Add batching to IterableDataset

[lappemic]

I've taken a try at implementing a batched IterableDataset as requested in issue #6279. This PR adds a new BatchedExamplesIterable class and a .batch() method to the IterableDataset class.

The main changes are:

1. A new BatchedExamplesIterable that groups examples into batches.
2. A .batch() method for IterableDataset to easily create batched versions.
3. Support for shuffling and sharding to work with PyTorch DataLoader and multiple workers.

I'm not sure if this is exactly what you had in mind and also have not fully tested it atm, so I'd really appreciate your feedback. Does this seem like it's heading in the right direction? I'm happy to make any changes or explore different approaches if needed.

Pinging @lhoestq

[lhoestq]

Cool ! Thanks for diving into it :)

Your implementation is great and indeed supports shuffling and batching, you just need to additionally account for state_dict (for dataset checkpointing+resuming)

That being said, I believe the implementation can be made simpler by relying on IterableDataset.map() which already implements all this. Maybe something like

def batch(self, batch_size: int, drop_last_batch: bool = False) -> "IterableDataset":
    def batch(unbatched: dict[str, list]) -> dict[str, list]:
        return {k: [v] for k, v in unbatched}

    return self.map(batch, batched=True, batch_size=batch_size, drop_last_batch=drop_last_batch)

And this way no need to reimplement everything !

(my only small concern is that it's not an Arrow-optimized function so it requires the examples to be manipulated as python objects even if the original data is in Arrow format (e.g. when streaming Parquet files) but it's not a big deal and we can see later if we need to optimize this)

[lappemic]

Thanks a lot for the feedback @lhoestq! I definitely could have saved some time looking into it properly first. 😅

Implemented the .batch() method, added a proper docsrtring for documentation, and added tests.

Let me know what you think and if this needs some update.

[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]

Nice thanks ! I added a few suggestions before we merge.

[lappemic]

Thanks for the feedbak @lhoestq!

Applied it and referenced the batched=True option in the map function and highlighted the difference. Hope i got this right.

[lhoestq]

lgtm !

[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)	0.005181 / 0.011353 (-0.006172)	0.003714 / 0.011008 (-0.007294)	0.063060 / 0.038508 (0.024552)	0.030885 / 0.023109 (0.007776)	0.239060 / 0.275898 (-0.036838)	0.262480 / 0.323480 (-0.061000)	0.004103 / 0.007986 (-0.003883)	0.002696 / 0.004328 (-0.001632)	0.048706 / 0.004250 (0.044456)	0.042577 / 0.037052 (0.005525)	0.249928 / 0.258489 (-0.008561)	0.283252 / 0.293841 (-0.010589)	0.029304 / 0.128546 (-0.099242)	0.012001 / 0.075646 (-0.063646)	0.204467 / 0.419271 (-0.214804)	0.035639 / 0.043533 (-0.007894)	0.243850 / 0.255139 (-0.011289)	0.261609 / 0.283200 (-0.021590)	0.018302 / 0.141683 (-0.123381)	1.096040 / 1.452155 (-0.356115)	1.135917 / 1.492716 (-0.356800)

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)	0.091976 / 0.018006 (0.073970)	0.296396 / 0.000490 (0.295906)	0.000203 / 0.000200 (0.000003)	0.000043 / 0.000054 (-0.000011)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.018405 / 0.037411 (-0.019007)	0.062470 / 0.014526 (0.047944)	0.073340 / 0.176557 (-0.103216)	0.119474 / 0.737135 (-0.617661)	0.075750 / 0.296338 (-0.220588)

