+gdV$ l d dl mZmZmZ ddlmZmZmZ ddlm Z ddl
m Z m Z m
Z
ddlmZ ddlmZ ej e Z edee Z dd ee d eee d
ee dee dee dee defdZ ddee dee dee dedef
dZd S ) )ListOptionalTypeVar )Dataset _concatenate_map_style_datasets _interleave_map_style_datasets) DatasetInfo)IterableDataset _concatenate_iterable_datasets _interleave_iterable_datasets)
NamedSplit)logging DatasetTypeNfirst_exhausteddatasets
probabilitiesseedinfosplitstopping_strategyreturnc : ddl m} ddlm} | st d t | d | }t | d | } || z s%t dt
| d | dd D ]\}
| rt |
| r|rFt |
| s6t dt
| d d t
|
d
]|d vrt | d | rt | |||||
S t | |||||
S )u
Interleave several datasets (sources) into a single dataset.
The new dataset is constructed by alternating between the sources to get the examples.
You can use this function on a list of [`Dataset`] objects, or on a list of [`IterableDataset`] objects.
- If `probabilities` is `None` (default) the new dataset is constructed by cycling between each source to get the examples.
- If `probabilities` is not `None`, the new dataset is constructed by getting examples from a random source at a time according to the provided probabilities.
The resulting dataset ends when one of the source datasets runs out of examples except when `oversampling` is `True`,
in which case, the resulting dataset ends when all datasets have ran out of examples at least one time.
Note for iterable datasets:
In a distributed setup or in PyTorch DataLoader workers, the stopping strategy is applied per process.
Therefore the "first_exhausted" strategy on an sharded iterable dataset can generate less samples in total (up to 1 missing sample per subdataset per worker).
Args:
datasets (`List[Dataset]` or `List[IterableDataset]`):
List of datasets to interleave.
probabilities (`List[float]`, *optional*, defaults to `None`):
If specified, the new dataset is constructed by sampling
examples from one source at a time according to these probabilities.
seed (`int`, *optional*, defaults to `None`):
The random seed used to choose a source for each example.
info ([`DatasetInfo`], *optional*):
Dataset information, like description, citation, etc.
split ([`NamedSplit`], *optional*):
Name of the dataset split.
stopping_strategy (`str`, *optional*, defaults to `first_exhausted`):
Two strategies are proposed right now, `first_exhausted` and `all_exhausted`.
By default, `first_exhausted` is an undersampling strategy, i.e the dataset construction is stopped as soon as one dataset has ran out of samples.
If the strategy is `all_exhausted`, we use an oversampling strategy, i.e the dataset construction is stopped as soon as every samples of every dataset has been added at least once.
Note that if the strategy is `all_exhausted`, the interleaved dataset size can get enormous:
- with no probabilities, the resulting dataset will have `max_length_datasets*nb_dataset` samples.
- with given probabilities, the resulting dataset will have more samples if some datasets have really low probability of visiting.
Returns:
[`Dataset`] or [`IterableDataset`]: Return type depends on the input `datasets`
parameter. `Dataset` if the input is a list of `Dataset`, `IterableDataset` if the input is a list of
`IterableDataset`.
