Using a registry instead of calling globals for fetching feature types

[psmyth94]

Hello,

When working with bio-data, each feature often has metadata associated with it (e.g. species, lineage, snp position, etc). To store this, I like to use the feature classes with the added metadata attribute. However, when saving or loading with custom features, you get an error since that class doesn't exist in the global namespace in datasets.features.features. Take for example,

from dataclasses import dataclass, field
from datasets import Dataset
from datasets.features.features import Value, Features

@dataclass
class FeatureA(Value):
    metadata: dict = field(default=dict)
    _type: str = field(default="FeatureA", init=False, repr=False)

@dataclass
class FeatureB(Value):
    metadata: dict = field(default_factory=dict)
    _type: str = field(default="FeatureB", init=False, repr=False)

test_data = {
    "a": [1, 2, 3],
    "b": [4, 5, 6],
}
test_data = Dataset.from_dict(
    test_data, 
    features=Features({
        "a": FeatureA("int32", metadata={"species": "lactobacillus acetotolerans"}),
        "b": FeatureB("int32", metadata={"species": "lactobacillus iners"}),
    })
)

# returns an error since FeatureA and FeatureB are not in the global namespace
test_data.save_to_disk('./test_data')

Saving the dataset (0/1 shards):   0%|          | 0/3 [00:00<?, ? examples/s]



---------------------------------------------------------------------------

KeyError                                  Traceback (most recent call last)

Cell In[2], line 28
     19 test_data = Dataset.from_dict(
     20     test_data, 
     21     features=Features({
   (...)
     24     })
     25 )
     27 # returns an error since FeatureA and FeatureB are not in the global namespace
---> 28 test_data.save_to_disk('./test_data')
...
File ~\Documents\datasets\src\datasets\features\features.py:1361, in generate_from_dict(obj)
   1359     return {key: generate_from_dict(value) for key, value in obj.items()}
   1360 obj = dict(obj)
-> 1361 class_type = globals()[obj.pop("_type")]
   1363 if class_type == Sequence:
   1364     return Sequence(feature=generate_from_dict(obj["feature"]), length=obj.get("length", -1))


KeyError: 'FeatureA'

We can avoid this by having a registry (like formatters) and doing

from datasets.features.features import register_feature
register_feature(FeatureA, "FeatureA")
register_feature(FeatureB, "FeatureB")
test_data.save_to_disk('./test_data')

Saving the dataset (1/1 shards): 100%|------| 3/3 [00:00<00:00, 211.13 examples/s]

and loading from disk returns with all metadata information

from datasets import load_from_disk
test_data = load_from_disk('./test_data')
test_data.features

{'a': FeatureA(dtype='int32', id=None, metadata={'species': 'lactobacillus acetotolerans'}),
 'b': FeatureB(dtype='int32', id=None, metadata={'species': 'lactobacillus iners'})}

[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]

Cool ! Though for now I'd keep this feature experimental if it's fine for you.

I just added a few comments:

[psmyth94]

looks like some files are missing in your google storage

[lhoestq]

cc @mariosasko is it related to #6474 ? The files should ideally not move for backward compatibility anyway

[mariosasko]

@lhoestq All the files are still there.

The problem is that the natural_questions is now a no-code dataset, so the test's paths are no longer correct (unless the revision is pinned to the previous version).

@psmyth94 This has been fixed on main, so you can make the CI tests green with the following:

git remote add upstream https://github.com/huggingface/datasets.git
git pull upstream main
git push

[lhoestq]

Thank you @mariosasko ! I'm updating this branch if you don't mind @psmyth94

[lhoestq]

Looks all good now, thanks !

[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)	0.004903 / 0.011353 (-0.006450)	0.003105 / 0.011008 (-0.007903)	0.061980 / 0.038508 (0.023471)	0.029726 / 0.023109 (0.006617)	0.243406 / 0.275898 (-0.032492)	0.262530 / 0.323480 (-0.060950)	0.003905 / 0.007986 (-0.004081)	0.002617 / 0.004328 (-0.001712)	0.047851 / 0.004250 (0.043601)	0.040397 / 0.037052 (0.003345)	0.259461 / 0.258489 (0.000972)	0.285059 / 0.293841 (-0.008782)	0.027321 / 0.128546 (-0.101225)	0.009876 / 0.075646 (-0.065770)	0.206999 / 0.419271 (-0.212273)	0.034906 / 0.043533 (-0.008626)	0.245120 / 0.255139 (-0.010019)	0.270490 / 0.283200 (-0.012710)	0.017341 / 0.141683 (-0.124342)	1.128182 / 1.452155 (-0.323973)	1.173024 / 1.492716 (-0.319693)

