metric / metric_utils.py
Elron's picture
Upload folder using huggingface_hub
058c80a verified
raw
history blame
No virus
12.7 kB
import json
from copy import deepcopy
from typing import Any, Dict, Generator, Iterable, List, Optional
from datasets import Features, Value
from numpy import nanmean
from .dataclass import Dataclass
from .dict_utils import dict_set
from .operator import (
MultiStreamOperator,
SequentialOperator,
SequentialOperatorInitializer,
StreamInitializerOperator,
)
from .operators import (
ApplyMetric,
ApplyOperatorsField,
Copy,
FlattenInstances,
MergeStreams,
RenameFields,
SplitByNestedGroup,
)
from .register import _reset_env_local_catalogs, register_all_artifacts
from .schema import UNITXT_DATASET_SCHEMA
from .settings_utils import get_settings
from .stream import DynamicStream, MultiStream
from .struct_data_operators import LoadJson
class MultiStreamScoreMean(MultiStreamOperator):
"""Given a multi-stream where each stream is already scored globally, generate a nested global score for the whole multi-stream.
The whole-ms-global-score is a nested structure, specifying (also) the individual global scores of the
individual streams participating in the input multi_stream.
The instances of all these individual streams are assumed to have the "group" field indicate the stream
they belong to.
Potentially, these individual streams were produced from a SplitByNestedGroup
operator that did not use the full length of the value in field "group" of the instances, but only the
first g components thereof, indicated by argument 'number_of_fusion_generations' of operator SplitByNestedGroup.
At any rate, a distinguishing prefix of the "group" value is recorded, by operator SplitByNestedGroup, in the stream_name.
The nested structure of the whole-ms-global-score is induced by these distinguishing prefixes,
by virtue of the global score of each individual stream sitting in the nested whole-ms-global-score,
deep in that dictionary, at the leaf lead to by a path being the distinguishing prefix indicated in the stream_name.
Thus, the global score of the stream becomes a leaf (though a dict by itself) of the whole-ms-global-score.
The ancestor nodes of the above leaves, in the whole-ms-global-score, contain each (in addition to dicts
leading down to leaves) a field named "score" whose value is set to be the mean of the values
sitting in field "score" of its immediate children nodes, and a field named "score_name" whose
value is set to be "group_mean".
When the input multistream consists of one single stream, it is returned as is, mainly for backward compatibility.
"""
def update_intermediate_level_scores(self, level: dict) -> float:
if "score" in level:
return level["score"]
# the global score of the stream participating in this MultiStream
sub_scores = []
for key in level:
if isinstance(level[key], dict):
sub_scores.append(self.update_intermediate_level_scores(level[key]))
level.update({"score": nanmean(sub_scores), "score_name": "groups_mean"})
return level["score"]
def process(self, multi_stream: MultiStream) -> MultiStream:
# each stream went through Metric which is a single-stream-operator , and ended up with all
# its instance["score"]["global"] linking to the same single dict object.
# Here we first generate a new, nested version, for the whole-ms-global_score, and then update
# each stream's global score with the new version
# but if only one stream in the multistream - we return it as is
if len(multi_stream) == 1:
return multi_stream
global_score = {}
first_instances = {}
iterators = {}
for stream_name, stream in multi_stream.items():
iterators[stream_name] = iter(stream)
try:
first_instances[stream_name] = next(iterators[stream_name])
except StopIteration:
continue # an empty stream, goto next stream
instance = first_instances[stream_name]
dict_set(
dic=global_score,
query=stream_name.split("~")[-1],
value=deepcopy(instance["score"]["global"]),
not_exist_ok=True,
)
self.update_intermediate_level_scores(global_score)
# update the global_score object for each stream. Recall that all instances
# in each stream link all to same python dict object
for stream_name in multi_stream.keys():
instance = first_instances[stream_name]
instance["score"]["global"].clear()
instance["score"]["global"].update(global_score)
def never_peek_twice_generator(
stream_name: str, first_instances: dict, iterators: dict
) -> Generator:
while True:
if stream_name in first_instances:
yield first_instances.pop(stream_name)
try:
yield next(iterators[stream_name])
except StopIteration:
return
return MultiStream(
{
stream_name: DynamicStream(
never_peek_twice_generator,
gen_kwargs={
"stream_name": stream_name,
"first_instances": first_instances,
"iterators": iterators,
},
)
for stream_name in multi_stream.keys()
}
)
class FromPredictionsAndOriginalData(StreamInitializerOperator):
def zip(self, predictions, references):
for prediction, original in zip(predictions, references):
yield {**original, "prediction": prediction}
def process(
self, predictions: List[str], references: Iterable, split_name: str = "all"
) -> MultiStream:
return MultiStream(
{
split_name: DynamicStream(
self.zip,
gen_kwargs={"predictions": predictions, "references": references},
)
}
)
