15883725135

Committed 25 Jun 2025 06:03PM UTC coverage: 81.096% (+1.3%) from 79.803%

Build # 15883725135

Build Type

push

github

Committed by

web-flow

Commit Message

Use full artifact representation as the cache key for the dataset (#1644)

* Use elaborated cache key and use it for filelock

Signed-off-by: elronbandel <elronbandel@gmail.com>

* Add documentation

Signed-off-by: elronbandel <elronbandel@gmail.com>

* Another try

Signed-off-by: elronbandel <elronbandel@gmail.com>

* Remove file lock

Signed-off-by: elronbandel <elronbandel@gmail.com>

* Update

Signed-off-by: elronbandel <elronbandel@gmail.com>

* Update

Signed-off-by: elronbandel <elronbandel@gmail.com>

* Update

Signed-off-by: elronbandel <elronbandel@gmail.com>

* Format

Signed-off-by: elronbandel <elronbandel@gmail.com>

---------

Signed-off-by: elronbandel <elronbandel@gmail.com>

Coverage Stats

1538 of 1909 branches covered (80.57%)

Branch coverage included in aggregate %.

10461 of 12887 relevant lines covered (81.17%)

0.81 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

78.0

src/unitxt/api.py

import hashlib
import inspect
import json
from datetime import datetime
from functools import lru_cache
from typing import Any, Dict, List, Optional, Union

from datasets import Dataset, DatasetDict, IterableDataset, IterableDatasetDict
from datasets.exceptions import DatasetGenerationError

from .artifact import fetch_artifact
from .benchmark import Benchmark
from .card import TaskCard
from .dataclass import to_dict
from .dataset_utils import get_dataset_artifact
from .error_utils import UnitxtError
from .inference import (
    InferenceEngine,
    LogProbInferenceEngine,
    OptionSelectingByLogProbsInferenceEngine,
)
from .loaders import LoadFromDictionary
from .logging_utils import get_logger
from .metric_utils import EvaluationResults, _compute, _inference_post_process
from .operator import SourceOperator
from .schema import loads_batch
from .settings_utils import get_constants, get_settings
from .standard import DatasetRecipe
from .task import Task

logger = get_logger()
constants = get_constants()
settings = get_settings()


def short_hex_hash(value, length=8):
    h = hashlib.sha256(value.encode()).hexdigest()  # Full 64-character hex
    return h[:length]


def _get_recipe_from_query(
    dataset_query: str, overwrite_kwargs: Optional[Dict[str, Any]] = None
) -> DatasetRecipe:
    try:
        dataset_stream, _ = fetch_artifact(
            dataset_query, overwrite_kwargs=overwrite_kwargs
        )
    except:
        dataset_stream = get_dataset_artifact(
            dataset_query, overwrite_kwargs=overwrite_kwargs
        )
    return dataset_stream


def _get_recipe_from_dict(dataset_params: Dict[str, Any]) -> DatasetRecipe:
    recipe_attributes = list(DatasetRecipe.__dict__["__fields__"].keys())
    for param in dataset_params.keys():
        assert param in recipe_attributes, (
            f"The parameter '{param}' is not an attribute of the 'DatasetRecipe' class. "
            f"Please check if the name is correct. The available attributes are: '{recipe_attributes}'."
        )
    return DatasetRecipe(**dataset_params)


def _verify_dataset_args(dataset_query: Optional[str] = None, dataset_args=None):
    if dataset_query and dataset_args:
        raise ValueError(
            "Cannot provide 'dataset_query' and key-worded arguments at the same time. "
            "If you want to load dataset from a card in local catalog, use query only. "
            "Otherwise, use key-worded arguments only to specify properties of dataset."
        )

    if dataset_query:
        if not isinstance(dataset_query, str):
            raise ValueError(
                f"If specified, 'dataset_query' must be a string, however, "
                f"'{dataset_query}' was provided instead, which is of type "
                f"'{type(dataset_query)}'."
            )

    if not dataset_query and not dataset_args:
        raise ValueError(
            "Either 'dataset_query' or key-worded arguments must be provided."
        )


