Faster column validation and reordering

[psmyth94]

I work with bioinformatics data and often these tables have thousands and even tens of thousands of features. These tables are also accompanied by metadata that I do not want to pass in the model. When I perform set_format('pt', columns=large_column_list) , it can take several minutes before it finishes. The culprit is when the following check is performed: any(col not in self._data.column_names for col in columns). Replacing this by set(columns) - (self._data.column_names) is more efficient.

[mariosasko]

Thanks for working on this!

Besides Dataset.set_format, we also use this inefficient check in the following methods:

• Dataset.select_columns
• Dataset.remove_columns
• Dataset.rename_columns
• Dataset.map (input_columns and remove_columns checks)
• IterableDataset.select_columns

So, let's also update them in this PR to be consistent across the codebase.

[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.

[psmyth94]

Thanks @mariosasko, I made the changes. However, I did some tests with map and I still saw that it took ~3.5 minutes per batch on 6000 features when using dataset.map(lambda x: x, batched=True). From the profile, the culprits were mainly with ArrowWriter.write_batch and ArrowWriter._build_writer. The slow down from _build_writer is due to updating existing features with the inferred ones. I don't think this can be optimized any further, but fortunately, I can avoid this by setting the features in map. On the other hand, write_batch selects cols based on intersection and difference between schema names and example keys using two for loops. The same exists in ArrowWriter.write_examples_on_file. Optimizing the column selection using set operations effectively brings it from 3.5 minutes per batch down to 6 seconds per batch. Can we add these changes along with this PR?

Edit: Ah just realized you can avoid the issue with inferring features altogether when you set the format to arrow (or pandas).

[mariosasko]

Indeed, let's also improve this column reordering logic to avoid quadratic time complexity.

Two comments to fix the CI failure.

[mariosasko]

I took the liberty to apply the code suggestions so that we can include this PR in the next release.

[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)	0.004990 / 0.011353 (-0.006363)	0.003138 / 0.011008 (-0.007870)	0.062368 / 0.038508 (0.023860)	0.028634 / 0.023109 (0.005524)	0.241297 / 0.275898 (-0.034601)	0.264433 / 0.323480 (-0.059047)	0.003133 / 0.007986 (-0.004852)	0.003444 / 0.004328 (-0.000885)	0.048522 / 0.004250 (0.044271)	0.043700 / 0.037052 (0.006648)	0.257054 / 0.258489 (-0.001435)	0.277551 / 0.293841 (-0.016290)	0.027132 / 0.128546 (-0.101414)	0.010395 / 0.075646 (-0.065251)	0.208003 / 0.419271 (-0.211269)	0.035814 / 0.043533 (-0.007719)	0.250098 / 0.255139 (-0.005041)	0.266726 / 0.283200 (-0.016474)	0.018424 / 0.141683 (-0.123259)	1.129242 / 1.452155 (-0.322912)	1.167674 / 1.492716 (-0.325042)

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)	0.091808 / 0.018006 (0.073802)	0.298726 / 0.000490 (0.298236)	0.000219 / 0.000200 (0.000019)	0.000051 / 0.000054 (-0.000004)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.019119 / 0.037411 (-0.018292)	0.061969 / 0.014526 (0.047443)	0.073392 / 0.176557 (-0.103165)	0.119460 / 0.737135 (-0.617675)	0.074072 / 0.296338 (-0.222266)

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)	0.281435 / 0.215209 (0.066226)	2.702094 / 2.077655 (0.624439)	1.411541 / 1.504120 (-0.092579)	1.284084 / 1.541195 (-0.257111)	1.302638 / 1.468490 (-0.165852)	0.562420 / 4.584777 (-4.022357)	2.364890 / 3.745712 (-1.380822)	2.744033 / 5.269862 (-2.525828)	1.699000 / 4.565676 (-2.866677)	0.062315 / 0.424275 (-0.361961)	0.004982 / 0.007607 (-0.002625)	0.334385 / 0.226044 (0.108341)	3.203268 / 2.268929 (0.934339)	1.766998 / 55.444624 (-53.677627)	1.497164 / 6.876477 (-5.379313)	1.509996 / 2.142072 (-0.632077)	0.633014 / 4.805227 (-4.172213)	0.115317 / 6.500664 (-6.385347)	0.041120 / 0.075469 (-0.034349)

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)	0.965877 / 1.841788 (-0.875911)	11.219909 / 8.074308 (3.145601)	9.333822 / 10.191392 (-0.857570)	0.136482 / 0.680424 (-0.543941)	0.013632 / 0.534201 (-0.520569)	0.287251 / 0.579283 (-0.292032)	0.262786 / 0.434364 (-0.171578)	0.322893 / 0.540337 (-0.217444)	0.418180 / 1.386936 (-0.968756)

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)	0.005444 / 0.011353 (-0.005909)	0.003147 / 0.011008 (-0.007862)	0.049242 / 0.038508 (0.010734)	0.030944 / 0.023109 (0.007834)	0.281901 / 0.275898 (0.006003)	0.303820 / 0.323480 (-0.019660)	0.004326 / 0.007986 (-0.003659)	0.002696 / 0.004328 (-0.001632)	0.048306 / 0.004250 (0.044055)	0.044145 / 0.037052 (0.007093)	0.297253 / 0.258489 (0.038764)	0.324062 / 0.293841 (0.030221)	0.046724 / 0.128546 (-0.081823)	0.010079 / 0.075646 (-0.065567)	0.057635 / 0.419271 (-0.361636)	0.033621 / 0.043533 (-0.009912)	0.282303 / 0.255139 (0.027164)	0.300761 / 0.283200 (0.017561)	0.017116 / 0.141683 (-0.124567)	1.156519 / 1.452155 (-0.295636)	1.216087 / 1.492716 (-0.276630)

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)	0.093011 / 0.018006 (0.075005)	0.301310 / 0.000490 (0.300820)	0.000223 / 0.000200 (0.000023)	0.000053 / 0.000054 (-0.000001)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.023112 / 0.037411 (-0.014299)	0.075192 / 0.014526 (0.060666)	0.086213 / 0.176557 (-0.090343)	0.125853 / 0.737135 (-0.611282)	0.087754 / 0.296338 (-0.208585)

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)	0.301095 / 0.215209 (0.085886)	2.911769 / 2.077655 (0.834114)	1.614708 / 1.504120 (0.110588)	1.494497 / 1.541195 (-0.046698)	1.506978 / 1.468490 (0.038488)	0.572743 / 4.584777 (-4.012034)	2.417142 / 3.745712 (-1.328570)	2.755338 / 5.269862 (-2.514523)	1.711026 / 4.565676 (-2.854650)	0.062732 / 0.424275 (-0.361543)	0.005031 / 0.007607 (-0.002576)	0.352343 / 0.226044 (0.126298)	3.465183 / 2.268929 (1.196255)	1.958795 / 55.444624 (-53.485829)	1.682239 / 6.876477 (-5.194238)	1.688897 / 2.142072 (-0.453176)	0.643311 / 4.805227 (-4.161916)	0.115426 / 6.500664 (-6.385238)	0.040338 / 0.075469 (-0.035131)

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)	1.005322 / 1.841788 (-0.836466)	11.779380 / 8.074308 (3.705072)	10.041574 / 10.191392 (-0.149818)	0.127617 / 0.680424 (-0.552807)	0.015840 / 0.534201 (-0.518361)	0.286905 / 0.579283 (-0.292378)	0.275180 / 0.434364 (-0.159183)	0.332498 / 0.540337 (-0.207840)	0.410719 / 1.386936 (-0.976217)