metric / operators.py
Elron's picture
Upload folder using huggingface_hub
cc5f321 verified
"""This section describes unitxt operators.
Operators: Building Blocks of Unitxt Processing Pipelines
==============================================================
Within the Unitxt framework, operators serve as the foundational elements used to assemble processing pipelines.
Each operator is designed to perform specific manipulations on dictionary structures within a stream.
These operators are callable entities that receive a MultiStream as input.
The output is a MultiStream, augmented with the operator's manipulations, which are then systematically applied to each instance in the stream when pulled.
Creating Custom Operators
-------------------------------
To enhance the functionality of Unitxt, users are encouraged to develop custom operators.
This can be achieved by inheriting from any of the existing operators listed below or from one of the fundamental :class:`base operators<unitxt.operator>`.
The primary task in any operator development is to implement the `process` function, which defines the unique manipulations the operator will perform.
General or Specialized Operators
--------------------------------
Some operators are specialized in specific data or specific operations such as:
- :class:`loaders<unitxt.loaders>` for accessing data from various sources.
- :class:`splitters<unitxt.splitters>` for fixing data splits.
- :class:`stream_operators<unitxt.stream_operators>` for changing joining and mixing streams.
- :class:`struct_data_operators<unitxt.struct_data_operators>` for structured data operators.
- :class:`collections_operators<unitxt.collections_operators>` for handling collections such as lists and dictionaries.
- :class:`dialog_operators<unitxt.dialog_operators>` for handling dialogs.
- :class:`string_operators<unitxt.string_operators>` for handling strings.
- :class:`span_labeling_operators<unitxt.span_labeling_operators>` for handling strings.
- :class:`fusion<unitxt.fusion>` for fusing and mixing datasets.
Other specialized operators are used by unitxt internally:
- :class:`templates<unitxt.templates>` for verbalizing data examples.
- :class:`formats<unitxt.formats>` for preparing data for models.
The rest of this section is dedicated to general operators.
General Operators List:
------------------------
"""
import operator
import uuid
import warnings
import zipfile
from abc import abstractmethod
from collections import Counter, defaultdict
from dataclasses import field
from itertools import zip_longest
from random import Random
from typing import (
Any,
Callable,
Dict,
Generator,
Iterable,
List,
Optional,
Tuple,
Union,
)
import requests
from .artifact import Artifact, fetch_artifact
from .dataclass import NonPositionalField, OptionalField
from .deprecation_utils import deprecation
from .dict_utils import dict_delete, dict_get, dict_set, is_subpath
from .operator import (
InstanceOperator,
MultiStream,
MultiStreamOperator,
PackageRequirementsMixin,
PagedStreamOperator,
SequentialOperator,
SideEffectOperator,
SingleStreamReducer,
SourceOperator,
StreamingOperator,
StreamInitializerOperator,
StreamOperator,
)
from .random_utils import new_random_generator
from .settings_utils import get_settings
from .stream import DynamicStream, ListStream, Stream
from .text_utils import nested_tuple_to_string
from .type_utils import isoftype
from .utils import (
deep_copy,
flatten_dict,
recursive_copy,
recursive_shallow_copy,
shallow_copy,
)
settings = get_settings()
class FromIterables(StreamInitializerOperator):
"""Creates a MultiStream from a dict of named iterables.
Example:
operator = FromIterables()
ms = operator.process(iterables)
"""
def process(self, iterables: Dict[str, Iterable]) -> MultiStream:
return MultiStream.from_iterables(iterables)
class IterableSource(SourceOperator):
"""Creates a MultiStream from a dict of named iterables.
It is a callable.
Args:
iterables (Dict[str, Iterable]): A dictionary mapping stream names to iterables.
Example:
operator = IterableSource(input_dict)
ms = operator()
"""
iterables: Dict[str, Iterable]
def process(self) -> MultiStream:
return MultiStream.from_iterables(self.iterables)
class MapInstanceValues(InstanceOperator):
"""A class used to map instance values into other values.
This class is a type of InstanceOperator,
it maps values of instances in a stream using predefined mappers.
Attributes:
mappers (Dict[str, Dict[str, Any]]): The mappers to use for mapping instance values.
Keys are the names of the fields to undergo mapping, and values are dictionaries
that define the mapping from old values to new values.
strict (bool): If True, the mapping is applied strictly. That means if a value
does not exist in the mapper, it will raise a KeyError. If False, values
that are not present in the mapper are kept as they are.
process_every_value (bool): If True, all fields to be mapped should be lists, and the mapping
is to be applied to their individual elements. If False, mapping is only applied to a field
containing a single value.
Examples:
MapInstanceValues(mappers={"a": {"1": "hi", "2": "bye"}})
replaces '1' with 'hi' and '2' with 'bye' in field 'a' in all instances of all streams:
instance {"a":"1", "b": 2} becomes {"a":"hi", "b": 2}.
MapInstanceValues(mappers={"a": {"1": "hi", "2": "bye"}}, process_every_value=True)
Assuming field 'a' is a list of values, potentially including "1"-s and "2"-s, this replaces
each such "1" with "hi" and "2" -- with "bye" in all instances of all streams:
instance {"a": ["1", "2"], "b": 2} becomes {"a": ["hi", "bye"], "b": 2}.
MapInstanceValues(mappers={"a": {"1": "hi", "2": "bye"}}, strict=True)
To ensure that all values of field 'a' are mapped in every instance, use strict=True.
Input instance {"a":"3", "b": 2} will raise an exception per the above call,
because "3" is not a key in the mapper of "a".
MapInstanceValues(mappers={"a": {str([1,2,3,4]): 'All', str([]): 'None'}}, strict=True)
replaces a list [1,2,3,4] with the string 'All' and an empty list by string 'None'.
Note that mapped values are defined by their string representation, so mapped values
must be converted to strings.
"""
mappers: Dict[str, Dict[str, str]]
strict: bool = True
process_every_value: bool = False
def verify(self):
# make sure the mappers are valid
for key, mapper in self.mappers.items():
assert isinstance(
mapper, dict
), f"Mapper for given field {key} should be a dict, got {type(mapper)}"
for k in mapper.keys():
assert isinstance(
k, str
), f'Key "{k}" in mapper for field "{key}" should be a string, got {type(k)}'
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
for key, mapper in self.mappers.items():
value = dict_get(instance, key)
if value is not None:
if (self.process_every_value is True) and (not isinstance(value, list)):
raise ValueError(
f"'process_every_field' == True is allowed only when all fields which have mappers, i.e., {list(self.mappers.keys())} are lists. Instance = {instance}"
)
if isinstance(value, list) and self.process_every_value:
for i, val in enumerate(value):
value[i] = self.get_mapped_value(instance, key, mapper, val)
else:
value = self.get_mapped_value(instance, key, mapper, value)
dict_set(
instance,
key,
value,
)
return instance
def get_mapped_value(self, instance, key, mapper, val):
val_as_str = str(val) # make sure the value is a string
if val_as_str in mapper:
return recursive_copy(mapper[val_as_str])
if self.strict:
raise KeyError(
f"value '{val}' in instance '{instance}' is not found in mapper '{mapper}', associated with field '{key}'."
)
return val
class FlattenInstances(InstanceOperator):
"""Flattens each instance in a stream, making nested dictionary entries into top-level entries.
Args:
parent_key (str): A prefix to use for the flattened keys. Defaults to an empty string.
sep (str): The separator to use when concatenating nested keys. Defaults to "_".
"""
parent_key: str = ""
sep: str = "_"
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
return flatten_dict(instance, parent_key=self.parent_key, sep=self.sep)
class Set(InstanceOperator):
"""Adds specified fields to each instance in a given stream or all streams (default) If fields exist, updates them.
Args:
fields (Dict[str, object]): The fields to add to each instance.