# The task_data field in the schema is defined as
# Sequence({"key": Value(dtype="string"), "value": Value("string")})
# When receiving instances from this scheme, the keys and values are returned as two separate
# lists, and are converted to a dictionary.
_post_process_steps = SequentialOperator(
steps=[
Copy(
field="prediction",
to_field="raw_prediction",
),
Copy(
field="references",
to_field="raw_references",
),
Copy(
field="source",
to_field="task_data/source",
),
ApplyOperatorsField(
operators_field="postprocessors",
),
Copy(
field="prediction",
to_field="processed_prediction",
),
Copy(
field="references",
to_field="processed_references",
),
]
)
class PostProcessRecipe(SequentialOperatorInitializer):
def prepare(self):
register_all_artifacts()
self.steps = [
FromPredictionsAndOriginalData(),
_post_process_steps,
]
def _post_process(
predictions: List[str],
references: Iterable,
split_name: str = "all",
):
_reset_env_local_catalogs()
register_all_artifacts()
recipe = PostProcessRecipe()
multi_stream = recipe(
predictions=predictions, references=references, split_name=split_name
)
return [instance["processed_prediction"] for instance in multi_stream[split_name]]
class MetricRecipe(SequentialOperatorInitializer):
calc_confidence_intervals: bool = True
number_of_fusion_generations: int = 2
def prepare(self):
register_all_artifacts()
self.steps = [
FromPredictionsAndOriginalData(),
LoadJson(field="task_data"),
_post_process_steps,
SplitByNestedGroup(
field_name_of_group="group",
number_of_fusion_generations=self.number_of_fusion_generations,
),
ApplyMetric(
"metrics",
calc_confidence_intervals=self.calc_confidence_intervals,
),
MultiStreamScoreMean(),
MergeStreams(),
RenameFields(
field="raw_prediction",
to_field="prediction",
),
RenameFields(
field="raw_references",
to_field="references",
),
Copy(
field="source",
to_field="task_data/source",
),
]
UNITXT_METRIC_SCHEMA = Features(
{"predictions": Value("string"), "references": dict(UNITXT_DATASET_SCHEMA)}
)
def _compute(
predictions: List[str],
references: Iterable,
flatten: bool = False,
split_name: str = "all",
calc_confidence_intervals: bool = True,
):
_reset_env_local_catalogs()
register_all_artifacts()
recipe = MetricRecipe(calc_confidence_intervals=calc_confidence_intervals)
multi_stream = recipe(
predictions=predictions, references=references, split_name=split_name
)
if flatten:
operator = FlattenInstances()
multi_stream = operator(multi_stream)
stream = multi_stream[split_name]
return list(stream)
"""
The API of a metric service:
- MetricRequest: A single input request to the metrics service.
- MetricResponse: A response returned from a metrics service.
"""
class InstanceInput(Dataclass):
"""A single instance inputted to a metric service."""
prediction: Any
references: List[Any]
additional_inputs: Optional[Dict] = None
class MetricRequest(Dataclass):
"""A request to a metrics service, includes a list of input instances."""
instance_inputs: List[InstanceInput]
class MetricResponse(Dataclass):
"""A response produced by a metrics service, includes the computed scores."""
# A list of instance score dictionaries. Each dictionary contains the
# score names and score values for a single instance.
instances_scores: List[Dict[str, Any]]
# The global scores dictionary, containing global score names and values.
# These are scores computed over the entire set of input instances, e.g.
# an average over a score computed per instance.
global_score: Dict[str, Any]
"""
Functionality for loading the remote metrics configuration from local environment variables.
"""
# A list of metrics to be executed remotely.
# For example: '["metrics.rag.context_relevance","metrics.rag.bert_k_precision"]'
# This value should be a valid json list
UNITXT_REMOTE_METRICS = "UNITXT_REMOTE_METRICS"
# The remote endpoint on which the remote metrics are available.
# For example, 'http://127.0.0.1:8000/compute'
UNITXT_REMOTE_METRICS_ENDPOINT = "UNITXT_REMOTE_METRICS_ENDPOINT"
def get_remote_metrics_names() -> List[str]:
"""Load the remote metrics names from an environment variable.
Returns:
List[str] - names of metrics to be executed remotely.
"""
settings = get_settings()
remote_metrics = settings.remote_metrics
if remote_metrics:
remote_metrics = json.loads(remote_metrics)
if not isinstance(remote_metrics, list):
raise RuntimeError(
f"Unexpected value {remote_metrics} for the '{UNITXT_REMOTE_METRICS}' environment variable. "
f"The value is expected to be a list of metric names in json format."
)
for remote_metric in remote_metrics:
if not isinstance(remote_metric, str):
raise RuntimeError(
f"Unexpected value {remote_metric} within the '{UNITXT_REMOTE_METRICS}' environment variable. "
f"The value is expected to be a string but its type is {type(remote_metric)}."
)
return remote_metrics
def get_remote_metrics_endpoint() -> str:
"""Load the remote metrics endpoint from an environment variable.
Returns:
str - The remote endpoint on which the remote metrics are available.
"""
settings = get_settings()
try:
remote_metrics_endpoint = settings.remote_metrics_endpoint
except AttributeError as e:
raise RuntimeError(
f"Unexpected None value for '{UNITXT_REMOTE_METRICS_ENDPOINT}'. "
f"Running remote metrics requires defining an "
f"endpoint in the environment variable '{UNITXT_REMOTE_METRICS_ENDPOINT}'."
) from e
return remote_metrics_endpoint