Add support for categorical/dictionary types

[EthanSteinberg]

Arrow has a very useful dictionary/categorical type (https://arrow.apache.org/docs/python/generated/pyarrow.dictionary.html). This data type has significant speed, memory and disk benefits over pa.string() when there are only a few unique text strings in a column.

Unfortunately, huggingface datasets currently does not support this type. So huggingface datasets cannot natively read many parquet files that use this datatype .This PR adds support for Huggingface Datasets to read categorical/dictionary data.

Note: This PR functions by simply converting those dictionary/categorical types to strings. This means that huggingface datasets cannot take advantage of the compute benefits of categoricals, but it significantly simplifies logic. At this time, I do not think it makes sense to optimize categorical support within huggingface datasets and that we should only try to optimize later, if necessary.

Closes #5706

[lhoestq]

Awesome ! I'll add a change to this PR to fix this if you don't mind:

import pyarrow as pa
from datasets import Dataset

animals = pa.array(["Flamingo", "Horse", "Brittle stars", "Centipede"]).cast(
    pa.dictionary(pa.int32(), pa.string())
)
t = pa.Table.from_arrays([animals], names=["animal"])
ds = Dataset(t)  # has 'string' feature type
ds.add_item({"animal": "Pikachu"})

currently it raises ArrowTypeError: Unable to merge: Field animal has incompatible types: dictionary<values=string, indices=int32, ordered=0> vs string because the ds.data is still the dictionary type

[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.

[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)	0.005388 / 0.011353 (-0.005965)	0.004004 / 0.011008 (-0.007005)	0.064037 / 0.038508 (0.025529)	0.031666 / 0.023109 (0.008557)	0.236493 / 0.275898 (-0.039405)	0.269047 / 0.323480 (-0.054432)	0.005008 / 0.007986 (-0.002977)	0.002964 / 0.004328 (-0.001364)	0.049926 / 0.004250 (0.045675)	0.048092 / 0.037052 (0.011039)	0.245563 / 0.258489 (-0.012926)	0.282614 / 0.293841 (-0.011227)	0.027488 / 0.128546 (-0.101058)	0.010904 / 0.075646 (-0.064742)	0.204892 / 0.419271 (-0.214379)	0.037161 / 0.043533 (-0.006372)	0.238488 / 0.255139 (-0.016651)	0.258192 / 0.283200 (-0.025008)	0.018819 / 0.141683 (-0.122864)	1.131573 / 1.452155 (-0.320582)	1.204084 / 1.492716 (-0.288632)

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)	0.095852 / 0.018006 (0.077846)	0.300225 / 0.000490 (0.299735)	0.000217 / 0.000200 (0.000017)	0.000052 / 0.000054 (-0.000002)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.018592 / 0.037411 (-0.018819)	0.062297 / 0.014526 (0.047772)	0.074344 / 0.176557 (-0.102212)	0.120654 / 0.737135 (-0.616481)	0.075567 / 0.296338 (-0.220772)

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)	0.287700 / 0.215209 (0.072491)	2.829536 / 2.077655 (0.751882)	1.446296 / 1.504120 (-0.057824)	1.320912 / 1.541195 (-0.220283)	1.362744 / 1.468490 (-0.105746)	0.563732 / 4.584777 (-4.021045)	2.399904 / 3.745712 (-1.345808)	2.676706 / 5.269862 (-2.593156)	1.744780 / 4.565676 (-2.820896)	0.062884 / 0.424275 (-0.361391)	0.004936 / 0.007607 (-0.002671)	0.338084 / 0.226044 (0.112040)	3.309532 / 2.268929 (1.040603)	1.792791 / 55.444624 (-53.651833)	1.502038 / 6.876477 (-5.374439)	1.662417 / 2.142072 (-0.479655)	0.642835 / 4.805227 (-4.162393)	0.117002 / 6.500664 (-6.383662)	0.041880 / 0.075469 (-0.033589)

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)	0.974814 / 1.841788 (-0.866974)	11.430883 / 8.074308 (3.356575)	10.314734 / 10.191392 (0.123342)	0.139838 / 0.680424 (-0.540586)	0.014939 / 0.534201 (-0.519262)	0.288048 / 0.579283 (-0.291235)	0.269146 / 0.434364 (-0.165218)	0.324300 / 0.540337 (-0.216037)	0.421612 / 1.386936 (-0.965324)

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)	0.005660 / 0.011353 (-0.005692)	0.003723 / 0.011008 (-0.007285)	0.049909 / 0.038508 (0.011401)	0.033079 / 0.023109 (0.009970)	0.270940 / 0.275898 (-0.004958)	0.291173 / 0.323480 (-0.032307)	0.004336 / 0.007986 (-0.003650)	0.002793 / 0.004328 (-0.001535)	0.049619 / 0.004250 (0.045368)	0.041062 / 0.037052 (0.004010)	0.285026 / 0.258489 (0.026537)	0.322119 / 0.293841 (0.028278)	0.029653 / 0.128546 (-0.098894)	0.010785 / 0.075646 (-0.064861)	0.058680 / 0.419271 (-0.360591)	0.033300 / 0.043533 (-0.010233)	0.269452 / 0.255139 (0.014313)	0.285426 / 0.283200 (0.002226)	0.017655 / 0.141683 (-0.124028)	1.144713 / 1.452155 (-0.307442)	1.196828 / 1.492716 (-0.295888)

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)	0.096719 / 0.018006 (0.078713)	0.303532 / 0.000490 (0.303042)	0.000223 / 0.000200 (0.000023)	0.000049 / 0.000054 (-0.000005)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.022620 / 0.037411 (-0.014791)	0.077057 / 0.014526 (0.062532)	0.088570 / 0.176557 (-0.087987)	0.128715 / 0.737135 (-0.608421)	0.090844 / 0.296338 (-0.205494)

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)	0.298101 / 0.215209 (0.082892)	2.919861 / 2.077655 (0.842206)	1.608945 / 1.504120 (0.104825)	1.487756 / 1.541195 (-0.053439)	1.520800 / 1.468490 (0.052310)	0.576615 / 4.584777 (-4.008162)	0.964250 / 3.745712 (-2.781462)	2.852968 / 5.269862 (-2.416893)	1.868768 / 4.565676 (-2.696908)	0.063934 / 0.424275 (-0.360341)	0.005093 / 0.007607 (-0.002514)	0.352984 / 0.226044 (0.126939)	3.507441 / 2.268929 (1.238513)	1.944467 / 55.444624 (-53.500158)	1.663985 / 6.876477 (-5.212492)	1.847029 / 2.142072 (-0.295043)	0.669228 / 4.805227 (-4.136000)	0.118990 / 6.500664 (-6.381675)	0.041788 / 0.075469 (-0.033681)

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)	1.004541 / 1.841788 (-0.837247)	12.525181 / 8.074308 (4.450873)	10.488167 / 10.191392 (0.296775)	0.141182 / 0.680424 (-0.539242)	0.016432 / 0.534201 (-0.517769)	0.283682 / 0.579283 (-0.295601)	0.128277 / 0.434364 (-0.306087)	0.321933 / 0.540337 (-0.218404)	0.416430 / 1.386936 (-0.970506)

[EthanSteinberg]

@lhoestq Thanks a ton for helping this get merged!