Use '/' to access inner fields
use_deepcopy (bool) : Deep copy the input value to avoid later modifications
Examples:
# Add a 'classes' field with a value of a list "positive" and "negative" to all streams
Set(fields={"classes": ["positive","negatives"]})
# Add a 'start' field under the 'span' field with a value of 0 to all streams
Set(fields={"span/start": 0}
# Add a 'classes' field with a value of a list "positive" and "negative" to 'train' stream
Set(fields={"classes": ["positive","negatives"], apply_to_stream=["train"]})
# Add a 'classes' field on a given list, prevent modification of original list
# from changing the instance.
Set(fields={"classes": alist}), use_deepcopy=True)
# if now alist is modified, still the instances remain intact.
"""
fields: Dict[str, object]
use_query: Optional[bool] = None
use_deepcopy: bool = False
def verify(self):
super().verify()
if self.use_query is not None:
depr_message = "Field 'use_query' is deprecated. From now on, default behavior is compatible to use_query=True. Please remove this field from your code."
warnings.warn(depr_message, DeprecationWarning, stacklevel=2)
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
for key, value in self.fields.items():
if self.use_deepcopy:
value = deep_copy(value)
dict_set(instance, key, value)
return instance
@deprecation(version="2.0.0", alternative=Set)
class AddFields(Set):
pass
class RemoveFields(InstanceOperator):
"""Remove specified fields from each instance in a stream.
Args:
fields (List[str]): The fields to remove from each instance.
"""
fields: List[str]
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
for field_name in self.fields:
del instance[field_name]
return instance
class SelectFields(InstanceOperator):
"""Keep only specified fields from each instance in a stream.
Args:
fields (List[str]): The fields to keep from each instance.
"""
fields: List[str]
def prepare(self):
super().prepare()
self.fields.extend(["data_classification_policy", "recipe_metadata"])
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
new_instance = {}
for selected_field in self.fields:
new_instance[selected_field] = instance[selected_field]
return new_instance
class DefaultPlaceHolder:
pass
default_place_holder = DefaultPlaceHolder()
class InstanceFieldOperator(InstanceOperator):
"""A general stream instance operator that processes the values of a field (or multiple ones).
Args:
field (Optional[str]): The field to process, if only a single one is passed. Defaults to None
to_field (Optional[str]): Field name to save result into, if only one field is processed, if None is passed the
operation would happen in-place and its result would replace the value of "field". Defaults to None
field_to_field (Optional[Union[List[List[str]], Dict[str, str]]]): Mapping from names of fields to process,
to names of fields to save the results into. Inner List, if used, should be of length 2.
A field is processed by feeding its value into method 'process_value' and storing the result in to_field that
is mapped to the field.
When the type of argument 'field_to_field' is List, the order by which the fields are processed is their order
in the (outer) List. But when the type of argument 'field_to_field' is Dict, there is no uniquely determined
order. The end result might depend on that order if either (1) two different fields are mapped to the same
to_field, or (2) a field shows both as a key and as a value in different mappings.
The operator throws an AssertionError in either of these cases.
field_to_field defaults to None
process_every_value (bool): Processes the values in a list instead of the list as a value, similar to *var. Defaults to False
Note: if 'field' and 'to_field' (or both members of a pair in 'field_to_field') are equal (or share a common
prefix if 'field' and 'to_field' contain a /), then the result of the operation is saved within 'field'
"""
field: Optional[str] = None
to_field: Optional[str] = None
field_to_field: Optional[Union[List[List[str]], Dict[str, str]]] = None
use_query: Optional[bool] = None
process_every_value: bool = False
get_default: Any = None
not_exist_ok: bool = False
not_exist_do_nothing: bool = False
def verify(self):
super().verify()
if self.use_query is not None:
depr_message = "Field 'use_query' is deprecated. From now on, default behavior is compatible to use_query=True. Please remove this field from your code."
warnings.warn(depr_message, DeprecationWarning, stacklevel=2)
def verify_field_definition(self):
if hasattr(self, "_field_to_field") and self._field_to_field is not None:
return
assert (
(self.field is None) != (self.field_to_field is None)
), "Must uniquely define the field to work on, through exactly one of either 'field' or 'field_to_field'"
assert (
self.to_field is None or self.field_to_field is None
), f"Can not apply operator to create both {self.to_field} and the to fields in the mapping {self.field_to_field}"
if self.field_to_field is None:
self._field_to_field = [
(self.field, self.to_field if self.to_field is not None else self.field)
]
else:
self._field_to_field = (
list(self.field_to_field.items())
if isinstance(self.field_to_field, dict)
else self.field_to_field
)
assert (
self.field is not None or self.field_to_field is not None
), "Must supply a field to work on"
assert (
self.to_field is None or self.field_to_field is None
), f"Can not apply operator to create both on {self.to_field} and on the mapping from fields to fields {self.field_to_field}"
assert (
self.field is None or self.field_to_field is None
), f"Can not apply operator both on {self.field} and on the from fields in the mapping {self.field_to_field}"
assert (
self._field_to_field is not None
), f"the from and to fields must be defined or implied from the other inputs got: {self._field_to_field}"
assert (
len(self._field_to_field) > 0
), f"'input argument 'field_to_field' should convey at least one field to process. Got {self.field_to_field}"
# self._field_to_field is built explicitly by pairs, or copied from argument 'field_to_field'
if self.field_to_field is None:
return
# for backward compatibility also allow list of tuples of two strings
if isoftype(self.field_to_field, List[List[str]]) or isoftype(
self.field_to_field, List[Tuple[str, str]]
):
for pair in self._field_to_field:
assert (
len(pair) == 2
), f"when 'field_to_field' is defined as a list of lists, the inner lists should all be of length 2. {self.field_to_field}"
# order of field processing is uniquely determined by the input field_to_field when a list
return
if isoftype(self.field_to_field, Dict[str, str]):
if len(self.field_to_field) < 2:
return
for ff, tt in self.field_to_field.items():
for f, t in self.field_to_field.items():
if f == ff:
continue
assert (
t != ff
), f"In input argument 'field_to_field': {self.field_to_field}, field {f} is mapped to field {t}, while the latter is mapped to {tt}. Whether {f} or {t} is processed first might impact end result."
assert (
tt != t
), f"In input argument 'field_to_field': {self.field_to_field}, two different fields: {ff} and {f} are mapped to field {tt}. Whether {ff} or {f} is processed last might impact end result."
return
raise ValueError(
"Input argument 'field_to_field': {self.field_to_field} is neither of type List{List[str]] nor of type Dict[str, str]."
)
@abstractmethod
def process_instance_value(self, value: Any, instance: Dict[str, Any]):
pass
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
self.verify_field_definition()
for from_field, to_field in self._field_to_field:
try:
old_value = dict_get(
instance,
from_field,
default=default_place_holder,
not_exist_ok=self.not_exist_ok or self.not_exist_do_nothing,
)
if old_value is default_place_holder:
if self.not_exist_do_nothing:
return instance
old_value = self.get_default
except Exception as e:
raise ValueError(
f"Failed to get '{from_field}' from {instance} due to : {e}"
) from e
try:
if self.process_every_value:
new_value = [
self.process_instance_value(value, instance)
for value in old_value
]
else:
new_value = self.process_instance_value(old_value, instance)
except Exception as e:
raise ValueError(
f"Failed to process '{from_field}' from {instance} due to : {e}"
) from e
dict_set(
instance,
to_field,
new_value,
not_exist_ok=True,
)
return instance
class FieldOperator(InstanceFieldOperator):
def process_instance_value(self, value: Any, instance: Dict[str, Any]):
return self.process_value(value)
@abstractmethod
def process_value(self, value: Any) -> Any:
pass
class MapValues(FieldOperator):
mapping: Dict[str, str]
def process_value(self, value: Any) -> Any:
return self.mapping[str(value)]
class Rename(FieldOperator):
"""Renames fields.
Move value from one field to another, potentially, if field name contains a /, from one branch into another.
Remove the from field, potentially part of it in case of / in from_field.
Examples:
Rename(field_to_field={"b": "c"})
will change inputs [{"a": 1, "b": 2}, {"a": 2, "b": 3}] to [{"a": 1, "c": 2}, {"a": 2, "c": 3}]
Rename(field_to_field={"b": "c/d"})
will change inputs [{"a": 1, "b": 2}, {"a": 2, "b": 3}] to [{"a": 1, "c": {"d": 2}}, {"a": 2, "c": {"d": 3}}]
Rename(field_to_field={"b": "b/d"})
will change inputs [{"a": 1, "b": 2}, {"a": 2, "b": 3}] to [{"a": 1, "b": {"d": 2}}, {"a": 2, "b": {"d": 3}}]
Rename(field_to_field={"b/c/e": "b/d"})
will change inputs [{"a": 1, "b": {"c": {"e": 2, "f": 20}}}] to [{"a": 1, "b": {"c": {"f": 20}, "d": 2}}]
"""
def process_value(self, value: Any) -> Any:
return value
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
res = super().process(instance=instance, stream_name=stream_name)
for from_field, to_field in self._field_to_field:
if (not is_subpath(from_field, to_field)) and (
not is_subpath(to_field, from_field)
):
dict_delete(res, from_field, remove_empty_ancestors=True)
return res
@deprecation(version="2.0.0", alternative=Rename)
class RenameFields(Rename):
pass
class AddConstant(FieldOperator):
"""Adds a constant, being argument 'add', to the processed value.
