ax.utils¶

Base¶

class ax.utils.common.base.Base[source]¶

Bases: object

Metaclass for core Ax classes. Provides an equality check and db_id property for SQA storage.

property db_id: Optional[int]¶

class ax.utils.common.base.SortableBase[source]¶

Bases: Base

Extension to the base class that also provides an inequality check.

Constants¶

class ax.utils.common.constants.Keys(value)[source]¶

Bases: str, Enum

Enum of reserved keys in options dicts etc, alphabetized.

NOTE: Useful for keys in dicts that correspond to kwargs to classes or functions and/or are used in multiple places.

ACQF_KWARGS = 'acquisition_function_kwargs'¶

BATCH_INIT_CONDITIONS = 'batch_initial_conditions'¶

CANDIDATE_SET = 'candidate_set'¶

CANDIDATE_SIZE = 'candidate_size'¶

COST_AWARE_UTILITY = 'cost_aware_utility'¶

COST_INTERCEPT = 'cost_intercept'¶

CURRENT_VALUE = 'current_value'¶

EXPAND = 'expand'¶

EXPECTED_ACQF_VAL = 'expected_acquisition_value'¶

FIDELITY_FEATURES = 'fidelity_features'¶

FIDELITY_WEIGHTS = 'fidelity_weights'¶

FRAC_RANDOM = 'frac_random'¶

FULL_PARAMETERIZATION = 'full_parameterization'¶

IMMUTABLE_SEARCH_SPACE_AND_OPT_CONF = 'immutable_search_space_and_opt_config'¶

MAXIMIZE = 'maximize'¶

METADATA = 'metadata'¶

METRIC_NAMES = 'metric_names'¶

NUM_FANTASIES = 'num_fantasies'¶

NUM_INNER_RESTARTS = 'num_inner_restarts'¶

NUM_RESTARTS = 'num_restarts'¶

NUM_TRACE_OBSERVATIONS = 'num_trace_observations'¶

OBJECTIVE = 'objective'¶

OPTIMIZER_KWARGS = 'optimizer_kwargs'¶

PREFERENCE_DATA = 'preference_data'¶

PROJECT = 'project'¶

QMC = 'qmc'¶

RAW_INNER_SAMPLES = 'raw_inner_samples'¶

RAW_SAMPLES = 'raw_samples'¶

REFIT_ON_UPDATE = 'refit_on_update'¶

SAMPLER = 'sampler'¶

SEED_INNER = 'seed_inner'¶

SEQUENTIAL = 'sequential'¶

STATE_DICT = 'state_dict'¶

SUBCLASS = 'subclass'¶

SUBSET_MODEL = 'subset_model'¶

TASK_FEATURES = 'task_features'¶

TRIAL_COMPLETION_TIMESTAMP = 'trial_completion_timestamp'¶

WARM_START_REFITTING = 'warm_start_refitting'¶

X_BASELINE = 'X_baseline'¶

Decorator¶

class ax.utils.common.decorator.ClassDecorator[source]¶

Bases: ABC

Template for making a decorator work as a class level decorator. That decorator should extend ClassDecorator. It must implement __init__ and decorate_callable. See disable_logger.decorate_callable for an example. decorate_callable should call self._call_func() instead of directly calling func to handle static functions. Note: _call_func is still imperfect and unit tests should be used to ensure everything is working properly. There is a lot of complexity in detecting classmethods and staticmethods and removing the self argument in the right situations. For best results always use keyword args in the decorated class.

DECORATE_PRIVATE can be set to determine whether private methods should be decorated. In the case of a logging decorator, you may only want to decorate things the user calls. But in the case of a disable logging decorator, you may want to decorate everything to ensure no logs escape.

DECORATE_PRIVATE = True¶

abstract decorate_callable(func: Callable[[...], T]) → Callable[[...], T][source]¶

decorate_class(klass: T) → T[source]¶

Docutils¶

Support functions for sphinx et. al

ax.utils.common.docutils.copy_doc(src: Callable[[...], Any]) → Callable[[_T], _T][source]¶

A decorator that copies the docstring of another object

Since sphinx actually loads the python modules to grab the docstrings this works with both sphinx and the help function.

class Cat(Mamal):

  @property
  @copy_doc(Mamal.is_feline)
  def is_feline(self) -> true:
      ...

Equality¶

ax.utils.common.equality.dataframe_equals(df1: pandas.DataFrame, df2: pandas.DataFrame) → bool[source]¶

Compare equality of two pandas dataframes.

ax.utils.common.equality.datetime_equals(dt1: Optional[datetime], dt2: Optional[datetime]) → bool[source]¶

Compare equality of two datetimes, ignoring microseconds.

ax.utils.common.equality.equality_typechecker(eq_func: Callable) → Callable[source]¶

A decorator to wrap all __eq__ methods to ensure that the inputs are of the right type.

ax.utils.common.equality.object_attribute_dicts_equal(one_dict: Dict[str, Any], other_dict: Dict[str, Any]) → bool[source]¶

Utility to check if all items in attribute dicts of two Ax objects are the same.