Example:
For regular datasets (map-style):
```python
>>> from datasets import Dataset, interleave_datasets
>>> d1 = Dataset.from_dict({"a": [0, 1, 2]})
>>> d2 = Dataset.from_dict({"a": [10, 11, 12]})
>>> d3 = Dataset.from_dict({"a": [20, 21, 22]})
>>> dataset = interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42, stopping_strategy="all_exhausted")
>>> dataset["a"]
[10, 0, 11, 1, 2, 20, 12, 10, 0, 1, 2, 21, 0, 11, 1, 2, 0, 1, 12, 2, 10, 0, 22]
>>> dataset = interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42)
>>> dataset["a"]
[10, 0, 11, 1, 2]
>>> dataset = interleave_datasets([d1, d2, d3])
>>> dataset["a"]
[0, 10, 20, 1, 11, 21, 2, 12, 22]
>>> dataset = interleave_datasets([d1, d2, d3], stopping_strategy="all_exhausted")
>>> dataset["a"]
[0, 10, 20, 1, 11, 21, 2, 12, 22]
>>> d1 = Dataset.from_dict({"a": [0, 1, 2]})
>>> d2 = Dataset.from_dict({"a": [10, 11, 12, 13]})
>>> d3 = Dataset.from_dict({"a": [20, 21, 22, 23, 24]})
>>> dataset = interleave_datasets([d1, d2, d3])
>>> dataset["a"]
[0, 10, 20, 1, 11, 21, 2, 12, 22]
>>> dataset = interleave_datasets([d1, d2, d3], stopping_strategy="all_exhausted")
>>> dataset["a"]
[0, 10, 20, 1, 11, 21, 2, 12, 22, 0, 13, 23, 1, 10, 24]
>>> dataset = interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42)
>>> dataset["a"]
[10, 0, 11, 1, 2]
>>> dataset = interleave_datasets([d1, d2, d3], probabilities=[0.7, 0.2, 0.1], seed=42, stopping_strategy="all_exhausted")
>>> dataset["a"]
[10, 0, 11, 1, 2, 20, 12, 13, ..., 0, 1, 2, 0, 24]
For datasets in streaming mode (iterable):
>>> from datasets import load_dataset, interleave_datasets
>>> d1 = load_dataset("oscar", "unshuffled_deduplicated_en", split="train", streaming=True)
>>> d2 = load_dataset("oscar", "unshuffled_deduplicated_fr", split="train", streaming=True)
>>> dataset = interleave_datasets([d1, d2])
>>> iterator = iter(dataset)
>>> next(iterator)
{'text': 'Mtendere Village was inspired by the vision...}
>>> next(iterator)
{'text': "Média de débat d'idées, de culture...}
```
r )r )r z/Unable to interleave an empty list of datasets.r `Expected a list of Dataset objects or a list of IterableDataset objects, but first element is a NzUnable to interleave a with a J. Expected a list of Dataset objects or a list of IterableDataset objects.)r
all_exhaustedz: is not supported. Please enter a valid stopping_strategy.)r r r )
arrow_datasetr iterable_datasetr
ValueError
isinstancetyper r
) r r r r r r r r iterable map_styledatasets 0lib/python3.11/site-packages/datasets/combine.pyinterleave_datasetsr' s H '&&&&&111111 LJKKK(1+77H8A;00I y
Cost|}~t pA pA
C
C
ABB< j' : : Q[\cetQuQu _$x{*;*; _ _T']] _ _ _
DDD -iiijjj
- mTEUf
- mTEUf
dsetsaxisc | st d t | d t }t | d t }||z s%t dt | d | dd D ]f}|rt |t r|rKt |t s6t dt | d dt | d g|rt | ||| S t
| ||| S )
a
Converts a list of [`Dataset`] with the same schema into a single [`Dataset`].
Args:
dsets (`List[datasets.Dataset]`):
List of Datasets to concatenate.
info (`DatasetInfo`, *optional*):
Dataset information, like description, citation, etc.
split (`NamedSplit`, *optional*):
Name of the dataset split.
axis (`{0, 1}`, defaults to `0`):
Axis to concatenate over, where `0` means over rows (vertically) and `1` means over columns
(horizontally).
Example:
```py
>>> ds3 = concatenate_datasets([ds1, ds2])
```
z0Unable to concatenate an empty list of datasets.r r r NzUnable to concatenate a r r )r r r* )r r! r r r" r r )r) r r r* r# r$ r% s r& concatenate_datasetsr, sO : MKLLL%(O44H58W--I y
ostyz{t|o}o}
9 j' : : Q[\cetQuQu ]4a>> ] ]4== ] ] ]
X.u4uSWXXXX-e$eRVWWWWr( )NNNNr )NNr )typingr r r r r r r r r
r r r r
splitsr utilsr
get_logger__name__loggerr floatintstrr' r, r( r& r7 s * * * * * * * * * * c c c c c c c c c c l l l l l l l l l l
H % % gmWo>>
,0 "&"&'8
}
}
; }
DK(}
3-}
;
}
J }
}
}
}
}
}
}
D #'"& -X -X -X
;
-X J -X -X
-X -X -X -X -X -Xr(