Args:
add: the constant to add.
"""
add: Any
def process_value(self, value: Any) -> Any:
return self.add + value
class ShuffleFieldValues(FieldOperator):
"""Shuffles a list of values found in a field."""
def process_value(self, value: Any) -> Any:
res = list(value)
random_generator = new_random_generator(sub_seed=res)
random_generator.shuffle(res)
return res
class JoinStr(FieldOperator):
"""Joins a list of strings (contents of a field), similar to str.join().
Args:
separator (str): text to put between values
"""
separator: str = ","
def process_value(self, value: Any) -> Any:
return self.separator.join(str(x) for x in value)
class Apply(InstanceOperator):
"""A class used to apply a python function and store the result in a field.
Args:
function (str): name of function.
to_field (str): the field to store the result
any additional arguments are field names whose values will be passed directly to the function specified
Examples:
Store in field "b" the uppercase string of the value in field "a"
Apply("a", function=str.upper, to_field="b")
Dump the json representation of field "t" and store back in the same field.
Apply("t", function=json.dumps, to_field="t")
Set the time in a field 'b'.
Apply(function=time.time, to_field="b")
"""
__allow_unexpected_arguments__ = True
function: Callable = NonPositionalField(required=True)
to_field: str = NonPositionalField(required=True)
def function_to_str(self, function: Callable) -> str:
parts = []
if hasattr(function, "__module__"):
parts.append(function.__module__)
if hasattr(function, "__qualname__"):
parts.append(function.__qualname__)
else:
parts.append(function.__name__)
return ".".join(parts)
def str_to_function(self, function_str: str) -> Callable:
parts = function_str.split(".", 1)
if len(parts) == 1:
return __builtins__[parts[0]]
module_name, function_name = parts
if module_name in __builtins__:
obj = __builtins__[module_name]
elif module_name in globals():
obj = globals()[module_name]
else:
obj = __import__(module_name)
for part in function_name.split("."):
obj = getattr(obj, part)
return obj
def prepare(self):
super().prepare()
if isinstance(self.function, str):
self.function = self.str_to_function(self.function)
self._init_dict["function"] = self.function_to_str(self.function)
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
argv = [instance[arg] for arg in self._argv]
kwargs = {key: instance[val] for key, val in self._kwargs}
result = self.function(*argv, **kwargs)
instance[self.to_field] = result
return instance
class ListFieldValues(InstanceOperator):
"""Concatenates values of multiple fields into a list, and assigns it to a new field."""
fields: List[str]
to_field: str
use_query: Optional[bool] = None
def verify(self):
super().verify()
if self.use_query is not None:
depr_message = "Field 'use_query' is deprecated. From now on, default behavior is compatible to use_query=True. Please remove this field from your code."
warnings.warn(depr_message, DeprecationWarning, stacklevel=2)
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
values = []
for field_name in self.fields:
values.append(dict_get(instance, field_name))
dict_set(instance, self.to_field, values)
return instance
class ZipFieldValues(InstanceOperator):
"""Zips values of multiple fields in a given instance, similar to list(zip(*fields)).
The value in each of the specified 'fields' is assumed to be a list. The lists from all 'fields'
are zipped, and stored into 'to_field'.
If 'longest'=False, the length of the zipped result is determined by the shortest input value.
If 'longest'=False, the length of the zipped result is determined by the longest input, padding shorter
inputs with None -s.
"""
fields: List[str]
to_field: str
longest: bool = False
use_query: Optional[bool] = None
def verify(self):
super().verify()
if self.use_query is not None:
depr_message = "Field 'use_query' is deprecated. From now on, default behavior is compatible to use_query=True. Please remove this field from your code."
warnings.warn(depr_message, DeprecationWarning, stacklevel=2)
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
values = []
for field_name in self.fields:
values.append(dict_get(instance, field_name))
if self.longest:
zipped = zip_longest(*values)
else:
zipped = zip(*values)
dict_set(instance, self.to_field, list(zipped))
return instance
class InterleaveListsToDialogOperator(InstanceOperator):
"""Interleaves two lists, one of user dialog turns and one of assistant dialog turns, into a single list of tuples, alternating between "user" and "assistant".
The list of tuples if of format (role, turn_content), where the role label is specified by
the 'user_role_label' and 'assistant_role_label' fields (default to "user" and "assistant").
The user turns and assistant turns field are specified in the arguments.
The value of each of the 'fields' is assumed to be a list.
"""
user_turns_field: str
assistant_turns_field: str
user_role_label: str = "user"
assistant_role_label: str = "assistant"
to_field: str
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
user_turns = instance[self.user_turns_field]
assistant_turns = instance[self.assistant_turns_field]
assert (
len(user_turns) == len(assistant_turns)
or (len(user_turns) - len(assistant_turns) == 1)
), "user_turns must have either the same length as assistant_turns or one more turn."
interleaved_dialog = []
i, j = 0, 0 # Indices for the user and assistant lists
# While either list has elements left, continue interleaving
while i < len(user_turns) or j < len(assistant_turns):
if i < len(user_turns):
interleaved_dialog.append((self.user_role_label, user_turns[i]))
i += 1
if j < len(assistant_turns):
interleaved_dialog.append(
(self.assistant_role_label, assistant_turns[j])
)
j += 1
instance[self.to_field] = interleaved_dialog
return instance
class IndexOf(InstanceOperator):
"""For a given instance, finds the offset of value of field 'index_of', within the value of field 'search_in'."""
search_in: str
index_of: str
to_field: str
use_query: Optional[bool] = None
def verify(self):
super().verify()
if self.use_query is not None:
depr_message = "Field 'use_query' is deprecated. From now on, default behavior is compatible to use_query=True. Please remove this field from your code."
warnings.warn(depr_message, DeprecationWarning, stacklevel=2)
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
lst = dict_get(instance, self.search_in)
item = dict_get(instance, self.index_of)
instance[self.to_field] = lst.index(item)
return instance
class TakeByField(InstanceOperator):
"""From field 'field' of a given instance, select the member indexed by field 'index', and store to field 'to_field'."""
field: str
index: str
to_field: str = None
use_query: Optional[bool] = None
def verify(self):
super().verify()
if self.use_query is not None:
depr_message = "Field 'use_query' is deprecated. From now on, default behavior is compatible to use_query=True. Please remove this field from your code."
warnings.warn(depr_message, DeprecationWarning, stacklevel=2)
def prepare(self):
if self.to_field is None:
self.to_field = self.field
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
value = dict_get(instance, self.field)
index_value = dict_get(instance, self.index)
instance[self.to_field] = value[index_value]
return instance
class Perturb(FieldOperator):
"""Slightly perturbs the contents of 'field'. Could be Handy for imitating prediction from given target.
When task was classification, argument 'select_from' can be used to list the other potential classes, as a
relevant perturbation
"""
select_from: List[Any] = []
percentage_to_perturb: int = 1 # 1 percent
def verify(self):
assert (
0 <= self.percentage_to_perturb and self.percentage_to_perturb <= 100
), f"'percentage_to_perturb' should be in the range 0..100. Received {self.percentage_to_perturb}"
def prepare(self):
super().prepare()
self.random_generator = new_random_generator(sub_seed="CopyWithPerturbation")
def process_value(self, value: Any) -> Any:
perturb = self.random_generator.randint(1, 100) <= self.percentage_to_perturb
if not perturb:
return value
if value in self.select_from:
# 80% of cases, return a decent class, otherwise, perturb the value itself as follows
if self.random_generator.random() < 0.8:
return self.random_generator.choice(self.select_from)
if isinstance(value, float):
return value * (0.5 + self.random_generator.random())
if isinstance(value, int):
perturb = 1 if self.random_generator.random() < 0.5 else -1
return value + perturb
if isinstance(value, str):
if len(value) < 2:
# give up perturbation
return value
# throw one char out
prefix_len = self.random_generator.randint(1, len(value) - 1)
return value[:prefix_len] + value[prefix_len + 1 :]
# and in any other case:
return value
class Copy(FieldOperator):
"""Copies values from specified fields to specified fields.