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)	0.279586 / 0.215209 (0.064377)	2.768542 / 2.077655 (0.690887)	1.449158 / 1.504120 (-0.054962)	1.328760 / 1.541195 (-0.212435)	1.336338 / 1.468490 (-0.132152)	0.732582 / 4.584777 (-3.852195)	2.325558 / 3.745712 (-1.420154)	2.898077 / 5.269862 (-2.371784)	1.893107 / 4.565676 (-2.672569)	0.078788 / 0.424275 (-0.345487)	0.005273 / 0.007607 (-0.002335)	0.334887 / 0.226044 (0.108842)	3.304173 / 2.268929 (1.035244)	1.834743 / 55.444624 (-53.609882)	1.527463 / 6.876477 (-5.349014)	1.538824 / 2.142072 (-0.603249)	0.785646 / 4.805227 (-4.019581)	0.134876 / 6.500664 (-6.365788)	0.042894 / 0.075469 (-0.032575)

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)	0.976635 / 1.841788 (-0.865152)	11.217156 / 8.074308 (3.142848)	9.616971 / 10.191392 (-0.574421)	0.127276 / 0.680424 (-0.553148)	0.014344 / 0.534201 (-0.519857)	0.301896 / 0.579283 (-0.277387)	0.259615 / 0.434364 (-0.174749)	0.340693 / 0.540337 (-0.199645)	0.429145 / 1.386936 (-0.957791)

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)	0.005534 / 0.011353 (-0.005819)	0.003795 / 0.011008 (-0.007213)	0.049761 / 0.038508 (0.011253)	0.031311 / 0.023109 (0.008202)	0.276032 / 0.275898 (0.000134)	0.297316 / 0.323480 (-0.026164)	0.004396 / 0.007986 (-0.003590)	0.002693 / 0.004328 (-0.001635)	0.049025 / 0.004250 (0.044775)	0.039707 / 0.037052 (0.002654)	0.284264 / 0.258489 (0.025775)	0.319962 / 0.293841 (0.026121)	0.031842 / 0.128546 (-0.096705)	0.012192 / 0.075646 (-0.063454)	0.059895 / 0.419271 (-0.359376)	0.033676 / 0.043533 (-0.009856)	0.275917 / 0.255139 (0.020778)	0.292637 / 0.283200 (0.009437)	0.017992 / 0.141683 (-0.123691)	1.199329 / 1.452155 (-0.252826)	1.259083 / 1.492716 (-0.233633)

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)	0.092770 / 0.018006 (0.074764)	0.313363 / 0.000490 (0.312873)	0.000212 / 0.000200 (0.000013)	0.000052 / 0.000054 (-0.000003)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.022977 / 0.037411 (-0.014434)	0.076839 / 0.014526 (0.062314)	0.088289 / 0.176557 (-0.088267)	0.128625 / 0.737135 (-0.608510)	0.089348 / 0.296338 (-0.206990)

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)	0.300881 / 0.215209 (0.085672)	2.946499 / 2.077655 (0.868845)	1.599686 / 1.504120 (0.095566)	1.479332 / 1.541195 (-0.061862)	1.476910 / 1.468490 (0.008420)	0.720536 / 4.584777 (-3.864241)	0.944822 / 3.745712 (-2.800890)	2.771864 / 5.269862 (-2.497998)	1.886573 / 4.565676 (-2.679103)	0.078462 / 0.424275 (-0.345813)	0.005392 / 0.007607 (-0.002215)	0.354984 / 0.226044 (0.128939)	3.516449 / 2.268929 (1.247520)	1.977033 / 55.444624 (-53.467592)	1.671922 / 6.876477 (-5.204555)	1.785755 / 2.142072 (-0.356318)	0.795330 / 4.805227 (-4.009897)	0.132895 / 6.500664 (-6.367769)	0.041178 / 0.075469 (-0.034291)

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)	1.031780 / 1.841788 (-0.810008)	11.855600 / 8.074308 (3.781292)	10.245599 / 10.191392 (0.054207)	0.140649 / 0.680424 (-0.539775)	0.015332 / 0.534201 (-0.518869)	0.299402 / 0.579283 (-0.279881)	0.120007 / 0.434364 (-0.314357)	0.337770 / 0.540337 (-0.202568)	0.433679 / 1.386936 (-0.953257)