keep more info in DatasetInfo.from_merge #6585

[JochenSiegWork]

• try not to merge DatasetInfos if they're equal

• fixes losing DatasetInfo during parallel Dataset.map

[thiagobarbosa]

@JochenSiegWork fyi, that seems to also affect the trainer.push_to_hub() method, which I guess also needs to parse that DatasetInfo from the kwargs used by push_to_hub.
There is short discussion about it here.
Would be great if you can check if your PR would also fix that!

[JochenSiegWork]

@JochenSiegWork fyi, that seems to also affect the trainer.push_to_hub() method, which I guess also needs to parse that DatasetInfo from the kwargs used by push_to_hub. There is short discussion about it here. Would be great if you can check if your PR would also fix that!

Hi @thiagobarbosa, it might be related but I didn't worked with push_to_hub yet. I don't see a minimal example reproducing the specific error in your link. However, if you have a running version producing the error locally you can test it by pulling this PR and run your specific example locally.

[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.

[lhoestq]

Thanks for the fix !

[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)	0.004729 / 0.011353 (-0.006624)	0.002983 / 0.011008 (-0.008025)	0.062482 / 0.038508 (0.023974)	0.028406 / 0.023109 (0.005297)	0.255896 / 0.275898 (-0.020002)	0.276423 / 0.323480 (-0.047057)	0.003828 / 0.007986 (-0.004157)	0.002601 / 0.004328 (-0.001728)	0.048954 / 0.004250 (0.044704)	0.040661 / 0.037052 (0.003609)	0.277710 / 0.258489 (0.019221)	0.290360 / 0.293841 (-0.003481)	0.027105 / 0.128546 (-0.101441)	0.010168 / 0.075646 (-0.065478)	0.206835 / 0.419271 (-0.212436)	0.035226 / 0.043533 (-0.008306)	0.262567 / 0.255139 (0.007428)	0.273979 / 0.283200 (-0.009221)	0.017576 / 0.141683 (-0.124106)	1.125588 / 1.452155 (-0.326566)	1.185018 / 1.492716 (-0.307698)

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)	0.092192 / 0.018006 (0.074186)	0.298350 / 0.000490 (0.297861)	0.000217 / 0.000200 (0.000017)	0.000051 / 0.000054 (-0.000003)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.017925 / 0.037411 (-0.019486)	0.060285 / 0.014526 (0.045759)	0.076579 / 0.176557 (-0.099978)	0.118830 / 0.737135 (-0.618305)	0.073017 / 0.296338 (-0.223322)

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)	0.288149 / 0.215209 (0.072940)	2.840004 / 2.077655 (0.762349)	1.495758 / 1.504120 (-0.008361)	1.362338 / 1.541195 (-0.178857)	1.389746 / 1.468490 (-0.078744)	0.576891 / 4.584777 (-4.007886)	2.375724 / 3.745712 (-1.369988)	2.707405 / 5.269862 (-2.562457)	1.719850 / 4.565676 (-2.845826)	0.067055 / 0.424275 (-0.357220)	0.005039 / 0.007607 (-0.002568)	0.346626 / 0.226044 (0.120581)	3.468346 / 2.268929 (1.199418)	1.860686 / 55.444624 (-53.583938)	1.582929 / 6.876477 (-5.293548)	1.613131 / 2.142072 (-0.528941)	0.659022 / 4.805227 (-4.146206)	0.118477 / 6.500664 (-6.382187)	0.041614 / 0.075469 (-0.033855)

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)	1.005062 / 1.841788 (-0.836726)	11.203210 / 8.074308 (3.128902)	10.320764 / 10.191392 (0.129372)	0.128541 / 0.680424 (-0.551883)	0.014646 / 0.534201 (-0.519555)	0.285280 / 0.579283 (-0.294003)	0.263613 / 0.434364 (-0.170751)	0.321161 / 0.540337 (-0.219177)	0.420565 / 1.386936 (-0.966371)

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)	0.005288 / 0.011353 (-0.006065)	0.003048 / 0.011008 (-0.007960)	0.049196 / 0.038508 (0.010688)	0.032104 / 0.023109 (0.008994)	0.279345 / 0.275898 (0.003447)	0.300194 / 0.323480 (-0.023286)	0.004045 / 0.007986 (-0.003941)	0.002594 / 0.004328 (-0.001735)	0.047680 / 0.004250 (0.043430)	0.044294 / 0.037052 (0.007241)	0.292330 / 0.258489 (0.033841)	0.318610 / 0.293841 (0.024769)	0.050417 / 0.128546 (-0.078129)	0.010326 / 0.075646 (-0.065320)	0.057372 / 0.419271 (-0.361899)	0.032985 / 0.043533 (-0.010548)	0.277717 / 0.255139 (0.022579)	0.295692 / 0.283200 (0.012493)	0.017756 / 0.141683 (-0.123927)	1.166277 / 1.452155 (-0.285877)	1.213337 / 1.492716 (-0.279380)

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)	0.091365 / 0.018006 (0.073359)	0.296261 / 0.000490 (0.295772)	0.000225 / 0.000200 (0.000025)	0.000043 / 0.000054 (-0.000011)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.021973 / 0.037411 (-0.015438)	0.074631 / 0.014526 (0.060106)	0.085645 / 0.176557 (-0.090911)	0.125181 / 0.737135 (-0.611955)	0.086893 / 0.296338 (-0.209445)

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)	0.294110 / 0.215209 (0.078901)	2.855531 / 2.077655 (0.777876)	1.583204 / 1.504120 (0.079084)	1.453911 / 1.541195 (-0.087284)	1.467031 / 1.468490 (-0.001460)	0.581214 / 4.584777 (-4.003562)	2.423626 / 3.745712 (-1.322086)	2.736665 / 5.269862 (-2.533197)	1.707000 / 4.565676 (-2.858676)	0.061171 / 0.424275 (-0.363104)	0.004789 / 0.007607 (-0.002818)	0.344546 / 0.226044 (0.118502)	3.530955 / 2.268929 (1.262027)	1.962532 / 55.444624 (-53.482092)	1.670207 / 6.876477 (-5.206270)	1.669041 / 2.142072 (-0.473031)	0.642298 / 4.805227 (-4.162929)	0.115503 / 6.500664 (-6.385161)	0.040729 / 0.075469 (-0.034740)

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)	0.973101 / 1.841788 (-0.868687)	11.823894 / 8.074308 (3.749586)	10.664592 / 10.191392 (0.473200)	0.139848 / 0.680424 (-0.540576)	0.015728 / 0.534201 (-0.518473)	0.289135 / 0.579283 (-0.290148)	0.271325 / 0.434364 (-0.163039)	0.332253 / 0.540337 (-0.208085)	0.416982 / 1.386936 (-0.969954)