Args (of parent class):
field_to_field (Union[List[List], Dict[str, str]]): A list of lists, where each sublist contains the source field and the destination field, or a dictionary mapping source fields to destination fields.
Examples:
An input instance {"a": 2, "b": 3}, when processed by
Copy(field_to_field={"a": "b"}
would yield {"a": 2, "b": 2}, and when processed by
Copy(field_to_field={"a": "c"} would yield
{"a": 2, "b": 3, "c": 2}
with field names containing / , we can also copy inside the field:
Copy(field="a/0",to_field="a")
would process instance {"a": [1, 3]} into {"a": 1}
"""
def process_value(self, value: Any) -> Any:
return value
class RecursiveCopy(FieldOperator):
def process_value(self, value: Any) -> Any:
return recursive_copy(value)
@deprecation(version="2.0.0", alternative=Copy)
class CopyFields(Copy):
pass
class GetItemByIndex(FieldOperator):
"""Get from the item list by the index in the field."""
items_list: List[Any]
def process_value(self, value: Any) -> Any:
return self.items_list[value]
class AddID(InstanceOperator):
"""Stores a unique id value in the designated 'id_field_name' field of the given instance."""
id_field_name: str = "id"
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
instance[self.id_field_name] = str(uuid.uuid4()).replace("-", "")
return instance
class Cast(FieldOperator):
"""Casts specified fields to specified types.
Args:
default (object): A dictionary mapping field names to default values for cases of casting failure.
process_every_value (bool): If true, all fields involved must contain lists, and each value in the list is then casted. Defaults to False.
"""
to: str
failure_default: Optional[Any] = "__UNDEFINED__"
def prepare(self):
self.types = {"int": int, "float": float, "str": str, "bool": bool}
def process_value(self, value):
try:
return self.types[self.to](value)
except ValueError as e:
if self.failure_default == "__UNDEFINED__":
raise ValueError(
f'Failed to cast value {value} to type "{self.to}", and no default value is provided.'
) from e
return self.failure_default
class CastFields(InstanceOperator):
"""Casts specified fields to specified types.
Args:
use_nested_query (bool): Whether to cast nested fields, expressed in dpath. Defaults to False.
fields (Dict[str, str]): A dictionary mapping field names to the names of the types to cast the fields to.
e.g: "int", "str", "float", "bool". Basic names of types
defaults (Dict[str, object]): A dictionary mapping field names to default values for cases of casting failure.
process_every_value (bool): If true, all fields involved must contain lists, and each value in the list is then casted. Defaults to False.
Examples:
CastFields(
fields={"a/d": "float", "b": "int"},
failure_defaults={"a/d": 0.0, "b": 0},
process_every_value=True,
use_nested_query=True
)
would process the input instance: {"a": {"d": ["half", "0.6", 1, 12]}, "b": ["2"]}
into {"a": {"d": [0.0, 0.6, 1.0, 12.0]}, "b": [2]}
"""
fields: Dict[str, str] = field(default_factory=dict)
failure_defaults: Dict[str, object] = field(default_factory=dict)
use_nested_query: bool = False
process_every_value: bool = False
def prepare(self):
self.types = {"int": int, "float": float, "str": str, "bool": bool}
def _cast_single(self, value, type, field):
try:
return self.types[type](value)
except Exception as e:
if field not in self.failure_defaults:
raise ValueError(
f'Failed to cast field "{field}" with value {value} to type "{type}", and no default value is provided.'
) from e
return self.failure_defaults[field]
def _cast_multiple(self, values, type, field):
return [self._cast_single(value, type, field) for value in values]
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
for field_name, type in self.fields.items():
value = dict_get(instance, field_name)
if self.process_every_value:
assert isinstance(
value, list
), f"'process_every_value' can be set to True only for fields that contain lists, whereas in instance {instance}, the contents of field '{field_name}' is of type '{type(value)}'"
casted_value = self._cast_multiple(value, type, field_name)
else:
casted_value = self._cast_single(value, type, field_name)
dict_set(instance, field_name, casted_value)
return instance
class DivideAllFieldsBy(InstanceOperator):
"""Recursively reach down to all fields that are float, and divide each by 'divisor'.
The given instance is viewed as a tree whose internal nodes are dictionaries and lists, and
the leaves are either 'float' and then divided, or other basic type, in which case, a ValueError is raised
if input flag 'strict' is True, or -- left alone, if 'strict' is False.
Args:
divisor (float) the value to divide by
strict (bool) whether to raise an error upon visiting a leaf that is not float. Defaults to False.
Example:
when instance {"a": 10.0, "b": [2.0, 4.0, 7.0], "c": 5} is processed by operator:
operator = DivideAllFieldsBy(divisor=2.0)
the output is: {"a": 5.0, "b": [1.0, 2.0, 3.5], "c": 5}
If the operator were defined with strict=True, through:
operator = DivideAllFieldsBy(divisor=2.0, strict=True),
the processing of the above instance would raise a ValueError, for the integer at "c".
"""
divisor: float = 1.0
strict: bool = False
def _recursive_divide(self, instance, divisor):
if isinstance(instance, dict):
for key, value in instance.items():
instance[key] = self._recursive_divide(value, divisor)
elif isinstance(instance, list):
for i, value in enumerate(instance):
instance[i] = self._recursive_divide(value, divisor)
elif isinstance(instance, float):
instance /= divisor
elif self.strict:
raise ValueError(f"Cannot divide instance of type {type(instance)}")
return instance
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
return self._recursive_divide(instance, self.divisor)
class ArtifactFetcherMixin:
"""Provides a way to fetch and cache artifacts in the system.
Args:
cache (Dict[str, Artifact]): A cache for storing fetched artifacts.
"""
cache: Dict[str, Artifact] = {}
@classmethod
def get_artifact(cls, artifact_identifier: str) -> Artifact:
if artifact_identifier not in cls.cache:
artifact, artifactory = fetch_artifact(artifact_identifier)
cls.cache[artifact_identifier] = artifact
return shallow_copy(cls.cache[artifact_identifier])
class ApplyOperatorsField(InstanceOperator):
"""Applies value operators to each instance in a stream based on specified fields.
Args:
operators_field (str): name of the field that contains a single name, or a list of names, of the operators to be applied,
one after the other, for the processing of the instance. Each operator is equipped with 'process_instance()'
method.
default_operators (List[str]): A list of default operators to be used if no operators are found in the instance.
Example:
when instance {"prediction": 111, "references": [222, 333] , "c": ["processors.to_string", "processors.first_character"]}
is processed by operator (please look up the catalog that these operators, they are tuned to process fields "prediction" and
"references"):
operator = ApplyOperatorsField(operators_field="c"),
the resulting instance is: {"prediction": "1", "references": ["2", "3"], "c": ["processors.to_string", "processors.first_character"]}
"""
operators_field: str
default_operators: List[str] = None
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
operator_names = instance.get(self.operators_field)
if operator_names is None:
assert (
self.default_operators is not None
), f"No operators found in field '{self.operators_field}', and no default operators provided."
operator_names = self.default_operators
if isinstance(operator_names, str):
operator_names = [operator_names]
# otherwise , operator_names is already a list
# we now have a list of nanes of operators, each is equipped with process_instance method.
operator = SequentialOperator(steps=operator_names)
return operator.process_instance(instance, stream_name=stream_name)
class FilterByCondition(StreamOperator):
"""Filters a stream, yielding only instances in which the values in required fields follow the required condition operator.
Raises an error if a required field name is missing from the input instance.
Args:
values (Dict[str, Any]): Field names and respective Values that instances must match according the condition, to be included in the output.
condition: the name of the desired condition operator between the specified (sub) field's value and the provided constant value. Supported conditions are ("gt", "ge", "lt", "le", "ne", "eq", "in","not in")
error_on_filtered_all (bool, optional): If True, raises an error if all instances are filtered out. Defaults to True.