def load_recipe(dataset_query: Optional[str] = None, **kwargs) -> DatasetRecipe:
    if isinstance(dataset_query, (DatasetRecipe, Benchmark)):
        return dataset_query

    if dataset_query:
        recipe = _get_recipe_from_query(dataset_query, kwargs)

    elif kwargs:
        recipe = _get_recipe_from_dict(kwargs)

    else:
        raise UnitxtError(
            "Specify either dataset recipe string artifact name or recipe args."
        )

    return recipe


def create_dataset(
    task: Union[str, Task],
    test_set: List[Dict[Any, Any]],
    train_set: Optional[List[Dict[Any, Any]]] = None,
    validation_set: Optional[List[Dict[Any, Any]]] = None,
    split: Optional[str] = None,
    data_classification_policy: Optional[List[str]] = None,
    **kwargs,
) -> Union[DatasetDict, IterableDatasetDict, Dataset, IterableDataset]:
    """Creates dataset from input data based on a specific task.

    Args:
        task:  The name of the task from the Unitxt Catalog (https://www.unitxt.ai/en/latest/catalog/catalog.tasks.__dir__.html)
        test_set : required list of instances
        train_set : optional train_set
        validation_set: optional validation set
        split: optional one split to choose
        data_classification_policy: data_classification_policy
        **kwargs: Arguments used to load dataset from provided datasets (see load_dataset())

    Returns:
        DatasetDict

    Example:
        template = Template(...)
        dataset = create_dataset(task="tasks.qa.open", template=template, format="formats.chatapi")
    """
    data = {"test": test_set}
    if train_set is not None:
        data["train"] = train_set
    if validation_set is not None:
        data["validation"] = validation_set
    task, _ = fetch_artifact(task)

    if "template" not in kwargs and task.default_template is None:
        raise Exception(
            f"No 'template' was passed to the create_dataset() and the given task ('{task.__id__}') has no 'default_template' field."
        )

    card = TaskCard(
        loader=LoadFromDictionary(
            data=data, data_classification_policy=data_classification_policy
        ),
        task=task,
    )
    return load_dataset(card=card, split=split, **kwargs)


def object_to_str_without_addresses(obj):
    """Generates a string representation of a Python object while removing memory address references.

    This function is useful for creating consistent and comparable string representations of objects
    that would otherwise include memory addresses (e.g., `<object_name at 0x123abc>`), which can vary
    between executions. By stripping the memory address, the function ensures that the representation
    is stable and independent of the object's location in memory.

    Args:
        obj: Any Python object to be converted to a string representation.

    Returns:
        str: A string representation of the object with memory addresses removed if present.

    Example:
        ```python
        class MyClass:
            pass

        obj = MyClass()
        print(str(obj))  # "<__main__.MyClass object at 0x7f8b9d4d6e20>"
        print(to_str_without_addresses(obj))  # "<__main__.MyClass object>"
        ```
    """
    obj_str = str(obj)
    if " at 0x" in obj_str:
        obj_str = obj_str.split(" at 0x")[0] + ">"
    return obj_str


def _source_to_dataset(
    source: SourceOperator,
    split=None,
    use_cache=False,
    streaming=False,
):
    from .dataset import Dataset as UnitxtDataset

    # Generate a unique signature for the source
    source_signature = json.dumps(
        to_dict(source, object_to_str_without_addresses), sort_keys=True
    )
    config_name = "recipe-" + short_hex_hash(source_signature)
    # Obtain data stream from the source
    stream = source()

    try:
        ds_builder = UnitxtDataset(
            dataset_name="unitxt",
            config_name=config_name,  # Dictate the cache name
            version=constants.version,
        )
        if split is not None:
            stream = {split: stream[split]}
        ds_builder._generators = stream

        try:
            ds_builder.download_and_prepare(
                verification_mode="no_checks",
                download_mode=None if use_cache else "force_redownload",
            )
        except DatasetGenerationError as e:
            if e.__cause__:
                raise e.__cause__ from None
            if e.__context__:
                raise e.__context__ from None
            raise

        if streaming:
            return ds_builder.as_streaming_dataset(split=split)

        return ds_builder.as_dataset(
            split=split, run_post_process=False, verification_mode="no_checks"
        )

    except DatasetGenerationError as e:
        raise e.__cause__


def load_dataset(
    dataset_query: Optional[str] = None,
    split: Optional[str] = None,
    streaming: bool = False,
    use_cache: Optional[bool] = False,
    **kwargs,
) -> Union[DatasetDict, IterableDatasetDict, Dataset, IterableDataset]:
    """Loads dataset.