NOTE: Special-cases some Ax object attributes, like “_experiment” or “_model”, where full equality is hard to check.

ax.utils.common.equality.object_attribute_dicts_find_unequal_fields(one_dict: Dict[str, Any], other_dict: Dict[str, Any], fast_return: bool = True, skip_db_id_check: bool = False) → Tuple[Dict[str, Tuple[Any, Any]], Dict[str, Tuple[Any, Any]]][source]¶

Utility for finding out what attributes of two objects’ attribute dicts are unequal.

Parameters:

• one_dict – First object’s attribute dict (obj.__dict__).

• other_dict – Second object’s attribute dict (obj.__dict__).

• fast_return – Boolean representing whether to return as soon as a single unequal attribute was found or to iterate over all attributes and collect all unequal ones.

Returns:

• attribute name to attribute values of unequal type (as a tuple),

• attribute name to attribute values of unequal value (as a tuple).

Return type:

Two dictionaries

ax.utils.common.equality.same_elements(list1: List[Any], list2: List[Any]) → bool[source]¶

Compare equality of two lists of core Ax objects.

Assumptions:

– The contents of each list are types that implement __eq__ – The lists do not contain duplicates

Checking equality is then the same as checking that the lists are the same length, and that one is a subset of the other.

Equality¶

ax.utils.common.executils.handle_exceptions_in_retries(no_retry_exceptions: Tuple[Type[Exception], ...], retry_exceptions: Tuple[Type[Exception], ...], suppress_errors: bool, check_message_contains: Optional[str], last_retry: bool, logger: Optional[Logger], wrap_error_message_in: Optional[str]) → Generator[None, None, None][source]¶

ax.utils.common.executils.retry_on_exception(exception_types: Optional[Tuple[Type[Exception], ...]] = None, no_retry_on_exception_types: Optional[Tuple[Type[Exception], ...]] = None, check_message_contains: Optional[List[str]] = None, retries: int = 3, suppress_all_errors: bool = False, logger: Optional[Logger] = None, default_return_on_suppression: Optional[Any] = None, wrap_error_message_in: Optional[str] = None, initial_wait_seconds: Optional[int] = None) → Optional[Any][source]¶

A decorator for instance methods or standalone functions that makes them retry on failure and allows to specify on which types of exceptions the function should and should not retry.

NOTE: If the argument suppress_all_errors is supplied and set to True, the error will be suppressed and default value returned.

Parameters:

• exception_types – A tuple of exception(s) types to catch in the decorated function. If none is provided, baseclass Exception will be used.

• no_retry_on_exception_types – Exception types to consider non-retryable even if their supertype appears in exception_types or the only exceptions to not retry on if no exception_types are specified.

• check_message_contains – A list of strings, against which to match error messages. If the error message contains any one of these strings, the exception will cause a retry. NOTE: This argument works in addition to exception_types; if those are specified, only the specified types of exceptions will be caught and retried on if they contain the strings provided as check_message_contains.

• retries – Number of retries to perform.

• suppress_all_errors – If true, after all the retries are exhausted, the error will still be suppressed and default_return_on_suppresion will be returned from the function. NOTE: If using this argument, the decorated function may not actually get fully executed, if it consistently raises an exception.

• logger – A handle for the logger to be used.

• default_return_on_suppression – If the error is suppressed after all the retries, then this default value will be returned from the function. Defaults to None.

• wrap_error_message_in – If raising the error message after all the retries, a string wrapper for the error message (useful for making error messages more user-friendly). NOTE: Format of resulting error will be: “<wrap_error_message_in>: <original_error_type>: <original_error_msg>”, with the stack trace of the original message.

• initial_wait_seconds – Initial length of time to wait between failures, doubled after each failure up to a maximum of 10 minutes. If unspecified then there is no wait between retries.

Kwargs¶

ax.utils.common.kwargs.consolidate_kwargs(kwargs_iterable: Iterable[Optional[Dict[str, Any]]], keywords: Iterable[str]) → Dict[str, Any][source]¶

Combine an iterable of kwargs into a single dict of kwargs, where kwargs by duplicate keys that appear later in the iterable get priority over the ones that appear earlier and only kwargs referenced in keywords will be used. This allows to combine somewhat redundant sets of kwargs, where a user-set kwarg, for instance, needs to override a default kwarg.

>>> consolidate_kwargs(
...     kwargs_iterable=[{'a': 1, 'b': 2}, {'b': 3, 'c': 4, 'd': 5}],
...     keywords=['a', 'b', 'd']
... )
{'a': 1, 'b': 3, 'd': 5}

ax.utils.common.kwargs.filter_kwargs(function: Callable, **kwargs: Any) → Any[source]¶

Filter out kwargs that are not applicable for a given function. Return a copy of given kwargs dict with only the required kwargs.

ax.utils.common.kwargs.get_function_argument_names(function: Callable, omit: Optional[List[str]] = None) → List[str][source]¶

Extract parameter names from function signature.

ax.utils.common.kwargs.get_function_default_arguments(function: Callable) → Dict[str, Any][source]¶

Extract default arguments from function signature.

ax.utils.common.kwargs.validate_kwarg_typing(typed_callables: List[Callable], **kwargs: Any) → None[source]¶

Check if keywords in kwargs exist in any of the typed_callables and if the type of each keyword value matches the type of corresponding arg in one of the callables

Note: this function expects the typed callables to have unique keywords for the arguments and will raise an error if repeat keywords are found.

ax.utils.common.kwargs.warn_on_kwargs(callable_with_kwargs: Callable, **kwargs: Any) → None[source]¶

Log a warning when a decoder function receives unexpected kwargs.