Examples:
FilterByCondition(values = {"a":4}, condition = "gt") will yield only instances where field "a" contains a value > 4
FilterByCondition(values = {"a":4}, condition = "le") will yield only instances where "a"<=4
FilterByCondition(values = {"a":[4,8]}, condition = "in") will yield only instances where "a" is 4 or 8
FilterByCondition(values = {"a":[4,8]}, condition = "not in") will yield only instances where "a" different from 4 or 8
FilterByCondition(values = {"a/b":[4,8]}, condition = "not in") will yield only instances where "a" is
a dict in which key "b" is mapped to a value that is neither 4 nor 8
FilterByCondition(values = {"a[2]":4}, condition = "le") will yield only instances where "a" is a list whose 3-rd
element is <= 4
"""
values: Dict[str, Any]
condition: str
condition_to_func = {
"gt": operator.gt,
"ge": operator.ge,
"lt": operator.lt,
"le": operator.le,
"eq": operator.eq,
"ne": operator.ne,
"in": None, # Handled as special case
"not in": None, # Handled as special case
}
error_on_filtered_all: bool = True
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
yielded = False
for instance in stream:
if self._is_required(instance):
yielded = True
yield instance
if not yielded and self.error_on_filtered_all:
raise RuntimeError(
f"{self.__class__.__name__} filtered out every instance in stream '{stream_name}'. If this is intended set error_on_filtered_all=False"
)
def verify(self):
if self.condition not in self.condition_to_func:
raise ValueError(
f"Unsupported condition operator '{self.condition}', supported {list(self.condition_to_func.keys())}"
)
for key, value in self.values.items():
if self.condition in ["in", "not it"] and not isinstance(value, list):
raise ValueError(
f"The filter for key ('{key}') in FilterByCondition with condition '{self.condition}' must be list but is not : '{value}'"
)
return super().verify()
def _is_required(self, instance: dict) -> bool:
for key, value in self.values.items():
try:
instance_key = dict_get(instance, key)
except ValueError as ve:
raise ValueError(
f"Required filter field ('{key}') in FilterByCondition is not found in {instance}"
) from ve
if self.condition == "in":
if instance_key not in value:
return False
elif self.condition == "not in":
if instance_key in value:
return False
else:
func = self.condition_to_func[self.condition]
if func is None:
raise ValueError(
f"Function not defined for condition '{self.condition}'"
)
if not func(instance_key, value):
return False
return True
class FilterByConditionBasedOnFields(FilterByCondition):
"""Filters a stream based on a condition between 2 fields values.
Raises an error if either of the required fields names is missing from the input instance.
Args:
values (Dict[str, str]): The fields names that the filter operation is based on.
condition: the name of the desired condition operator between the specified field's values. Supported conditions are ("gt", "ge", "lt", "le", "ne", "eq", "in","not in")
error_on_filtered_all (bool, optional): If True, raises an error if all instances are filtered out. Defaults to True.
Examples:
FilterByCondition(values = {"a":"b}, condition = "gt") will yield only instances where field "a" contains a value greater then the value in field "b".
FilterByCondition(values = {"a":"b}, condition = "le") will yield only instances where "a"<="b"
"""
def _is_required(self, instance: dict) -> bool:
for key, value in self.values.items():
try:
instance_key = dict_get(instance, key)
except ValueError as ve:
raise ValueError(
f"Required filter field ('{key}') in FilterByCondition is not found in {instance}"
) from ve
try:
instance_value = dict_get(instance, value)
except ValueError as ve:
raise ValueError(
f"Required filter field ('{value}') in FilterByCondition is not found in {instance}"
) from ve
if self.condition == "in":
if instance_key not in instance_value:
return False
elif self.condition == "not in":
if instance_key in instance_value:
return False
else:
func = self.condition_to_func[self.condition]
if func is None:
raise ValueError(
f"Function not defined for condition '{self.condition}'"
)
if not func(instance_key, instance_value):
return False
return True
class ComputeExpressionMixin(Artifact):
"""Computes an expression expressed over fields of an instance.
Args:
expression (str): the expression, in terms of names of fields of an instance
imports_list (List[str]): list of names of imports needed for the evaluation of the expression
"""
expression: str
imports_list: List[str] = OptionalField(default_factory=list)
def verify(self):
PackageRequirementsMixin.check_missing_requirements(self, self.imports_list)
def prepare(self):
# can not do the imports here, because object does not pickle with imports
self.globals = {
module_name: __import__(module_name) for module_name in self.imports_list
}
def compute_expression(self, instance: dict) -> Any:
if settings.allow_unverified_code:
return eval(self.expression, {**self.globals, **instance})
raise ValueError(
f"Cannot evaluate expression in {self} when unitxt.settings.allow_unverified_code=False - either set it to True or set {settings.allow_unverified_code_key} environment variable."
"\nNote: If using test_card() with the default setting, increase loader_limit to avoid missing conditions due to limited data sampling."
)
class FilterByExpression(StreamOperator, ComputeExpressionMixin):
"""Filters a stream, yielding only instances which fulfil a condition specified as a string to be python's eval-uated.
Raises an error if a field participating in the specified condition is missing from the instance
Args:
expression (str): a condition over fields of the instance, to be processed by python's eval()
imports_list (List[str]): names of imports needed for the eval of the query (e.g. 're', 'json')
error_on_filtered_all (bool, optional): If True, raises an error if all instances are filtered out. Defaults to True.
Examples:
FilterByExpression(expression = "a > 4") will yield only instances where "a">4
FilterByExpression(expression = "a <= 4 and b > 5") will yield only instances where the value of field "a" is not exceeding 4 and in field "b" -- greater than 5
FilterByExpression(expression = "a in [4, 8]") will yield only instances where "a" is 4 or 8
FilterByExpression(expression = "a not in [4, 8]") will yield only instances where "a" is neither 4 nor 8
FilterByExpression(expression = "a['b'] not in [4, 8]") will yield only instances where "a" is a dict in which key 'b' is mapped to a value that is neither 4 nor 8
"""
error_on_filtered_all: bool = True
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
yielded = False
for instance in stream:
if self.compute_expression(instance):
yielded = True
yield instance
if not yielded and self.error_on_filtered_all:
raise RuntimeError(
f"{self.__class__.__name__} filtered out every instance in stream '{stream_name}'. If this is intended set error_on_filtered_all=False"
)
class ExecuteExpression(InstanceOperator, ComputeExpressionMixin):
"""Compute an expression, specified as a string to be eval-uated, over the instance's fields, and store the result in field to_field.
Raises an error if a field mentioned in the query is missing from the instance.
Args:
expression (str): an expression to be evaluated over the fields of the instance
to_field (str): the field where the result is to be stored into
imports_list (List[str]): names of imports needed for the eval of the query (e.g. 're', 'json')
Examples:
When instance {"a": 2, "b": 3} is process-ed by operator
ExecuteExpression(expression="a+b", to_field = "c")
the result is {"a": 2, "b": 3, "c": 5}
When instance {"a": "hello", "b": "world"} is process-ed by operator
ExecuteExpression(expression = "a+' '+b", to_field = "c")
the result is {"a": "hello", "b": "world", "c": "hello world"}
"""
to_field: str
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
instance[self.to_field] = self.compute_expression(instance)
return instance
class ExtractMostCommonFieldValues(MultiStreamOperator):
field: str
stream_name: str
overall_top_frequency_percent: Optional[int] = 100
min_frequency_percent: Optional[int] = 0
to_field: str
process_every_value: Optional[bool] = False
"""
Extract the unique values of a field ('field') of a given stream ('stream_name') and store (the most frequent of) them
as a list in a new field ('to_field') in all streams.
More specifically, sort all the unique values encountered in field 'field' by decreasing order of frequency.
When 'overall_top_frequency_percent' is smaller than 100, trim the list from bottom, so that the total frequency of
the remaining values makes 'overall_top_frequency_percent' of the total number of instances in the stream.
When 'min_frequency_percent' is larger than 0, remove from the list any value whose relative frequency makes
less than 'min_frequency_percent' of the total number of instances in the stream.
At most one of 'overall_top_frequency_percent' and 'min_frequency_percent' is allowed to move from their default values.
