 55.444624 (-53.492913)	1.663874 / 6.876477 (-5.212602)	1.657363 / 2.142072 (-0.484709)	0.653488 / 4.805227 (-4.151739)	0.117055 / 6.500664 (-6.383609)	0.040687 / 0.075469 (-0.034782)

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)	0.969485 / 1.841788 (-0.872303)	12.064793 / 8.074308 (3.990485)	10.851531 / 10.191392 (0.660139)	0.129060 / 0.680424 (-0.551364)	0.015339 / 0.534201 (-0.518862)	0.287215 / 0.579283 (-0.292069)	0.276545 / 0.434364 (-0.157819)	0.322748 / 0.540337 (-0.217589)	0.421363 / 1.386936 (-0.965573)

[kopyl]

@lhoestq

it's still not fixed =(

[kopyl]

@lhoestq i was thinking uninstalling datasets and then pip install git+https://github.com/huggingface/datasets.git has to fix it. Buuuuut. I'm not sure what's going on actually...

Now instead of showing progress bars one after another it seems to be downloading the dataset way way way faster (like 4 mins instead of 58, thank you very much) but does not show any progress bars related to downloading at all.