NOTE: This mainly caters to the use case where an older version of Ax is used to decode objects, serialized to JSON by a newer version of Ax (and therefore potentially containing new fields). In that case, the decoding function should not fail when encountering those additional fields, but rather just ignore them and log a warning using this function.

Logger¶

class ax.utils.common.logger.AxOutputNameFilter(name='')[source]¶

Bases: Filter

This is a filter which sets the record’s output_name, if not configured

filter(record: LogRecord) → bool[source]¶

Determine if the specified record is to be logged.

Returns True if the record should be logged, or False otherwise. If deemed appropriate, the record may be modified in-place.

ax.utils.common.logger.build_file_handler(filepath: str, level: int = 20) → StreamHandler[source]¶

Build a file handle that logs entries to the given file, using the same formatting as the stream handler.

Parameters:

• filepath – Location of the file to log output to. If the file exists, output will be appended. If it does not exist, a new file will be created.

• level – The log level. By default, sets level to INFO

Returns:

A logging.FileHandler instance

ax.utils.common.logger.build_stream_handler(level: int = 20) → StreamHandler[source]¶

Build the default stream handler used for most Ax logging. Sets default level to INFO, instead of WARNING.

Parameters:

level – The log level. By default, sets level to INFO

Returns:

A logging.StreamHandler instance

class ax.utils.common.logger.disable_logger(name: str, level: int = 40)[source]¶

Bases: ClassDecorator

decorate_callable(func: Callable[[...], T]) → Callable[[...], T][source]¶

class ax.utils.common.logger.disable_loggers(names: List[str], level: int = 40)[source]¶

Bases: ClassDecorator

decorate_callable(func: Callable[[...], T]) → Callable[[...], T][source]¶

ax.utils.common.logger.get_logger(name: str, level: int = 20) → Logger[source]¶

Get an Axlogger.

To set a human-readable “output_name” that appears in logger outputs, add {“output_name”: “[MY_OUTPUT_NAME]”} to the logger’s contextual information. By default, we use the logger’s name

NOTE: To change the log level on particular outputs (e.g. STDERR logs), set the proper log level on the relevant handler, instead of the logger e.g. logger.handers[0].setLevel(INFO)

Parameters:

name – The name of the logger.

Returns:

The logging.Logger object.

ax.utils.common.logger.make_indices_str(indices: Iterable[int]) → str[source]¶

Generate a string representation of an iterable of indices; if indices are contiguous, returns a string formatted like like ‘<min_idx> - <max_idx>’, otherwise a string formatted like ‘[idx_1, idx_2, …, idx_n’].

ax.utils.common.logger.set_stderr_log_level(level: int) → None[source]¶

Set the log level for stream handler, such that logs of given level are printed to STDERR by the root logger

Serialization¶

class ax.utils.common.serialization.SerializationMixin[source]¶

Bases: ABC

classmethod deserialize_init_args(args: Dict[str, Any]) → Dict[str, Any][source]¶

Given a dictionary, deserialize the properties needed to initialize the object. Used for storage.

classmethod serialize_init_args(obj: Any) → Dict[str, Any][source]¶

Serialize the properties needed to initialize the object. Used for storage.

ax.utils.common.serialization.callable_from_reference(path: str) → Callable[source]¶

Retrieves a callable by its path.

ax.utils.common.serialization.callable_to_reference(callable: Callable) → str[source]¶

Obtains path to the callable of form <module>.<name>.

ax.utils.common.serialization.extract_init_args(args: Dict[str, Any], class_: Type) → Dict[str, Any][source]¶

Given a dictionary, extract the arguments required for the given class’s constructor.

ax.utils.common.serialization.named_tuple_to_dict(data: Any) → Any[source]¶

Recursively convert NamedTuples to dictionaries.

ax.utils.common.serialization.serialize_init_args(object: Any, exclude_fields: Optional[List[str]] = None) → Dict[str, Any][source]¶

Given an object, return a dictionary of the arguments that are needed by its constructor.

Testutils¶

Support functions for tests

class ax.utils.common.testutils.TestCase(methodName: str = 'runTest')[source]¶

Bases: TestCase

The base Ax test case, contains various helper functions to write unittests.

assertEqual(first: Any, second: Any, msg: Optional[str] = None) → None[source]¶

Fail if the two objects are unequal as determined by the ‘==’ operator.

assertRaisesOn(exc: Type[Exception], line: Optional[str] = None, regex: Optional[str] = None) → AbstractContextManager[None][source]¶

Assert that an exception is raised on a specific line.

static silence_stderr() → Generator[None, None, None][source]¶

A context manager that silences stderr for part of a test.