Examples:
ExtractMostCommonFieldValues(stream_name="train", field="label", to_field="classes") - extracts all the unique values of
field 'label', sorts them by decreasing frequency, and stores the resulting list in field 'classes' of each and
every instance in all streams.
ExtractMostCommonFieldValues(stream_name="train", field="labels", to_field="classes", process_every_value=True) -
in case that field 'labels' contains a list of values (and not a single value) - track the occurrences of all the possible
value members in these lists, and report the most frequent values.
if process_every_value=False, track the most frequent whole lists, and report those (as a list of lists) in field
'to_field' of each instance of all streams.
ExtractMostCommonFieldValues(stream_name="train", field="label", to_field="classes",overall_top_frequency_percent=80) -
extracts the most frequent possible values of field 'label' that together cover at least 80% of the instances of stream_name,
and stores them in field 'classes' of each instance of all streams.
ExtractMostCommonFieldValues(stream_name="train", field="label", to_field="classes",min_frequency_percent=5) -
extracts all possible values of field 'label' that cover, each, at least 5% of the instances.
Stores these values, sorted by decreasing order of frequency, in field 'classes' of each instance in all streams.
"""
def verify(self):
assert (
self.overall_top_frequency_percent <= 100
and self.overall_top_frequency_percent >= 0
), "'overall_top_frequency_percent' must be between 0 and 100"
assert (
self.min_frequency_percent <= 100 and self.min_frequency_percent >= 0
), "'min_frequency_percent' must be between 0 and 100"
assert not (
self.overall_top_frequency_percent < 100 and self.min_frequency_percent > 0
), "At most one of 'overall_top_frequency_percent' and 'min_frequency_percent' is allowed to move from their default value"
super().verify()
def process(self, multi_stream: MultiStream) -> MultiStream:
stream = multi_stream[self.stream_name]
counter = Counter()
for instance in stream:
if (not isinstance(instance[self.field], list)) and (
self.process_every_value is True
):
raise ValueError(
"'process_every_field' is allowed to change to 'True' only for fields whose contents are lists"
)
if (not isinstance(instance[self.field], list)) or (
self.process_every_value is False
):
# either not a list, or is a list but process_every_value == False : view contetns of 'field' as one entity whose occurrences are counted.
counter.update(
[(*instance[self.field],)]
if isinstance(instance[self.field], list)
else [instance[self.field]]
) # convert to a tuple if list, to enable the use of Counter which would not accept
# a list as an hashable entity to count its occurrences
else:
# content of 'field' is a list and process_every_value == True: add one occurrence on behalf of each individual value
counter.update(instance[self.field])
# here counter counts occurrences of individual values, or tuples.
values_and_counts = counter.most_common()
if self.overall_top_frequency_percent < 100:
top_frequency = (
sum(counter.values()) * self.overall_top_frequency_percent / 100.0
)
sum_counts = 0
for _i, p in enumerate(values_and_counts):
sum_counts += p[1]
if sum_counts >= top_frequency:
break
values_and_counts = counter.most_common(_i + 1)
if self.min_frequency_percent > 0:
min_frequency = self.min_frequency_percent * sum(counter.values()) / 100.0
while values_and_counts[-1][1] < min_frequency:
values_and_counts.pop()
values_to_keep = [
[*ele[0]] if isinstance(ele[0], tuple) else ele[0]
for ele in values_and_counts
]
addmostcommons = Set(fields={self.to_field: values_to_keep})
return addmostcommons(multi_stream)
class ExtractFieldValues(ExtractMostCommonFieldValues):
def verify(self):
super().verify()
def prepare(self):
self.overall_top_frequency_percent = 100
self.min_frequency_percent = 0
class Intersect(FieldOperator):
"""Intersects the value of a field, which must be a list, with a given list.
Args:
allowed_values (list) - list to intersect.
"""
allowed_values: List[Any]
def verify(self):
super().verify()
if self.process_every_value:
raise ValueError(
"'process_every_value=True' is not supported in Intersect operator"
)
if not isinstance(self.allowed_values, list):
raise ValueError(
f"The allowed_values is not a list but '{self.allowed_values}'"
)
def process_value(self, value: Any) -> Any:
super().process_value(value)
if not isinstance(value, list):
raise ValueError(f"The value in field is not a list but '{value}'")
return [e for e in value if e in self.allowed_values]
class RemoveValues(FieldOperator):
"""Removes elements in a field, which must be a list, using a given list of unallowed.
Args:
unallowed_values (list) - values to be removed.
"""
unallowed_values: List[Any]
def verify(self):
super().verify()
if not isinstance(self.unallowed_values, list):
raise ValueError(
f"The unallowed_values is not a list but '{self.unallowed_values}'"
)
def process_value(self, value: Any) -> Any:
if not isinstance(value, list):
raise ValueError(f"The value in field is not a list but '{value}'")
return [e for e in value if e not in self.unallowed_values]
class Unique(SingleStreamReducer):
"""Reduces a stream to unique instances based on specified fields.
Args:
fields (List[str]): The fields that should be unique in each instance.
"""
fields: List[str] = field(default_factory=list)
@staticmethod
def to_tuple(instance: dict, fields: List[str]) -> tuple:
result = []
for field_name in fields:
value = instance[field_name]
if isinstance(value, list):
value = tuple(value)
result.append(value)
return tuple(result)
def process(self, stream: Stream) -> Stream:
seen = set()
for instance in stream:
values = self.to_tuple(instance, self.fields)
if values not in seen:
seen.add(values)
return list(seen)
class SplitByValue(MultiStreamOperator):
"""Splits a MultiStream into multiple streams based on unique values in specified fields.
Args:
fields (List[str]): The fields to use when splitting the MultiStream.
"""
fields: List[str] = field(default_factory=list)
def process(self, multi_stream: MultiStream) -> MultiStream:
uniques = Unique(fields=self.fields)(multi_stream)
result = {}
for stream_name, stream in multi_stream.items():
stream_unique_values = uniques[stream_name]
for unique_values in stream_unique_values:
filtering_values = dict(zip(self.fields, unique_values))
filtered_streams = FilterByCondition(
values=filtering_values, condition="eq"
)._process_single_stream(stream)
filtered_stream_name = (
stream_name + "_" + nested_tuple_to_string(unique_values)
)
result[filtered_stream_name] = filtered_streams
return MultiStream(result)
class SplitByNestedGroup(MultiStreamOperator):
"""Splits a MultiStream that is small - for metrics, hence: whole stream can sit in memory, split by the value of field 'group'.
Args:
number_of_fusion_generations: int
the value in field group is of the form "sourcen/sourcenminus1/..." describing the sources in which the instance sat
when these were fused, potentially several phases of fusion. the name of the most recent source sits first in this value.
(See BaseFusion and its extensions)
number_of_fuaion_generations specifies the length of the prefix by which to split the stream.
E.g. for number_of_fusion_generations = 1, only the most recent fusion in creating this multi_stream, affects the splitting.
For number_of_fusion_generations = -1, take the whole history written in this field, ignoring number of generations.
"""
field_name_of_group: str = "group"
number_of_fusion_generations: int = 1
def process(self, multi_stream: MultiStream) -> MultiStream:
result = defaultdict(list)
for stream_name, stream in multi_stream.items():
for instance in stream:
if self.field_name_of_group not in instance:
raise ValueError(
f"Field {self.field_name_of_group} is missing from instance {instance}"
)
signature = (
stream_name
+ "~" # a sign that does not show within group values
+ (
"/".join(
instance[self.field_name_of_group].split("/")[
: self.number_of_fusion_generations
]
)
if self.number_of_fusion_generations >= 0
# for values with a smaller number of generations - take up to their last generation
else instance[self.field_name_of_group]
# for each instance - take all its generations
)
)
result[signature].append(instance)
return MultiStream.from_iterables(result)
class ApplyStreamOperatorsField(StreamOperator, ArtifactFetcherMixin):
"""Applies stream operators to a stream based on specified fields in each instance.
Args:
field (str): The field containing the operators to be applied.
reversed (bool): Whether to apply the operators in reverse order.
"""
field: str
reversed: bool = False
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
first_instance = stream.peek()
operators = first_instance.get(self.field, [])
if isinstance(operators, str):
operators = [operators]
if self.reversed:
operators = list(reversed(operators))
for operator_name in operators:
operator = self.get_artifact(operator_name)
assert isinstance(
operator, StreamingOperator
), f"Operator {operator_name} must be a StreamOperator"
stream = operator(MultiStream({stream_name: stream}))[stream_name]
yield from stream
class ApplyMetric(StreamOperator, ArtifactFetcherMixin):
"""Applies metric operators to a stream based on a metric field specified in each instance.
Args:
metric_field (str): The field containing the metrics to be applied.
calc_confidence_intervals (bool): Whether the applied metric should calculate confidence intervals or not.
"""
metric_field: str
calc_confidence_intervals: bool
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
from .metrics import Metric
# Number of instances in input stream is assumed to be small. This is why
# each metric consumes all of them and lays them in its main memory, and even generates
# some 1000 copies thereof for the sake of CI.
# So we start with deep copying here, to make a 'frozen' status of the stream, having
# passed the preprocess_steps of the task, and inference, and now getting to be evaluated,
# a frozen status to be fed into each of the metrics listed in metric_field,
# so that the evaluation of one does not affect the evaluation of another
# (typically, affecting via change of instance as part of
# preprocess_steps of MetricPipeline, as illustrated in docs/adding_metrics/Using Metric Pipelines).
instances_upon_entrance_to_metrics_evaluations = []
for instance in stream:
instances_upon_entrance_to_metrics_evaluations.append(
recursive_copy(instance)
)
first_instance = instances_upon_entrance_to_metrics_evaluations[0]
metric_names = first_instance.get(self.metric_field, [])
if not metric_names:
raise RuntimeError(
f"Missing metric names in field '{self.metric_field}' and instance '{first_instance}'."
)
if isinstance(metric_names, str):
metric_names = [metric_names]
# Each metric operator computes its score and then sets the main score, overwriting
# the previous main score value (if any). So, we need to reverse the order of the listed metrics.
# This will cause the first listed metric to run last, and the main score will be set
# by the first listed metric (as desired).
metric_names = list(reversed(metric_names))
for metric_name in metric_names:
metric = self.get_artifact(metric_name)
assert isinstance(
metric, Metric
), f"Operator {metric_name} must be a Metric"
if not self.calc_confidence_intervals:
metric.disable_confidence_interval_calculation()
multi_stream = MultiStream(
{
"tmp": ListStream(
instances_list=instances_upon_entrance_to_metrics_evaluations,
copying=True, # ensures deep copy when iterating over instances
)
}
)
multi_stream = metric(multi_stream)
for evaluated_instance, freezed_instance in zip(
multi_stream["tmp"], instances_upon_entrance_to_metrics_evaluations
):
freezed_instance["score"] = recursive_shallow_copy(
evaluated_instance["score"]
)
yield from instances_upon_entrance_to_metrics_evaluations
class MergeStreams(MultiStreamOperator):
"""Merges multiple streams into a single stream.
Args:
new_stream_name (str): The name of the new stream resulting from the merge.
add_origin_stream_name (bool): Whether to add the origin stream name to each instance.
origin_stream_name_field_name (str): The field name for the origin stream name.
"""
streams_to_merge: List[str] = None
new_stream_name: str = "all"
add_origin_stream_name: bool = True
origin_stream_name_field_name: str = "origin"
def merge(self, multi_stream) -> Generator:
for stream_name, stream in multi_stream.items():
if self.streams_to_merge is None or stream_name in self.streams_to_merge:
for instance in stream:
if self.add_origin_stream_name:
instance[self.origin_stream_name_field_name] = stream_name
yield instance
def process(self, multi_stream: MultiStream) -> MultiStream:
return MultiStream(
{
self.new_stream_name: DynamicStream(
self.merge, gen_kwargs={"multi_stream": multi_stream}
)
}
)
class Shuffle(PagedStreamOperator):
"""Shuffles the order of instances in each page of a stream.
Args (of superclass):
page_size (int): The size of each page in the stream. Defaults to 1000.
"""
random_generator: Random = None
def before_process_multi_stream(self):
super().before_process_multi_stream()
self.random_generator = new_random_generator(sub_seed="shuffle")
def process(self, page: List[Dict], stream_name: Optional[str] = None) -> Generator:
self.random_generator.shuffle(page)
yield from page
class FeatureGroupedShuffle(Shuffle):
"""Class for shuffling an input dataset by instance 'blocks', not on the individual instance level.
Example is if the dataset consists of questions with paraphrases of it, and each question falls into a topic.
All paraphrases have the same ID value as the original.
In this case, we may want to shuffle on grouping_features = ['question ID'],
to keep the paraphrases and original question together.
We may also want to group by both 'question ID' and 'topic', if the question IDs are repeated between topics.
In this case, grouping_features = ['question ID', 'topic']
Args:
grouping_features (list of strings): list of feature names to use to define the groups.
a group is defined by each unique observed combination of data values for features in grouping_features
shuffle_within_group (bool): whether to further shuffle the instances within each group block, keeping the block order
Args (of superclass):
page_size (int): The size of each page in the stream. Defaults to 1000.
Note: shuffle_by_grouping_features determines the unique groups (unique combinations of values of grouping_features)
separately by page (determined by page_size). If a block of instances in the same group are split
into separate pages (either by a page break falling in the group, or the dataset was not sorted by
grouping_features), these instances will be shuffled separately and thus the grouping may be
broken up by pages. If the user wants to ensure the shuffle does the grouping and shuffling
across all pages, set the page_size to be larger than the dataset size.
See outputs_2features_bigpage and outputs_2features_smallpage in test_grouped_shuffle.
"""
grouping_features: List[str] = None
shuffle_within_group: bool = False
def process(self, page: List[Dict], stream_name: Optional[str] = None) -> Generator:
if self.grouping_features is None:
super().process(page, stream_name)
else:
yield from self.shuffle_by_grouping_features(page)
def shuffle_by_grouping_features(self, page):
import itertools
from collections import defaultdict
groups_to_instances = defaultdict(list)
for item in page:
groups_to_instances[
tuple(item[ff] for ff in self.grouping_features)
].append(item)
# now extract the groups (i.e., lists of dicts with order preserved)
page_blocks = list(groups_to_instances.values())
# and now shuffle the blocks
self.random_generator.shuffle(page_blocks)
if self.shuffle_within_group:
blocks = []
# reshuffle the instances within each block, but keep the blocks in order
for block in page_blocks:
self.random_generator.shuffle(block)
blocks.append(block)
page_blocks = blocks
# now flatten the list so it consists of individual dicts, but in (randomized) block order
return list(itertools.chain(*page_blocks))
class EncodeLabels(InstanceOperator):
"""Encode each value encountered in any field in 'fields' into the integers 0,1,...
Encoding is determined by a str->int map that is built on the go, as different values are
first encountered in the stream, either as list members or as values in single-value fields.
Args:
fields (List[str]): The fields to encode together.
Example: applying
EncodeLabels(fields = ["a", "b/*"])
on input stream = [{"a": "red", "b": ["red", "blue"], "c":"bread"},
{"a": "blue", "b": ["green"], "c":"water"}] will yield the
output stream = [{'a': 0, 'b': [0, 1], 'c': 'bread'}, {'a': 1, 'b': [2], 'c': 'water'}]
Note: qpath is applied here, and hence, fields that are lists, should be included in
input 'fields' with the appendix "/*" as in the above example.
"""
fields: List[str]
def _process_multi_stream(self, multi_stream: MultiStream) -> MultiStream:
self.encoder = {}
return super()._process_multi_stream(multi_stream)
def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
for field_name in self.fields:
values = dict_get(instance, field_name)
values_was_a_list = isinstance(values, list)
if not isinstance(values, list):
values = [values]
for value in values:
if value not in self.encoder:
self.encoder[value] = len(self.encoder)
new_values = [self.encoder[value] for value in values]
if not values_was_a_list:
new_values = new_values[0]
dict_set(
instance,
field_name,
new_values,
not_exist_ok=False, # the values to encode where just taken from there
set_multiple="*" in field_name
and isinstance(new_values, list)
and len(new_values) > 0,
)
return instance
class StreamRefiner(StreamOperator):
"""Discard from the input stream all instances beyond the leading 'max_instances' instances.
Thereby, if the input stream consists of no more than 'max_instances' instances, the resulting stream is the whole of the
input stream. And if the input stream consists of more than 'max_instances' instances, the resulting stream only consists
of the leading 'max_instances' of the input stream.
Args: max_instances (int)
apply_to_streams (optional, list(str)): names of streams to refine.
Examples:
when input = [{"a": 1},{"a": 2},{"a": 3},{"a": 4},{"a": 5},{"a": 6}] is fed into
StreamRefiner(max_instances=4)
the resulting stream is [{"a": 1},{"a": 2},{"a": 3},{"a": 4}]
"""
max_instances: int = None
apply_to_streams: Optional[List[str]] = None
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
if self.max_instances is not None:
yield from stream.take(self.max_instances)
else:
yield from stream
class DeterministicBalancer(StreamRefiner):
"""A class used to balance streams deterministically.
For each instance, a signature is constructed from the values of the instance in specified input 'fields'.
By discarding instances from the input stream, DeterministicBalancer maintains equal number of instances for all signatures.
When also input 'max_instances' is specified, DeterministicBalancer maintains a total instance count not exceeding
'max_instances'. The total number of discarded instances is as few as possible.
Attributes:
fields (List[str]): A list of field names to be used in producing the instance's signature.
max_instances (Optional, int)
Usage:
balancer = DeterministicBalancer(fields=["field1", "field2"], max_instances=200)
balanced_stream = balancer.process(stream)
Example:
When input [{"a": 1, "b": 1},{"a": 1, "b": 2},{"a": 2},{"a": 3},{"a": 4}] is fed into
DeterministicBalancer(fields=["a"])
the resulting stream will be: [{"a": 1, "b": 1},{"a": 2},{"a": 3},{"a": 4}]
"""
fields: List[str]
def signature(self, instance):
return str(tuple(dict_get(instance, field) for field in self.fields))
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
counter = Counter()
for instance in stream:
counter[self.signature(instance)] += 1
if len(counter) == 0:
return
lowest_count = counter.most_common()[-1][-1]
max_total_instances_per_sign = lowest_count
if self.max_instances is not None:
max_total_instances_per_sign = min(
lowest_count, self.max_instances // len(counter)
)
counter = Counter()
for instance in stream:
sign = self.signature(instance)
if counter[sign] < max_total_instances_per_sign:
counter[sign] += 1
yield instance
class MinimumOneExamplePerLabelRefiner(StreamRefiner):
"""A class used to return a specified number instances ensuring at least one example per label.
For each instance, a signature value is constructed from the values of the instance in specified input 'fields'.
MinimumOneExamplePerLabelRefiner takes first instance that appears from each label (each unique signature), and then adds more elements up to the max_instances limit. In general, the refiner takes the first elements in the stream that meet the required conditions.
MinimumOneExamplePerLabelRefiner then shuffles the results to avoid having one instance
from each class first and then the rest . If max instance is not set, the original stream will be used
Attributes:
fields (List[str]): A list of field names to be used in producing the instance's signature.
max_instances (Optional, int): Number of elements to select. Note that max_instances of StreamRefiners that are passed to the recipe (e.g. 'train_refiner'. `test_refiner`) are overridden by the recipe parameters ( `max_train_instances`, `max_test_instances`)
Usage:
balancer = MinimumOneExamplePerLabelRefiner(fields=["field1", "field2"], max_instances=200)
balanced_stream = balancer.process(stream)
Example:
When input [{"a": 1, "b": 1},{"a": 1, "b": 2},{"a": 1, "b": 3},{"a": 1, "b": 4},{"a": 2, "b": 5}] is fed into
MinimumOneExamplePerLabelRefiner(fields=["a"], max_instances=3)
the resulting stream will be:
[{'a': 1, 'b': 1}, {'a': 1, 'b': 2}, {'a': 2, 'b': 5}] (order may be different)
"""
fields: List[str]
def signature(self, instance):
return str(tuple(dict_get(instance, field) for field in self.fields))
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
if self.max_instances is None:
for instance in stream:
yield instance
counter = Counter()
for instance in stream:
counter[self.signature(instance)] += 1
all_keys = counter.keys()
if len(counter) == 0:
return
if self.max_instances is not None and len(all_keys) > self.max_instances:
raise Exception(
f"Can not generate a stream with at least one example per label, because the max instances requested {self.max_instances} is smaller than the number of different labels {len(all_keys)}"
f" ({len(all_keys)}"
)
counter = Counter()
used_indices = set()
selected_elements = []
# select at least one per class
for idx, instance in enumerate(stream):
sign = self.signature(instance)
if counter[sign] == 0:
counter[sign] += 1
used_indices.add(idx)
selected_elements.append(
instance
) # collect all elements first to allow shuffling of both groups
# select more to reach self.max_instances examples
for idx, instance in enumerate(stream):
if idx not in used_indices:
if self.max_instances is None or len(used_indices) < self.max_instances:
used_indices.add(idx)
selected_elements.append(
instance
) # collect all elements first to allow shuffling of both groups
# shuffle elements to avoid having one element from each class appear first
random_generator = new_random_generator(sub_seed=selected_elements)
random_generator.shuffle(selected_elements)
yield from selected_elements
class LengthBalancer(DeterministicBalancer):
"""Balances by a signature that reflects the total length of the fields' values, quantized into integer segments.
Args:
segments_boundaries (List[int]): distinct integers sorted in increasing order, that maps a given total length
into the index of the least of them that exceeds the total length. (If none exceeds -- into one index
beyond, namely, the length of segments_boundaries)
fields (Optional, List[str])
Example:
when input [{"a": [1, 3], "b": 0, "id": 0}, {"a": [1, 3], "b": 0, "id": 1}, {"a": [], "b": "a", "id": 2}] is fed into
.. code-block::
LengthBalancer(fields=["a"], segments_boundaries=[1])
input instances will be counted and balanced against two categories: empty total length (less than 1), and non-empty.
"""
segments_boundaries: List[int]
fields: Optional[List[str]]
def signature(self, instance):
total_len = 0
for field_name in self.fields:
total_len += len(dict_get(instance, field_name))
for i, val in enumerate(self.segments_boundaries):
if total_len < val:
return i
return i + 1
class DownloadError(Exception):
def __init__(
self,
message,
):
self.__super__(message)
class UnexpectedHttpCodeError(Exception):
def __init__(self, http_code):
self.__super__(f"unexpected http code {http_code}")
class DownloadOperator(SideEffectOperator):
"""Operator for downloading a file from a given URL to a specified local path.
Attributes:
source (str): URL of the file to be downloaded.
target (str): Local path where the downloaded file should be saved.
"""
source: str
target: str
def process(self):
try:
response = requests.get(self.source, allow_redirects=True)
except Exception as e:
raise DownloadError(f"Unabled to download {self.source}") from e
if response.status_code != 200:
raise UnexpectedHttpCodeError(response.status_code)
with open(self.target, "wb") as f:
f.write(response.content)
class ExtractZipFile(SideEffectOperator):
"""Operator for extracting files from a zip archive.
Attributes:
zip_file (str): Path of the zip file to be extracted.
target_dir (str): Directory where the contents of the zip file will be extracted.
"""
zip_file: str
target_dir: str
def process(self):
with zipfile.ZipFile(self.zip_file) as zf:
zf.extractall(self.target_dir)
class DuplicateInstances(StreamOperator):
"""Operator which duplicates each instance in stream a given number of times.
Attributes:
num_duplications (int): How many times each instance should be duplicated (1 means no duplication).
duplication_index_field (Optional[str]):
If given, then additional field with specified name is added to each duplicated instance,
which contains id of a given duplication. Defaults to None, so no field is added.
"""
num_duplications: int
duplication_index_field: Optional[str] = None
def process(self, stream: Stream, stream_name: Optional[str] = None) -> Generator:
for instance in stream:
for idx in range(self.num_duplications):
duplicate = recursive_shallow_copy(instance)
if self.duplication_index_field:
duplicate.update({self.duplication_index_field: idx})
yield duplicate
def verify(self):
if not isinstance(self.num_duplications, int) or self.num_duplications < 1:
raise ValueError(
f"num_duplications must be an integer equal to or greater than 1. "
f"Got: {self.num_duplications}."
)
if self.duplication_index_field is not None and not isinstance(
self.duplication_index_field, str
):
raise ValueError(
f"If given, duplication_index_field must be a string. "
f"Got: {self.duplication_index_field}"
)