#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# pyre-strict
from __future__ import annotations
import json
from abc import abstractmethod
from collections.abc import Iterable
from copy import deepcopy
from hashlib import md5
from io import StringIO
from typing import Any, TypeVar
import numpy as np
import pandas as pd
from ax.core.types import TFidelityTrialEvaluation, TTrialEvaluation
from ax.utils.common.base import Base
from ax.utils.common.serialization import (
extract_init_args,
SerializationMixin,
serialize_init_args,
TClassDecoderRegistry,
TDecoderRegistry,
)
from ax.utils.common.typeutils import checked_cast
from pyre_extensions import none_throws
TBaseData = TypeVar("TBaseData", bound="BaseData")
DF_REPR_MAX_LENGTH = 1000
[docs]
class BaseData(Base, SerializationMixin):
"""Class storing data for an experiment.
The dataframe is retrieved via the `df` property. The data can be stored
to an external store for future use by attaching it to an experiment using
`experiment.attach_data()` (this requires a description to be set.)
Attributes:
df: DataFrame with underlying data, and required columns. For BaseData, the
one required column is "arm_name".
description: Human-readable description of data.
"""
REQUIRED_COLUMNS = {"arm_name"}
COLUMN_DATA_TYPES: dict[str, Any] = {
# Ubiquitous columns.
"arm_name": str,
# Metric data-related columns.
"metric_name": str,
"mean": np.float64,
"sem": np.float64,
# Metadata columns available for all subclasses.
"trial_index": int,
"start_time": pd.Timestamp,
"end_time": pd.Timestamp,
"n": int,
# Metadata columns available for only some subclasses.
"frac_nonnull": np.float64,
"random_split": int,
"fidelities": str, # Dictionary stored as json
}
_df: pd.DataFrame
def __init__(
self: TBaseData,
df: pd.DataFrame | None = None,
description: str | None = None,
) -> None:
"""Init Data.
Args:
df: DataFrame with underlying data, and required columns.
description: Human-readable description of data.
"""
# Initialize with barebones DF.
if df is None:
self._df = pd.DataFrame(columns=list(self.required_columns()))
else:
columns = set(df.columns)
missing_columns = self.required_columns() - columns
if missing_columns:
raise ValueError(
f"Dataframe must contain required columns {list(missing_columns)}."
)
extra_columns = columns - self.supported_columns()
if extra_columns:
raise ValueError(f"Columns {list(extra_columns)} are not supported.")
df = df.dropna(axis=0, how="all").reset_index(drop=True)
df = self._safecast_df(df=df)
# Reorder the columns for easier viewing
col_order = [c for c in self.column_data_types() if c in df.columns]
self._df = df[col_order]
self.description = description
@classmethod
def _safecast_df(
cls: type[TBaseData],
df: pd.DataFrame,
# pyre-fixme[24]: Generic type `type` expects 1 type parameter, use
# `typing.Type` to avoid runtime subscripting errors.
extra_column_types: dict[str, type] | None = None,
) -> pd.DataFrame:
"""Function for safely casting df to standard data types.
Needed because numpy does not support NaNs in integer arrays.
Allows `Any` to be specified as a type, and will skip casting for that column.
Args:
df: DataFrame to safe-cast.
extra_column_types: types of columns only specified at instantiation-time.
Returns:
safe_df: DataFrame cast to standard dtypes.
"""
extra_column_types = extra_column_types or {}
dtype = {
# Pandas timestamp handlng is weird
col: "datetime64[ns]" if coltype is pd.Timestamp else coltype
for col, coltype in cls.column_data_types(
extra_column_types=extra_column_types
).items()
if col in df.columns.values
and not (
cls.column_data_types(extra_column_types)[col] is int
and df.loc[:, col].isnull().any()
)
and coltype is not Any
}
return checked_cast(pd.DataFrame, df.astype(dtype=dtype))
[docs]
@classmethod
def required_columns(cls) -> set[str]:
"""Names of columns that must be present in the underlying ``DataFrame``."""
return cls.REQUIRED_COLUMNS
[docs]
@classmethod
def supported_columns(
cls, extra_column_names: Iterable[str] | None = None
) -> set[str]:
"""Names of columns supported (but not necessarily required) by this class."""