    If the 'dataset_query' argument is provided, then dataset is loaded from a card
    in local catalog based on parameters specified in the query.

    Alternatively, dataset is loaded from a provided card based on explicitly
    given parameters.

    If both are given, then the textual recipe is loaded with the key word args overriding the textual recipe args.

    Args:
        dataset_query (str, optional):
            A string query which specifies a dataset to load from
            local catalog or name of specific recipe or benchmark in the catalog. For
            example, ``"card=cards.wnli,template=templates.classification.multi_class.relation.default"``.
        streaming (bool, False):
            When True yields the data as a stream.
            This is useful when loading very large datasets.
            Loading datasets as streams avoid loading all the data to memory, but requires the dataset's loader to support streaming.
        split (str, optional):
            The split of the data to load
        use_cache (bool, optional):
            If set to True, the returned Huggingface dataset is cached on local disk such that if the same dataset is loaded again, it will be loaded from local disk, resulting in faster runs.
            If set to False (default), the returned dataset is not cached.
            Note that if caching is enabled and the dataset card definition is changed, the old version in the cache may be returned.
            Enable caching only if you are sure you are working with fixed Unitxt datasets and definitions (e.g. running using predefined datasets from the Unitxt catalog).
        **kwargs:
            Arguments used to load dataset from provided card, which is not present in local catalog.

    Returns:
        DatasetDict

    :Example:

        .. code-block:: python

            dataset = load_dataset(
                dataset_query="card=cards.stsb,template=templates.regression.two_texts.simple,max_train_instances=5"
            )  # card and template must be present in local catalog

            # or built programmatically
            card = TaskCard(...)
            template = Template(...)
            loader_limit = 10
            dataset = load_dataset(card=card, template=template, loader_limit=loader_limit)

    """
    recipe = load_recipe(dataset_query, **kwargs)

    dataset = _source_to_dataset(
        source=recipe, split=split, use_cache=use_cache, streaming=streaming
    )

    frame = inspect.currentframe()
    args, _, _, values = inspect.getargvalues(frame)
    all_kwargs = {key: values[key] for key in args if key != "kwargs"}
    all_kwargs.update(kwargs)
    metadata = fill_metadata(**all_kwargs)
    if isinstance(dataset, dict):
        for ds in dataset.values():
            ds.info.description = metadata.copy()
    else:
        dataset.info.description = metadata
    return dataset


def fill_metadata(**kwargs):
    metadata = kwargs.copy()
    metadata["unitxt_version"] = get_constants().version
    metadata["creation_time"] = datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f")[:-3]
    return metadata


def evaluate(
    predictions,
    dataset: Union[Dataset, IterableDataset] = None,
    data=None,
    calc_confidence_intervals: bool = True,
) -> EvaluationResults:
    if dataset is None and data is None:
        raise UnitxtError(message="Specify 'dataset' in evaluate")
    if data is not None:
        dataset = data  # for backward compatibility
    evaluation_result = _compute(
        predictions=predictions,
        references=dataset,
        calc_confidence_intervals=calc_confidence_intervals,
    )
    if hasattr(dataset, "info") and hasattr(dataset.info, "description"):
        evaluation_result.metadata["dataset"] = dataset.info.description
    if hasattr(predictions, "metadata"):
        evaluation_result.metadata["predictions"] = predictions.metadata
    evaluation_result.metadata["creation_time"] = datetime.now().strftime(
        "%Y-%m-%d %H:%M:%S.%f"
    )[:-3]
    return evaluation_result


def post_process(predictions, data) -> List[Dict[str, Any]]:
    return _inference_post_process(predictions=predictions, references=data)