If any exception passes through this context manager the stderr will be printed, otherwise it will be discarded.

tearDown() → None[source]¶

Hook method for deconstructing the test fixture after testing it.

ax.utils.common.testutils.disable_profiler(f: Callable[[...], T]) → Callable[[...], T][source]¶

Not all tests are compatible with yappi. This wrapper can be used on a test to stop profiling instead of disabling it for the entire test case

Timeutils¶

ax.utils.common.timeutils.current_timestamp_in_millis() → int[source]¶

Grab current timestamp in milliseconds as an int.

ax.utils.common.timeutils.timestamps_in_range(start: datetime, end: datetime, delta: timedelta) → Generator[datetime, None, None][source]¶

Generator of timestamps in range [start, end], at intervals delta.

ax.utils.common.timeutils.to_ds(ts: datetime) → str[source]¶

Convert a datetime to a DS string.

ax.utils.common.timeutils.to_ts(ds: str) → datetime[source]¶

Convert a DS string to a datetime.

Typeutils¶

ax.utils.common.typeutils.checked_cast(typ: Type[T], val: V) → T[source]¶

Cast a value to a type (with a runtime safety check).

Returns the value unchanged and checks its type at runtime. This signals to the typechecker that the value has the designated type.

Like typing.cast check_cast performs no runtime conversion on its argument, but, unlike typing.cast, checked_cast will throw an error if the value is not of the expected type. The type passed as an argument should be a python class.

Parameters:

• typ – the type to cast to

• val – the value that we are casting

Returns:

the val argument, unchanged

ax.utils.common.typeutils.checked_cast_complex(typ: Type[T], val: V, message: Optional[str] = None) → T[source]¶

Cast a value to a type (with a runtime safety check). Used for subscripted generics which isinstance cannot run against.

Returns the value unchanged and checks its type at runtime. This signals to the typechecker that the value has the designated type.

Like typing.cast check_cast performs no runtime conversion on its argument, but, unlike typing.cast, checked_cast will throw an error if the value is not of the expected type.

Parameters:

• typ – the type to cast to

• val – the value that we are casting

• message – message to print on error

Returns:

the val argument casted to typ

ax.utils.common.typeutils.checked_cast_dict(key_typ: Type[K], value_typ: Type[V], d: Dict[X, Y]) → Dict[K, V][source]¶

Calls checked_cast on all keys and values in the dictionary.

ax.utils.common.typeutils.checked_cast_list(typ: Type[T], old_l: List[V]) → List[T][source]¶

Calls checked_cast on all items in a list.

ax.utils.common.typeutils.checked_cast_optional(typ: Type[T], val: Optional[V]) → Optional[T][source]¶

Calls checked_cast only if value is not None.

ax.utils.common.typeutils.checked_cast_to_tuple(typ: Tuple[Type[V], ...], val: V) → T[source]¶

Cast a value to a union of multiple types (with a runtime safety check). This function is similar to checked_cast, but allows for the type to be defined as a tuple of types, in which case the value is cast as a union of the types in the tuple.

Parameters:

• typ – the tuple of types to cast to

• val – the value that we are casting

Returns:

the val argument, unchanged

ax.utils.common.typeutils.not_none(val: Optional[T], message: Optional[str] = None) → T[source]¶

Unbox an optional type.

Parameters:

• val – the value to cast to a non None type

• message – optional override of the default error message

Returns:

val when val is not None

Return type:

V

Throws:

ValueError if val is None

ax.utils.common.typeutils.numpy_type_to_python_type(value: Any) → Any[source]¶

If value is a Numpy int or float, coerce to a Python int or float. This is necessary because some of our transforms return Numpy values.

Typeutils Torch¶

ax.utils.common.typeutils_torch.torch_type_from_str(identifier: str, type_name: str) → Union[dtype, device][source]¶

ax.utils.common.typeutils_torch.torch_type_to_str(value: Any) → str[source]¶

Converts torch types, commonly used in Ax, to string representations.

Docstring Checker¶

ax.utils.flake8_plugins.docstring_checker.A000(node: AST) → Error¶

class ax.utils.flake8_plugins.docstring_checker.DocstringChecker(tree, filename)[source]¶

Bases: object

A flake8 plug-in that makes sure all public functions have a docstring

fikename: str¶

name: str = 'docstring checker'¶

run()[source]¶

tree: Module¶

version: str = '1.0.0'¶

class ax.utils.flake8_plugins.docstring_checker.DocstringCheckerVisitor[source]¶

Bases: NodeVisitor

check_A000(node: AST) → None[source]¶

errors: List[Error]¶

visit_AsyncFunctionDef(node: ClassDef) → None[source]¶

visit_ClassDef(node: ClassDef) → None[source]¶

visit_FunctionDef(node: FunctionDef) → None[source]¶

class ax.utils.flake8_plugins.docstring_checker.Error(lineno, col, message, type)[source]¶

Bases: tuple

col: int¶

Alias for field number 1

lineno: int¶

Alias for field number 0

message: str¶

Alias for field number 2

type: Type¶

Alias for field number 3

ax.utils.flake8_plugins.docstring_checker.is_copy_doc_call(c)[source]¶

Tries to guess if this is a call to the copy_doc decorator. This is a purely syntactic check so if the decorator was aliased as another name] or wrapped in another function we will fail.

ax.utils.flake8_plugins.docstring_checker.new_error(errorid: str, msg: str) → Callable[[AST], Error][source]¶

ax.utils.flake8_plugins.docstring_checker.should_check(filename)[source]¶

Synthetic Functions¶

class ax.utils.measurement.synthetic_functions.Aug_Branin[source]¶

Bases: SyntheticFunction

Augmented Branin function (3-dimensional with infinitely many global minima).