extra_column_names = set(extra_column_names or [])
extra_column_types: dict[str, Any] = {name: Any for name in extra_column_names}
return cls.REQUIRED_COLUMNS.union(
cls.column_data_types(extra_column_types=extra_column_types)
)
[docs]
@classmethod
def column_data_types(
cls,
# pyre-fixme[24]: Generic type `type` expects 1 type parameter, use
# `typing.Type` to avoid runtime subscripting errors.
extra_column_types: dict[str, type] | None = None,
excluded_columns: Iterable[str] | None = None,
# pyre-fixme[24]: Generic type `type` expects 1 type parameter, use
# `typing.Type` to avoid runtime subscripting errors.
) -> dict[str, type]:
"""Type specification for all supported columns."""
extra_column_types = extra_column_types or {}
excluded_columns = excluded_columns or []
columns = {**cls.COLUMN_DATA_TYPES, **extra_column_types}
for column in excluded_columns:
if column in columns:
del columns[column]
return columns
[docs]
@classmethod
# pyre-fixme[2]: Parameter annotation cannot be `Any`.
def serialize_init_args(cls, obj: Any) -> dict[str, Any]:
"""Serialize the class-dependent properties needed to initialize this Data.
Used for storage and to help construct new similar Data.
"""
data = checked_cast(cls, obj)
return serialize_init_args(obj=data)
[docs]
@classmethod
def deserialize_init_args(
cls,
args: dict[str, Any],
decoder_registry: TDecoderRegistry | None = None,
class_decoder_registry: TClassDecoderRegistry | None = None,
) -> dict[str, Any]:
"""Given a dictionary, extract the properties needed to initialize the object.
Used for storage.
"""
# Extract `df` only if present, since certain inputs to this fn, e.g.
# SQAData.structure_metadata_json, don't have a `df` attribute.
if "df" in args and not isinstance(args["df"], pd.DataFrame):
# NOTE: Need dtype=False, otherwise infers arm_names like
# "4_1" should be int 41.
args["df"] = pd.read_json(StringIO(args["df"]["value"]), dtype=False)
return extract_init_args(args=args, class_=cls)
@property
def true_df(self) -> pd.DataFrame:
"""Return the ``DataFrame`` being used as the source of truth (avoid using
except for caching).
"""
return self._df
@property
def df(self) -> pd.DataFrame:
return self._df
@property
def df_hash(self) -> str:
"""Compute hash of pandas DataFrame.
This first serializes the DataFrame and computes the md5 hash on the
resulting string. Note that this may cause performance issue for very large
DataFrames.
Args:
df: The DataFrame for which to compute the hash.
Returns
str: The hash of the DataFrame.
"""
return md5(none_throws(self.df.to_json()).encode("utf-8")).hexdigest()
[docs]
def get_filtered_results(
self: TBaseData, **filters: dict[str, Any]
) -> pd.DataFrame:
"""Return filtered subset of data.
Args:
filter: Column names and values they must match.
Returns
df: The filtered DataFrame.
"""
df = self.df.copy()
if df.empty:
return df
columns = df.columns
for colname, value in filters.items():
if colname not in columns:
raise ValueError(
f"{colname} not in the set of columns: {columns}"
f"in this data object of type: {str(type(self))}."
)
df = df[df[colname] == value]
return df
[docs]
@classmethod
def from_multiple(
cls: type[TBaseData],
data: Iterable[TBaseData],
) -> TBaseData:
"""Combines multiple objects into one (with the concatenated
underlying dataframe).
Args:
data: Iterable of Ax objects of this class to combine.
"""
dfs = []
for datum in data:
if not isinstance(datum, cls):
raise TypeError(
f"All data objects must be instances of class {cls}. Got "
f"{type(datum)}."
)
dfs.append(datum.df)
if len(dfs) == 0:
return cls()
return cls(df=pd.concat(dfs, axis=0, sort=True))
[docs]
@classmethod
def from_evaluations(
cls: type[TBaseData],
evaluations: dict[str, TTrialEvaluation],
trial_index: int,
sample_sizes: dict[str, int] | None = None,
start_time: int | str | None = None,
end_time: int | str | None = None,
) -> TBaseData:
"""
Convert dict of evaluations to Ax data object.
Args:
evaluations: Map from arm name to outcomes, which itself is a mapping of
outcome names to values, means, or tuples of mean and SEM. If SEM is
not specified, it will be set to None and inferred from data.
trial_index: Trial index to which this data belongs.
sample_sizes: Number of samples collected for each arm.
start_time: Optional start time of run of the trial that produced this
data, in milliseconds or iso format. Milliseconds will be automatically
converted to iso format because iso format automatically works with the
pandas column type `Timestamp`.
end_time: Optional end time of run of the trial that produced this
data, in milliseconds or iso format. Milliseconds will be automatically
converted to iso format because iso format automatically works with the
pandas column type `Timestamp`.
Returns:
Ax object of the enclosing class.
"""
records = cls._get_records(evaluations=evaluations, trial_index=trial_index)
records = cls._add_cols_to_records(
records=records,
sample_sizes=sample_sizes,
start_time=start_time,
end_time=end_time,
)
return cls(df=pd.DataFrame(records))
@staticmethod
@abstractmethod
def _get_records(
evaluations: dict[str, TTrialEvaluation], trial_index: int
) -> list[dict[str, Any]]:
pass
[docs]
@classmethod
def from_fidelity_evaluations(
cls: type[TBaseData],
evaluations: dict[str, TFidelityTrialEvaluation],
trial_index: int,
sample_sizes: dict[str, int] | None = None,
start_time: int | None = None,
end_time: int | None = None,
) -> TBaseData:
"""
Convert dict of fidelity evaluations to Ax data object.
Args:
evaluations: Map from arm name to list of (fidelity, outcomes)
where outcomes is itself a mapping of outcome names to values, means,
or tuples of mean and SEM. If SEM is not specified, it will be set
to None and inferred from data.
trial_index: Trial index to which this data belongs.
sample_sizes: Number of samples collected for each arm.
start_time: Optional start time of run of the trial that produced this
data, in milliseconds.
end_time: Optional end time of run of the trial that produced this
data, in milliseconds.
Returns:
Ax object of type ``cls``.
"""
records = cls._get_fidelity_records(
evaluations=evaluations, trial_index=trial_index
)
records = cls._add_cols_to_records(
records=records,
sample_sizes=sample_sizes,
start_time=start_time,
end_time=end_time,
)
return cls(df=pd.DataFrame(records))
@staticmethod
@abstractmethod
def _get_fidelity_records(
evaluations: dict[str, TFidelityTrialEvaluation], trial_index: int
) -> list[dict[str, Any]]:
pass
@staticmethod
def _add_cols_to_records(
records: list[dict[str, Any]],
sample_sizes: dict[str, int] | None = None,
start_time: int | str | None = None,
end_time: int | str | None = None,
) -> list[dict[str, Any]]:
"""Adds to records metadata columns that are available for all
BaseData subclasses.
"""
if start_time is not None or end_time is not None:
if isinstance(start_time, int):
start_time = _ms_epoch_to_isoformat(start_time)
if isinstance(end_time, int):
end_time = _ms_epoch_to_isoformat(end_time)
for record in records:
record.update({"start_time": start_time, "end_time": end_time})
if sample_sizes:
for record in records:
record["n"] = sample_sizes[str(record["arm_name"])]
return records
[docs]
def copy_structure_with_df(self: TBaseData, df: pd.DataFrame) -> TBaseData:
"""Serialize the structural properties needed to initialize this class.
Used for storage and to help construct new similar objects. All kwargs
other than ``df`` and ``description`` are considered structural.
"""
cls = type(self)
return cls(df=df, **cls.serialize_init_args(self))
def __repr__(self) -> str:
"""String representation of the subclass, inheriting from this base."""
df_markdown = self.df.to_markdown()
if len(df_markdown) > DF_REPR_MAX_LENGTH:
df_markdown = df_markdown[:DF_REPR_MAX_LENGTH] + "..."
return f"{self.__class__.__name__}(df=\n{df_markdown})"
[docs]
class Data(BaseData):
"""Class storing numerical data for an experiment.
The dataframe is retrieved via the `df` property. The data can be stored
to an external store for future use by attaching it to an experiment using
`experiment.attach_data()` (this requires a description to be set.)
Attributes:
df: DataFrame with underlying data, and required columns. For BaseData, the
required columns are "arm_name", "metric_name", "mean", and "sem", the
latter two of which must be numeric.
description: Human-readable description of data.
"""
# Note: Although the SEM (standard error of the mean) is a required column in data,
# downstream models can infer missing SEMs. Simply specify NaN as the SEM value,
# either in your Metric class or in Data explicitly.
REQUIRED_COLUMNS: set[str] = BaseData.REQUIRED_COLUMNS.union(
{"metric_name", "mean", "sem"}
)
@staticmethod
def _get_records(
evaluations: dict[str, TTrialEvaluation], trial_index: int
) -> list[dict[str, Any]]:
return [
{
"arm_name": name,
"metric_name": metric_name,
"mean": value[0] if isinstance(value, tuple) else value,
"sem": value[1] if isinstance(value, tuple) else None,
"trial_index": trial_index,
}
for name, evaluation in evaluations.items()
for metric_name, value in evaluation.items()
]
@staticmethod
def _get_fidelity_records(
evaluations: dict[str, TFidelityTrialEvaluation], trial_index: int
) -> list[dict[str, Any]]:
return [
{
"arm_name": name,
"metric_name": metric_name,
"mean": value[0] if isinstance(value, tuple) else value,
"sem": value[1] if isinstance(value, tuple) else None,
"trial_index": trial_index,
"fidelities": json.dumps(fidelity),
}
for name, fidelity_and_metrics_list in evaluations.items()
for fidelity, evaluation in fidelity_and_metrics_list
for metric_name, value in evaluation.items()
]
@property
def metric_names(self) -> set[str]:
"""Set of metric names that appear in the underlying dataframe of
this object.
"""
return set() if self.df.empty else set(self.df["metric_name"].values)
[docs]
def filter(
self,
trial_indices: Iterable[int] | None = None,
metric_names: Iterable[str] | None = None,
) -> Data:
"""Construct a new object with the subset of rows corresponding to the
provided trial indices AND metric names. If either trial_indices or
metric_names are not provided, that dimension will not be filtered.
"""
return self.__class__(
df=self._filter_df(
df=self.df, trial_indices=trial_indices, metric_names=metric_names
)
)
@staticmethod
def _filter_df(
df: pd.DataFrame,
trial_indices: Iterable[int] | None = None,
metric_names: Iterable[str] | None = None,
) -> pd.DataFrame:
trial_indices_mask = (
df["trial_index"].isin(trial_indices)
if trial_indices is not None
else pd.Series([True] * len(df))
)
metric_names_mask = (
df["metric_name"].isin(metric_names)
if metric_names is not None
else pd.Series([True] * len(df))
)
return df.loc[trial_indices_mask & metric_names_mask]
[docs]
@classmethod
def from_multiple_data(
cls, data: Iterable[Data], subset_metrics: Iterable[str] | None = None
) -> Data:
"""Combines multiple objects into one (with the concatenated
underlying dataframe).
Args:
data: Iterable of Ax objects of this class to combine.
subset_metrics: If specified, combined object will only contain
metrics, names of which appear in this iterable,
in the underlying dataframe.
"""
data_out = cls.from_multiple(data=data)
if len(data_out.df.index) == 0:
return data_out
if subset_metrics:
data_out._df = data_out.df.loc[
data_out.df["metric_name"].isin(subset_metrics)
]
return data_out
[docs]
def clone(self) -> Data:
"""Returns a new Data object with the same underlying dataframe."""
return Data(df=deepcopy(self.df), description=self.description)
[docs]
def set_single_trial(data: Data) -> Data:
"""Returns a new Data object where we set all rows to have the same
trial index (i.e. 0). This is meant to be used with our IVW transform,
which will combine multiple observations of the same outcome.
"""
df = data._df.copy()
if "trial_index" in df:
df["trial_index"] = 0
return Data(df=df)
[docs]
def clone_without_metrics(data: Data, excluded_metric_names: Iterable[str]) -> Data:
"""Returns a new data object where rows containing the outcomes specified by
`metric_names` are filtered out. Used to sanitize data before using it as
training data for a model that requires data rectangularity.
Args:
data: Original data to clone.
excluded_metric_names: Metrics to avoid copying
Returns:
new version of the original data without specified metrics.
"""
return Data(
df=data.df[
data.df["metric_name"].apply(lambda n: n not in excluded_metric_names)
].copy()
)
def _ms_epoch_to_isoformat(epoch: int) -> str:
return pd.Timestamp(epoch, unit="ms").isoformat()
[docs]
def custom_data_class(
# pyre-fixme[24]: Generic type `type` expects 1 type parameter, use
# `typing.Type` to avoid runtime subscripting errors.
column_data_types: dict[str, type] | None = None,
required_columns: set[str] | None = None,
time_columns: set[str] | None = None,
) -> type[Data]:
"""Creates a custom data class with additional columns.
All columns and their designations on the base data class are preserved,
the inputs here are appended to the definitions on the base class.
Args:
column_data_types: Dict from column name to column type.
required_columns: Set of additional columns required for this data object.
time_columns: Set of additional columns to cast to timestamp.
Returns:
New data subclass with amended column definitions.
"""
class CustomData(Data):
@classmethod
def required_columns(cls) -> set[str]:
return (required_columns or set()).union(Data.REQUIRED_COLUMNS)
@classmethod
def column_data_types(
cls, extra_column_types: dict[str, type] | None = None
) -> dict[str, type]:
return super().column_data_types(
{**(extra_column_types or {}), **(column_data_types or {})}
)
return CustomData