@lru_cache
def _get_produce_with_cache(dataset_query: Optional[str] = None, **kwargs):
    return load_recipe(dataset_query, **kwargs).produce


def produce(
    instance_or_instances, dataset_query: Optional[str] = None, **kwargs
) -> Union[Dataset, Dict[str, Any]]:
    is_list = isinstance(instance_or_instances, list)
    if not is_list:
        instance_or_instances = [instance_or_instances]
    result = _get_produce_with_cache(dataset_query, **kwargs)(instance_or_instances)
    if not is_list:
        return result[0]
    return Dataset.from_list(result).with_transform(loads_batch)


def infer(
    instance_or_instances,
    engine: InferenceEngine,
    dataset_query: Optional[str] = None,
    return_data: bool = False,
    return_log_probs: bool = False,
    return_meta_data: bool = False,
    previous_messages: Optional[List[Dict[str, str]]] = None,
    **kwargs,
):
    dataset = produce(instance_or_instances, dataset_query, **kwargs)
    if previous_messages is not None:

        def add_previous_messages(example, index):
            example["source"] = previous_messages[index] + example["source"]
            return example

        dataset = dataset.map(add_previous_messages, with_indices=True)
    engine, _ = fetch_artifact(engine)
    if return_log_probs:
        if not isinstance(engine, LogProbInferenceEngine):
            raise NotImplementedError(
                f"Error in infer: return_log_probs set to True but supplied engine "
                f"{engine.__class__.__name__} does not support logprobs."
            )
        infer_outputs = engine.infer_log_probs(dataset, return_meta_data)
        raw_predictions = (
            [output.prediction for output in infer_outputs]
            if return_meta_data
            else infer_outputs
        )
        raw_predictions = [
            json.dumps(raw_prediction) for raw_prediction in raw_predictions
        ]
    else:
        infer_outputs = engine.infer(dataset, return_meta_data)
        raw_predictions = (
            [output.prediction for output in infer_outputs]
            if return_meta_data
            else infer_outputs
        )
    predictions = post_process(raw_predictions, dataset)
    if return_data:
        if return_meta_data:
            infer_output_list = [
                infer_output.__dict__ for infer_output in infer_outputs
            ]
            for infer_output in infer_output_list:
                del infer_output["prediction"]
            dataset = dataset.add_column("infer_meta_data", infer_output_list)
        dataset = dataset.add_column("prediction", predictions)
        return dataset.add_column("raw_prediction", raw_predictions)
    return predictions


def select(
    instance_or_instances,
    engine: OptionSelectingByLogProbsInferenceEngine,
    dataset_query: Optional[str] = None,
    return_data: bool = False,
    previous_messages: Optional[List[Dict[str, str]]] = None,
    **kwargs,
):
    dataset = produce(instance_or_instances, dataset_query, **kwargs)
    if previous_messages is not None:

        def add_previous_messages(example, index):
            example["source"] = previous_messages[index] + example["source"]
            return example

        dataset = dataset.map(add_previous_messages, with_indices=True)
    engine, _ = fetch_artifact(engine)
    predictions = engine.select(dataset)
    # predictions = post_process(raw_predictions, dataset)
    if return_data:
        return dataset.add_column("prediction", predictions)
    return predictions