class ax.utils.measurement.synthetic_functions.Aug_Hartmann6[source]¶

Bases: Hartmann6

Augmented Hartmann6 function (7-dimensional with 1 global minimum).

class ax.utils.measurement.synthetic_functions.Branin[source]¶

Bases: SyntheticFunction

Branin function (2-dimensional with 3 global minima).

class ax.utils.measurement.synthetic_functions.FromBotorch(botorch_synthetic_function: SyntheticTestFunction)[source]¶

Bases: SyntheticFunction

property name: str¶

class ax.utils.measurement.synthetic_functions.Hartmann6[source]¶

Bases: SyntheticFunction

Hartmann6 function (6-dimensional with 1 global minimum).

class ax.utils.measurement.synthetic_functions.SyntheticFunction[source]¶

Bases: ABC

property domain: Any¶

f(X: ndarray) → Union[float, ndarray][source]¶

Synthetic function implementation.

Parameters:

X (numpy.ndarray) – an n by d array, where n represents the number of observations and d is the dimensionality of the inputs.

Returns:

an n-dimensional array.

Return type:

numpy.ndarray

property fmax: Any¶

property fmin: Any¶

property maximums: Any¶

property minimums: Any¶

property name: Any¶

property required_dimensionality: Any¶

ax.utils.measurement.synthetic_functions.from_botorch(botorch_synthetic_function: SyntheticTestFunction) → SyntheticFunction[source]¶

Utility to generate Ax synthetic functions from BoTorch synthetic functions.

ax.utils.measurement.synthetic_functions.informative_failure_on_none(func: Callable) → Any[source]¶

Plotting¶

ax.utils.notebook.plotting.init_notebook_plotting(offline: bool = False) → None[source]¶

Initialize plotting in notebooks, either in online or offline mode.

ax.utils.notebook.plotting.render(plot_config: AxPlotConfig, inject_helpers: bool = False) → None[source]¶

Render plot config.

Render¶

ax.utils.report.render.h2_html(text: str) → str[source]¶

Embed text in subheading tag.

ax.utils.report.render.h3_html(text: str) → str[source]¶

Embed text in subsubheading tag.

ax.utils.report.render.link_html(text: str, href: str) → str[source]¶

Embed text and reference address into link tag.

ax.utils.report.render.list_item_html(text: str) → str[source]¶

Embed text in list element tag.

ax.utils.report.render.p_html(text: str) → str[source]¶

Embed text in paragraph tag.

ax.utils.report.render.render_report_elements(experiment_name: str, html_elements: List[str], header: bool = True, offline: bool = False, notebook_env: bool = False) → str[source]¶

Generate Ax HTML report for a given experiment from HTML elements.

Uses Jinja2 for template. Injects Plotly JS for graph rendering.

Example:

html_elements = [
    h2_html("Subsection with plot"),
    p_html("This is an example paragraph."),
    plot_html(plot_fitted(gp_model, 'perf_metric')),
    h2_html("Subsection with table"),
    pandas_html(data.df),
]
html = render_report_elements('My experiment', html_elements)

Parameters:

• experiment_name – the name of the experiment to use for title.

• html_elements – list of HTML strings to render in report body.

• header – if True, render experiment title as a header. Meant to be used for standalone reports (e.g. via email), as opposed to served on the front-end.

• offline – if True, entire Plotly library is bundled with report.

• notebook_env – if True, caps the report width to 700px for viewing in a notebook environment.

Returns:

HTML string.

Return type:

str

ax.utils.report.render.table_cell_html(text: str, width: Optional[str] = None) → str[source]¶

Embed text or an HTML element into table cell tag.

ax.utils.report.render.table_heading_cell_html(text: str) → str[source]¶

Embed text or an HTML element into table heading cell tag.

ax.utils.report.render.table_html(table_rows: List[str]) → str[source]¶

Embed list of HTML elements into table tag.

ax.utils.report.render.table_row_html(table_cells: List[str]) → str[source]¶

Embed list of HTML elements into table row tag.

ax.utils.report.render.unordered_list_html(list_items: List[str]) → str[source]¶

Embed list of html elements into an unordered list tag.

Derivative GP¶

ax.utils.sensitivity.derivative_gp.get_KXX_inv(gp: Model) → Tensor[source]¶

Get the inverse matrix of K(X,X). :param gp: Botorch model.

Returns:

The inverse of K(X,X).

ax.utils.sensitivity.derivative_gp.get_KxX_dx(gp: Model, x: Tensor, kernel_type: str = 'rbf') → Tensor[source]¶

Computes the analytic derivative of the kernel K(x,X) w.r.t. x. :param gp: Botorch model. :param x: (n x D) Test points. :param kernel_type: Takes “rbf” or “matern_l1” or “matern_l2”

Returns:

Tensor (n x D) The derivative of the kernel K(x,X) w.r.t. x.

ax.utils.sensitivity.derivative_gp.get_Kxx_dx2(gp: Model, kernel_type: str = 'rbf') → Tensor[source]¶

Computes the analytic second derivative of the kernel w.r.t. the training data :param gp: Botorch model. :param kernel_type: Takes “rbf” or “matern_l1” or “matern_l2”

Returns:

Tensor (n x D x D) The second derivative of the kernel w.r.t. the training data.

ax.utils.sensitivity.derivative_gp.posterior_derivative(gp: Model, x: Tensor, kernel_type: str = 'rbf') → MultivariateNormal[source]¶

Computes the posterior of the derivative of the GP w.r.t. the given test points x. This follows the derivation used by GIBO in Sarah Muller, Alexander von Rohr, Sebastian Trimpe. “Local policy search with Bayesian optimization”, Advances in Neural Information Processing Systems 34, NeurIPS 2021. :param gp: Botorch model :param x: (n x D) Test points. :param kernel_type: Takes “rbf” or “matern_l1” or “matern_l2”

Returns:

A Botorch Posterior.

Derivative Measures¶

class ax.utils.sensitivity.derivative_measures.GpDGSMGpMean(model: Model, bounds: Tensor, derivative_gp: bool = False, kernel_type: Optional[str] = None, Y_scale: float = 1.0, num_mc_samples: int = 10000, input_qmc: bool = False, dtype: dtype = torch.float64, num_bootstrap_samples: int = 1)[source]¶

Bases: object

aggregation(transform_fun: Callable[[Tensor], Tensor]) → Tensor[source]¶

bootstrap_indices: Optional[Tensor] = None¶

gradient_absolute_measure() → Tensor[source]¶

Computes the gradient absolute measure:

Returns:

if self.num_bootstrap_samples > 1

Tensor: (values, var_mc, stderr_mc) x dim

else

Tensor: (values) x dim

gradient_measure() → Tensor[source]¶

Computes the gradient measure:

Returns:

if self.num_bootstrap_samples > 1

Tensor: (values, var_mc, stderr_mc) x dim

else

Tensor: (values) x dim

gradients_square_measure() → Tensor[source]¶

Computes the gradient square measure:

Returns:

if num_bootstrap_samples > 1

Tensor: (values, var_mc, stderr_mc) x dim

else

Tensor: (values) x dim

mean_gradients: Optional[Tensor] = None¶

mean_gradients_btsp: Optional[List[Tensor]] = None¶

class ax.utils.sensitivity.derivative_measures.GpDGSMGpSampling(model: Model, bounds: Tensor, num_gp_samples: int, derivative_gp: bool = False, kernel_type: Optional[str] = None, Y_scale: float = 1.0, num_mc_samples: int = 10000, input_qmc: bool = False, gp_sample_qmc: bool = False, dtype: dtype = torch.float64, num_bootstrap_samples: int = 1)[source]¶

Bases: GpDGSMGpMean

aggregation(transform_fun: Callable[[Tensor], Tensor]) → Tensor[source]¶

samples_gradients: Optional[Tensor] = None¶

samples_gradients_btsp: Optional[List[Tensor]] = None¶

Sobol Measures¶

class ax.utils.sensitivity.sobol_measures.SobolSensitivity(dim: int, bounds: Tensor, input_function: Optional[Callable[[Tensor], Tensor]] = None, num_mc_samples: int = 10000, input_qmc: bool = False, second_order: bool = False, num_bootstrap_samples: int = 1, bootstrap_array: bool = False)[source]¶

Bases: object

evalute_function(f_A_B_ABi: Optional[Tensor] = None) → None[source]¶

evaluates the objective function and devides the evaluation into

torch.Tensors needed for the indices computation.

Parameters:

f_A_B_ABi – Function evaluations on the entire grid of size M(d+2).

first_order_indices() → Tensor[source]¶

Computes the first order Sobol indices:

Returns:

if num_bootstrap_samples>1

Tensor: (values,var_mc,stderr_mc)x dim

else

Tensor: (values)x dim

generate_all_input_matrix() → Tensor[source]¶

second_order_indices(first_order_idxs: Optional[Tensor] = None, first_order_idxs_btsp: Optional[Tensor] = None)[source]¶

Computes the Second order Sobol indices: :param first_order_idxs: Tensor of first order indices. :param first_order_idxs_btsp: Tensor of all first order indices given by bootstrap.

Returns:

if num_bootstrap_samples>1

Tensor: (values,var_mc,stderr_mc)x dim

else

Tensor: (values)x dim

total_order_indices()[source]¶

Computes the total Sobol indices:

Returns:

if num_bootstrap_samples>1

Tensor: (values,var_mc,stderr_mc)x dim

else

Tensor: (values)x dim

class ax.utils.sensitivity.sobol_measures.SobolSensitivityGPMean(model: Model, bounds: Tensor, num_mc_samples: int = 10000, second_order: bool = False, input_qmc: bool = False, num_bootstrap_samples: int = 1)[source]¶

Bases: object

first_order_indices() → Tensor[source]¶

Computes the first order Sobol indices:

Returns:

if num_bootstrap_samples>1

Tensor: (values,var_mc,stderr_mc)x dim

else

Tensor: (values)x dim

second_order_indices()[source]¶

Computes the Second order Sobol indices:

Returns:

if num_bootstrap_samples>1

Tensor: (values,var_mc,stderr_mc)x dim(dim-1)/2

else

Tensor: (values)x dim(dim-1)/2

total_order_indices()[source]¶

Computes the total Sobol indices:

Returns:

if num_bootstrap_samples>1

Tensor: (values,var_mc,stderr_mc)x dim

else

Tensor: (values)x dim

class ax.utils.sensitivity.sobol_measures.SobolSensitivityGPSampling(model: Model, bounds: Tensor, num_gp_samples: int = 1000, num_mc_samples: int = 10000, second_order: bool = False, input_qmc: bool = False, gp_sample_qmc: bool = False, num_bootstrap_samples: int = 1)[source]¶

Bases: object

first_order_indices() → Tensor[source]¶

Computes the first order Sobol indices:

Returns:

if num_bootstrap_samples>1

Tensor: (values,var_gp,stderr_gp,var_mc,stderr_mc)x dim

else

Tensor: (values,var,stderr)x dim

second_order_indices() → Tensor[source]¶

Computes the Second order Sobol indices:

Returns:

if num_bootstrap_samples>1

Tensor: (values,var_gp,stderr_gp,var_mc,stderr_mc)x dim(dim-1)/2

else

Tensor: (values,var,stderr)x dim(dim-1)/2

total_order_indices() → Tensor[source]¶

Computes the total Sobol indices:

Returns:

if num_bootstrap_samples>1

Tensor: (values,var_gp,stderr_gp,var_mc,stderr_mc)x dim

else

Tensor: (values,var,stderr)x dim

Statstools¶

ax.utils.stats.statstools.agresti_coull_sem(n_numer: Union[pandas.Series, ndarray, int], n_denom: Union[pandas.Series, ndarray, int], prior_successes: int = 2, prior_failures: int = 2) → Union[ndarray, float][source]¶

Compute the Agresti-Coull style standard error for a binomial proportion.

Reference: Agresti, Alan, and Brent A. Coull. Approximate Is Better than ‘Exact’ for Interval Estimation of Binomial Proportions.” The American Statistician, vol. 52, no. 2, 1998, pp. 119-126. JSTOR, www.jstor.org/stable/2685469.

ax.utils.stats.statstools.inverse_variance_weight(means: ndarray, variances: ndarray, conflicting_noiseless: str = 'warn') → Tuple[float, float][source]¶

Perform inverse variance weighting.

Parameters:

• means – The means of the observations.

• variances – The variances of the observations.

• conflicting_noiseless – How to handle the case of multiple observations with zero variance but different means. Options are “warn” (default), “ignore” or “raise”.

ax.utils.stats.statstools.marginal_effects(df: pandas.DataFrame) → pandas.DataFrame[source]¶

This method calculates the relative (in %) change in the outcome achieved by using any individual factor level versus randomizing across all factor levels. It does this by estimating a baseline under the experiment by marginalizing over all factors/levels. For each factor level, then, it conditions on that level for the individual factor and then marginalizes over all levels for all other factors.

Parameters:

df – Dataframe containing columns named mean and sem. All other columns are assumed to be factors for which to calculate marginal effects.

Returns:

A dataframe containing columns “Name”, “Level”, “Beta” and “SE”

corresponding to the factor, level, effect and standard error. Results are relativized as percentage changes.

ax.utils.stats.statstools.positive_part_james_stein(means: Union[ndarray, List[float]], sems: Union[ndarray, List[float]]) → Tuple[ndarray, ndarray][source]¶

Estimation method for Positive-part James-Stein estimator.

This method takes a vector of K means (y_i) and standard errors (sigma_i) and calculates the positive-part James Stein estimator.

Resulting estimates are the shrunk means and standard errors. The positive part James-Stein estimator shrinks each constituent average to the grand average:

y_i - phi_i * y_i + phi_i * ybar

The variable phi_i determines the amount of shrinkage. For phi_i = 1, mu_hat is equal to ybar (the mean of all y_i), while for phi_i = 0, mu_hat is equal to y_i. It can be shown that restricting phi_i <= 1 dominates the unrestricted estimator, so this method restricts phi_i in this manner. The amount of shrinkage, phi_i, is determined by:

(K - 3) * sigma2_i / s2

That is, less shrinkage is applied when individual means are estimated with greater precision, and more shrinkage is applied when individual means are very tightly clustered together. We also restrict phi_i to never be larger than 1.

The variance of the mean estimator is:

(1 - phi_i) * sigma2_i + phi * sigma2_i / K + 2 * phi_i ** 2 * (y_i - ybar)^2 / (K - 3)

The first term is the variance component from y_i, the second term is the contribution from the mean of all y_i, and the third term is the contribution from the uncertainty in the sum of squared deviations of y_i from the mean of all y_i.

For more information, see https://ax.dev/docs/models.html#empirical-bayes-and-thompson-sampling.

Parameters:

• means – Means of each arm

• sems – Standard errors of each arm

Returns:

Empirical Bayes estimate of each arm’s mean sem_i: Empirical Bayes estimate of each arm’s sem

Return type:

mu_hat_i

ax.utils.stats.statstools.relativize(means_t: Union[ndarray, List[float], float], sems_t: Union[ndarray, List[float], float], mean_c: float, sem_c: float, bias_correction: bool = True, cov_means: Union[ndarray, List[float], float] = 0.0, as_percent: bool = False) → Tuple[ndarray, ndarray][source]¶

Ratio estimator based on the delta method.

This uses the delta method (i.e. a Taylor series approximation) to estimate the mean and standard deviation of the sampling distribution of the ratio between test and control – that is, the sampling distribution of an estimator of the true population value under the assumption that the means in test and control have a known covariance:

(mu_t / mu_c) - 1.

Under a second-order Taylor expansion, the sampling distribution of the relative change in empirical means, which is m_t / m_c - 1, is approximately normally distributed with mean

[(mu_t - mu_c) / mu_c] - [(sigma_c)^2 * mu_t] / (mu_c)^3

and variance

(sigma_t / mu_c)^2 - 2 * mu_t _ sigma_tc / mu_c^3 + [(sigma_c * mu_t)^2 / (mu_c)^4]

as the higher terms are assumed to be close to zero in the full Taylor series. To estimate these parameters, we plug in the empirical means and standard errors. This gives us the estimators:

[(m_t - m_c) / m_c] - [(s_c)^2 * m_t] / (m_c)^3

and

(s_t / m_c)^2 - 2 * m_t * s_tc / m_c^3 + [(s_c * m_t)^2 / (m_c)^4]

Note that the delta method does NOT take as input the empirical standard deviation of a metric, but rather the standard error of the mean of that metric – that is, the standard deviation of the metric after division by the square root of the total number of observations.

Parameters:

• means_t – Sample means (test)

• sems_t – Sample standard errors of the means (test)

• mean_c – Sample mean (control)

• sem_c – Sample standard error of the mean (control)

• cov_means – Sample covariance between test and control

• as_percent – If true, return results in percent (* 100)

Returns:

Inferred means of the sampling distribution of

the relative change (mean_t / mean_c) - 1

sem_hat: Inferred standard deviation of the sampling

distribution of rel_hat – i.e. the standard error.

Return type:

rel_hat

ax.utils.stats.statstools.relativize_data(data: Data, status_quo_name: str = 'status_quo', as_percent: bool = False, include_sq: bool = False) → Data[source]¶

Relativize a data object w.r.t. a status_quo arm.

Parameters:

• data – The data object to be relativized.

• status_quo_name – The name of the status_quo arm.

• as_percent – If True, return results as percentage change.

• include_sq – Include status quo in final df.

Returns:

The new data object with the relativized metrics (excluding the

status_quo arm)

ax.utils.stats.statstools.total_variance(means: ndarray, variances: ndarray, sample_sizes: ndarray) → float[source]¶

Compute total variance.

Deletion¶

Backend Scheduler¶

Backend Simulator¶

class ax.utils.testing.backend_simulator.BackendSimulator(options: Optional[BackendSimulatorOptions] = None, queued: Optional[List[SimTrial]] = None, running: Optional[List[SimTrial]] = None, failed: Optional[List[SimTrial]] = None, completed: Optional[List[SimTrial]] = None, verbose_logging: bool = True)[source]¶

Bases: object

Simulator for a backend deployment with concurrent dispatch and a queue.

property all_trials: List[SimTrial]¶

All trials on the simulator.

classmethod from_state(state: BackendSimulatorState)[source]¶

Construct a simulator from a state.

Parameters:

state – A BackendSimulatorState to set the simulator to.

Returns:

A BackendSimulator with the desired state.

get_sim_trial_by_index(trial_index: int) → Optional[SimTrial][source]¶

Get a SimTrial by trial_index.

Parameters:

trial_index – The index of the trial to return.

Returns:

A SimTrial with the index trial_index or None if not found.

lookup_trial_index_status(trial_index: int) → Optional[TrialStatus][source]¶

Lookup the trial status of a trial_index.

Parameters:

trial_index – The index of the trial to check.

Returns:

A TrialStatus.

new_trial(trial: SimTrial, status: TrialStatus) → None[source]¶

Register a trial into the simulator.

Parameters:

• trial – A new trial to add.

• status – The status of the new trial, either STAGED (add to self._queued) or RUNNING (add to self._running).

property num_completed: int¶

The number of completed trials.

property num_failed: int¶

The number of failed trials.

property num_queued: int¶

The number of queued trials (to run as soon as capacity is available).

property num_running: int¶

The number of currently running trials.

reset() → None[source]¶

Reset the simulator.

run_trial(trial_index: int, runtime: float) → None[source]¶

Run a simulated trial.

Parameters:

• trial_index – The index of the trial (usually the Ax trial index)

• runtime – The runtime of the simulation. Typically sampled from the runtime model of a simulation model.

Internally, the runtime is scaled by the time_scaling factor, so that the simulation can run arbitrarily faster than the underlying evaluation.

state() → BackendSimulatorState[source]¶

Return a BackendSimulatorState containing the state of the simulator.

status() → SimStatus[source]¶

Return the internal status of the simulator.

Returns:

A SimStatus object representing the current simulator state.

stop_trial(trial_index: int) → None[source]¶

Stop a simulated trial by setting the completed time to the current time.

Parameters:

trial_index – The index of the trial to stop.

property time: float¶

The current time.

update() → None[source]¶

Update the state of the simulator.

property use_internal_clock: bool¶

Whether or not we are using the internal clock.