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)	0.089337 / 0.018006 (0.071331)	0.298256 / 0.000490 (0.297767)	0.000216 / 0.000200 (0.000016)	0.000047 / 0.000054 (-0.000007)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.018179 / 0.037411 (-0.019233)	0.061275 / 0.014526 (0.046749)	0.073137 / 0.176557 (-0.103419)	0.119603 / 0.737135 (-0.617532)	0.073969 / 0.296338 (-0.222370)

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)	0.283109 / 0.215209 (0.067900)	2.765441 / 2.077655 (0.687787)	1.471276 / 1.504120 (-0.032844)	1.346365 / 1.541195 (-0.194830)	1.360668 / 1.468490 (-0.107822)	0.549947 / 4.584777 (-4.034830)	2.344213 / 3.745712 (-1.401499)	2.700905 / 5.269862 (-2.568956)	1.689936 / 4.565676 (-2.875741)	0.061985 / 0.424275 (-0.362290)	0.004923 / 0.007607 (-0.002684)	0.329833 / 0.226044 (0.103788)	3.277580 / 2.268929 (1.008652)	1.833987 / 55.444624 (-53.610638)	1.571023 / 6.876477 (-5.305454)	1.573259 / 2.142072 (-0.568813)	0.627504 / 4.805227 (-4.177723)	0.114106 / 6.500664 (-6.386558)	0.041197 / 0.075469 (-0.034272)

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)	0.967400 / 1.841788 (-0.874388)	11.046527 / 8.074308 (2.972219)	9.542214 / 10.191392 (-0.649178)	0.140745 / 0.680424 (-0.539679)	0.013627 / 0.534201 (-0.520574)	0.288429 / 0.579283 (-0.290855)	0.260509 / 0.434364 (-0.173855)	0.324704 / 0.540337 (-0.215633)	0.419366 / 1.386936 (-0.967570)

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)	0.005123 / 0.011353 (-0.006230)	0.003119 / 0.011008 (-0.007890)	0.048931 / 0.038508 (0.010423)	0.032067 / 0.023109 (0.008958)	0.276825 / 0.275898 (0.000927)	0.297589 / 0.323480 (-0.025890)	0.004075 / 0.007986 (-0.003911)	0.002579 / 0.004328 (-0.001750)	0.047862 / 0.004250 (0.043612)	0.044032 / 0.037052 (0.006980)	0.289469 / 0.258489 (0.030980)	0.327269 / 0.293841 (0.033428)	0.029369 / 0.128546 (-0.099177)	0.010180 / 0.075646 (-0.065466)	0.057111 / 0.419271 (-0.362161)	0.051046 / 0.043533 (0.007513)	0.276758 / 0.255139 (0.021619)	0.296084 / 0.283200 (0.012884)	0.017376 / 0.141683 (-0.124306)	1.154486 / 1.452155 (-0.297669)	1.192699 / 1.492716 (-0.300018)

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)	0.085981 / 0.018006 (0.067974)	0.296956 / 0.000490 (0.296466)	0.000211 / 0.000200 (0.000011)	0.000050 / 0.000054 (-0.000004)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.021239 / 0.037411 (-0.016172)	0.074851 / 0.014526 (0.060326)	0.085676 / 0.176557 (-0.090881)	0.125876 / 0.737135 (-0.611259)	0.087573 / 0.296338 (-0.208765)

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)	0.289220 / 0.215209 (0.074011)	2.812342 / 2.077655 (0.734688)	1.572886 / 1.504120 (0.068766)	1.446442 / 1.541195 (-0.094752)	1.458737 / 1.468490 (-0.009753)	0.562010 / 4.584777 (-4.022767)	2.422896 / 3.745712 (-1.322816)	2.578408 / 5.269862 (-2.691454)	1.689998 / 4.565676 (-2.875678)	0.064782 / 0.424275 (-0.359493)	0.005051 / 0.007607 (-0.002556)	0.339982 / 0.226044 (0.113938)	3.309882 / 2.268929 (1.040953)	1.910273 / 55.444624 (-53.534351)	1.649723 / 6.876477 (-5.226753)	1.744073 / 2.142072 (-0.397999)	0.651905 / 4.805227 (-4.153323)	0.114606 / 6.500664 (-6.386058)	0.040030 / 0.075469 (-0.035439)

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)	1.008374 / 1.841788 (-0.833414)	11.547300 / 8.074308 (3.472992)	9.966061 / 10.191392 (-0.225331)	0.144874 / 0.680424 (-0.535550)	0.014400 / 0.534201 (-0.519801)	0.285435 / 0.579283 (-0.293848)	0.274755 / 0.434364 (-0.159609)	0.323105 / 0.540337 (-0.217232)	0.439172 / 1.386936 (-0.947764)