1	import hashlib	1✔
2	import inspect	1✔
3	import json	1✔
4	from datetime import datetime	1✔
5	from functools import lru_cache	1✔
6	from typing import Any, Dict, List, Optional, Union	1✔
7
8	from datasets import Dataset, DatasetDict, IterableDataset, IterableDatasetDict	1✔
9	from datasets.exceptions import DatasetGenerationError	1✔
10
11	from .artifact import fetch_artifact	1✔
12	from .benchmark import Benchmark	1✔
13	from .card import TaskCard	1✔
14	from .dataclass import to_dict	1✔
15	from .dataset_utils import get_dataset_artifact	1✔
16	from .error_utils import UnitxtError	1✔
17	from .inference import (	1✔
18	InferenceEngine,
19	LogProbInferenceEngine,
20	OptionSelectingByLogProbsInferenceEngine,
21	)
22	from .loaders import LoadFromDictionary	1✔
23	from .logging_utils import get_logger	1✔
24	from .metric_utils import EvaluationResults, _compute, _inference_post_process	1✔
25	from .operator import SourceOperator	1✔
26	from .schema import loads_batch	1✔
27	from .settings_utils import get_constants, get_settings	1✔
28	from .standard import DatasetRecipe	1✔
29	from .task import Task	1✔
30
31	logger = get_logger()
32	constants = get_constants()	1✔
33	settings = get_settings()	1✔
34
35
36	def short_hex_hash(value, length=8):	1✔
37	h = hashlib.sha256(value.encode()).hexdigest() # Full 64-character hex	1✔
38	return h[:length]	1✔
39
40
41	def _get_recipe_from_query(	1✔
42	dataset_query: str, overwrite_kwargs: Optional[Dict[str, Any]] = None
43	) -> DatasetRecipe:
44	try:	1✔
45	dataset_stream, _ = fetch_artifact(	1✔
46	dataset_query, overwrite_kwargs=overwrite_kwargs
47	)
48	except:	1✔
49	dataset_stream = get_dataset_artifact(	1✔
50	dataset_query, overwrite_kwargs=overwrite_kwargs
51	)
52	return dataset_stream	1✔
53
54
55	def _get_recipe_from_dict(dataset_params: Dict[str, Any]) -> DatasetRecipe:	1✔
56	recipe_attributes = list(DatasetRecipe.__dict__["__fields__"].keys())	1✔
57	for param in dataset_params.keys():	1✔
58	assert param in recipe_attributes, (	1✔
59	f"The parameter '{param}' is not an attribute of the 'DatasetRecipe' class. "
60	f"Please check if the name is correct. The available attributes are: '{recipe_attributes}'."
61	)
62	return DatasetRecipe(**dataset_params)	1✔
63
64
65	def _verify_dataset_args(dataset_query: Optional[str] = None, dataset_args=None):	1✔
66	if dataset_query and dataset_args:	×
67	raise ValueError(
68	"Cannot provide 'dataset_query' and key-worded arguments at the same time. "
69	"If you want to load dataset from a card in local catalog, use query only. "
70	"Otherwise, use key-worded arguments only to specify properties of dataset."
71	)
72
73	if dataset_query:	×
74	if not isinstance(dataset_query, str):	×
75	raise ValueError(
76	f"If specified, 'dataset_query' must be a string, however, "
77	f"'{dataset_query}' was provided instead, which is of type "
78	f"'{type(dataset_query)}'."
79	)
80
81	if not dataset_query and not dataset_args:	×
82	raise ValueError(
83	"Either 'dataset_query' or key-worded arguments must be provided."
84	)
85
86
87	def load_recipe(dataset_query: Optional[str] = None, **kwargs) -> DatasetRecipe:	1✔
88	if isinstance(dataset_query, (DatasetRecipe, Benchmark)):	1✔
89	return dataset_query	×
90
91	if dataset_query:	1✔
92	recipe = _get_recipe_from_query(dataset_query, kwargs)	1✔
93
94	elif kwargs:	1✔
95	recipe = _get_recipe_from_dict(kwargs)	1✔
96
97	else:
98	raise UnitxtError(	×
99	"Specify either dataset recipe string artifact name or recipe args."
100	)
101
102	return recipe	1✔
103
104
105	def create_dataset(	1✔
106	task: Union[str, Task],
107	test_set: List[Dict[Any, Any]],
108	train_set: Optional[List[Dict[Any, Any]]] = None,
109	validation_set: Optional[List[Dict[Any, Any]]] = None,
110	split: Optional[str] = None,
111	data_classification_policy: Optional[List[str]] = None,
112	**kwargs,
113	) -> Union[DatasetDict, IterableDatasetDict, Dataset, IterableDataset]:
114	"""Creates dataset from input data based on a specific task.
115
116	Args:
117	task: The name of the task from the Unitxt Catalog (https://www.unitxt.ai/en/latest/catalog/catalog.tasks.__dir__.html)
118	test_set : required list of instances
119	train_set : optional train_set
120	validation_set: optional validation set
121	split: optional one split to choose
122	data_classification_policy: data_classification_policy
123	**kwargs: Arguments used to load dataset from provided datasets (see load_dataset())
124
125	Returns:
126	DatasetDict
127
128	Example:
129	template = Template(...)
130	dataset = create_dataset(task="tasks.qa.open", template=template, format="formats.chatapi")
131	"""
132	data = {"test": test_set}	1✔
133	if train_set is not None:	1✔
134	data["train"] = train_set	1✔
135	if validation_set is not None:	1✔
136	data["validation"] = validation_set	1✔
137	task, _ = fetch_artifact(task)	1✔
138
139	if "template" not in kwargs and task.default_template is None:	1✔
140	raise Exception(
141	f"No 'template' was passed to the create_dataset() and the given task ('{task.__id__}') has no 'default_template' field."
142	)
143
144	card = TaskCard(	1✔
145	loader=LoadFromDictionary(
146	data=data, data_classification_policy=data_classification_policy
147	),
148	task=task,
149	)
150	return load_dataset(card=card, split=split, **kwargs)	1✔
151
152
153	def object_to_str_without_addresses(obj):	1✔
154	"""Generates a string representation of a Python object while removing memory address references.
155
156	This function is useful for creating consistent and comparable string representations of objects
157	that would otherwise include memory addresses (e.g., `<object_name at 0x123abc>`), which can vary
158	between executions. By stripping the memory address, the function ensures that the representation
159	is stable and independent of the object's location in memory.
160
161	Args:
162	obj: Any Python object to be converted to a string representation.
163
164	Returns:
165	str: A string representation of the object with memory addresses removed if present.
166
167	Example:
168	```python
169	class MyClass:
170	pass
171
172	obj = MyClass()
173	print(str(obj)) # "<__main__.MyClass object at 0x7f8b9d4d6e20>"
174	print(to_str_without_addresses(obj)) # "<__main__.MyClass object>"
175	```
176	"""
177	obj_str = str(obj)	1✔
178	if " at 0x" in obj_str:	1✔
179	obj_str = obj_str.split(" at 0x")[0] + ">"	1✔
180	return obj_str	1✔
181
182
183	def _source_to_dataset(	1✔
184	source: SourceOperator,
185	split=None,
186	use_cache=False,
187	streaming=False,
188	):
189	from .dataset import Dataset as UnitxtDataset	1✔
190
191	# Generate a unique signature for the source
192	source_signature = json.dumps(	1✔
193	to_dict(source, object_to_str_without_addresses), sort_keys=True
194	)
195	config_name = "recipe-" + short_hex_hash(source_signature)	1✔
196	# Obtain data stream from the source
197	stream = source()	1✔
198
199	try:	1✔
200	ds_builder = UnitxtDataset(	1✔
201	dataset_name="unitxt",
202	config_name=config_name, # Dictate the cache name
203	version=constants.version,
204	)
205	if split is not None:	1✔
206	stream = {split: stream[split]}	1✔
207	ds_builder._generators = stream	1✔
208
209	try:	1✔
210	ds_builder.download_and_prepare(	1✔
211	verification_mode="no_checks",
212	download_mode=None if use_cache else "force_redownload",
213	)
214	except DatasetGenerationError as e:	×
215	if e.__cause__:	×
216	raise e.__cause__ from None	×
217	if e.__context__:	×
218	raise e.__context__ from None	×
219	raise	×
220
221	if streaming:	1✔
222	return ds_builder.as_streaming_dataset(split=split)	×
223
224	return ds_builder.as_dataset(	1✔
225	split=split, run_post_process=False, verification_mode="no_checks"
226	)
227
228	except DatasetGenerationError as e:	×
229	raise e.__cause__	×
230
231
232	def load_dataset(	1✔
233	dataset_query: Optional[str] = None,
234	split: Optional[str] = None,
235	streaming: bool = False,
236	use_cache: Optional[bool] = False,
237	**kwargs,
238	) -> Union[DatasetDict, IterableDatasetDict, Dataset, IterableDataset]:
239	"""Loads dataset.
240
241	If the 'dataset_query' argument is provided, then dataset is loaded from a card
242	in local catalog based on parameters specified in the query.
243
244	Alternatively, dataset is loaded from a provided card based on explicitly
245	given parameters.
246
247	If both are given, then the textual recipe is loaded with the key word args overriding the textual recipe args.
248
249	Args:
250	dataset_query (str, optional):
251	A string query which specifies a dataset to load from
252	local catalog or name of specific recipe or benchmark in the catalog. For
253	example, ``"card=cards.wnli,template=templates.classification.multi_class.relation.default"``.
254	streaming (bool, False):
255	When True yields the data as a stream.
256	This is useful when loading very large datasets.
257	Loading datasets as streams avoid loading all the data to memory, but requires the dataset's loader to support streaming.
258	split (str, optional):
259	The split of the data to load
260	use_cache (bool, optional):
261	If set to True, the returned Huggingface dataset is cached on local disk such that if the same dataset is loaded again, it will be loaded from local disk, resulting in faster runs.
262	If set to False (default), the returned dataset is not cached.
263	Note that if caching is enabled and the dataset card definition is changed, the old version in the cache may be returned.
264	Enable caching only if you are sure you are working with fixed Unitxt datasets and definitions (e.g. running using predefined datasets from the Unitxt catalog).
265	**kwargs:
266	Arguments used to load dataset from provided card, which is not present in local catalog.
267
268	Returns:
269	DatasetDict
270
271	:Example:
272
273	.. code-block:: python
274
275	dataset = load_dataset(
276	dataset_query="card=cards.stsb,template=templates.regression.two_texts.simple,max_train_instances=5"
277	) # card and template must be present in local catalog
278
279	# or built programmatically
280	card = TaskCard(...)
281	template = Template(...)
282	loader_limit = 10
283	dataset = load_dataset(card=card, template=template, loader_limit=loader_limit)
284
285	"""
286	recipe = load_recipe(dataset_query, **kwargs)	1✔
287
288	dataset = _source_to_dataset(	1✔
289	source=recipe, split=split, use_cache=use_cache, streaming=streaming
290	)
291
292	frame = inspect.currentframe()	1✔
293	args, _, _, values = inspect.getargvalues(frame)	1✔
294	all_kwargs = {key: values[key] for key in args if key != "kwargs"}	1✔
295	all_kwargs.update(kwargs)	1✔
296	metadata = fill_metadata(**all_kwargs)	1✔
297	if isinstance(dataset, dict):	1✔
298	for ds in dataset.values():	1✔
299	ds.info.description = metadata.copy()	1✔
300	else:
301	dataset.info.description = metadata	1✔
302	return dataset	1✔
303
304
305	def fill_metadata(**kwargs):	1✔
306	metadata = kwargs.copy()	1✔
307	metadata["unitxt_version"] = get_constants().version	1✔
308	metadata["creation_time"] = datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f")[:-3]	1✔
309	return metadata	1✔
310
311
312	def evaluate(	1✔
313	predictions,
314	dataset: Union[Dataset, IterableDataset] = None,
315	data=None,
316	calc_confidence_intervals: bool = True,
317	) -> EvaluationResults:
318	if dataset is None and data is None:	1✔
319	raise UnitxtError(message="Specify 'dataset' in evaluate")	×
320	if data is not None:	1✔
321	dataset = data # for backward compatibility	1✔
322	evaluation_result = _compute(	1✔
323	predictions=predictions,
324	references=dataset,
325	calc_confidence_intervals=calc_confidence_intervals,
326	)
327	if hasattr(dataset, "info") and hasattr(dataset.info, "description"):	1✔
328	evaluation_result.metadata["dataset"] = dataset.info.description	1✔
329	if hasattr(predictions, "metadata"):	1✔
330	evaluation_result.metadata["predictions"] = predictions.metadata	×
331	evaluation_result.metadata["creation_time"] = datetime.now().strftime(	1✔
332	"%Y-%m-%d %H:%M:%S.%f"
333	)[:-3]
334	return evaluation_result	1✔
335
336
337	def post_process(predictions, data) -> List[Dict[str, Any]]:	1✔
338	return _inference_post_process(predictions=predictions, references=data)	1✔
339
340
341	@lru_cache	1✔
342	def _get_produce_with_cache(dataset_query: Optional[str] = None, **kwargs):	1✔
343	return load_recipe(dataset_query, **kwargs).produce	1✔
344
345
346	def produce(	1✔
347	instance_or_instances, dataset_query: Optional[str] = None, **kwargs
348	) -> Union[Dataset, Dict[str, Any]]:
349	is_list = isinstance(instance_or_instances, list)	1✔
350	if not is_list:	1✔
351	instance_or_instances = [instance_or_instances]	1✔
352	result = _get_produce_with_cache(dataset_query, **kwargs)(instance_or_instances)	1✔
353	if not is_list:	1✔
354	return result[0]	1✔
355	return Dataset.from_list(result).with_transform(loads_batch)	1✔
356
357
358	def infer(	1✔
359	instance_or_instances,
360	engine: InferenceEngine,
361	dataset_query: Optional[str] = None,
362	return_data: bool = False,
363	return_log_probs: bool = False,
364	return_meta_data: bool = False,
365	previous_messages: Optional[List[Dict[str, str]]] = None,
366	**kwargs,
367	):
368	dataset = produce(instance_or_instances, dataset_query, **kwargs)	1✔
369	if previous_messages is not None:	1✔
370
371	def add_previous_messages(example, index):	1✔
372	example["source"] = previous_messages[index] + example["source"]	1✔
373	return example	1✔
374
375	dataset = dataset.map(add_previous_messages, with_indices=True)	1✔
376	engine, _ = fetch_artifact(engine)	1✔
377	if return_log_probs:	1✔
378	if not isinstance(engine, LogProbInferenceEngine):	×
379	raise NotImplementedError(	×
380	f"Error in infer: return_log_probs set to True but supplied engine "
381	f"{engine.__class__.__name__} does not support logprobs."
382	)
383	infer_outputs = engine.infer_log_probs(dataset, return_meta_data)	×
384	raw_predictions = (	×
385	[output.prediction for output in infer_outputs]
386	if return_meta_data
387	else infer_outputs
388	)
389	raw_predictions = [	×
390	json.dumps(raw_prediction) for raw_prediction in raw_predictions
391	]
392	else:
393	infer_outputs = engine.infer(dataset, return_meta_data)	1✔
394	raw_predictions = (	1✔
395	[output.prediction for output in infer_outputs]
396	if return_meta_data
397	else infer_outputs
398	)
399	predictions = post_process(raw_predictions, dataset)	1✔
400	if return_data:	1✔
401	if return_meta_data:	1✔
402	infer_output_list = [	×
403	infer_output.__dict__ for infer_output in infer_outputs
404	]
405	for infer_output in infer_output_list:	×
406	del infer_output["prediction"]	×
407	dataset = dataset.add_column("infer_meta_data", infer_output_list)	×
408	dataset = dataset.add_column("prediction", predictions)	1✔
409	return dataset.add_column("raw_prediction", raw_predictions)	1✔
410	return predictions	1✔
411
412
413	def select(	1✔
414	instance_or_instances,
415	engine: OptionSelectingByLogProbsInferenceEngine,
416	dataset_query: Optional[str] = None,
417	return_data: bool = False,
418	previous_messages: Optional[List[Dict[str, str]]] = None,
419	**kwargs,
420	):
421	dataset = produce(instance_or_instances, dataset_query, **kwargs)	×
422	if previous_messages is not None:	×
423
424	def add_previous_messages(example, index):	×
425	example["source"] = previous_messages[index] + example["source"]	×
426	return example	×
427
428	dataset = dataset.map(add_previous_messages, with_indices=True)	×
429	engine, _ = fetch_artifact(engine)	×
430	predictions = engine.select(dataset)	×
431	# predictions = post_process(raw_predictions, dataset)
432	if return_data:	×
433	return dataset.add_column("prediction", predictions)	×
434	return predictions	×

IBM / unitxt / 15883725135

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous