forked from M-Labs/artiq
1
0
Fork 0
artiq/artiq/language/environment.py

504 lines
18 KiB
Python

from collections import OrderedDict
from inspect import isclass
from sipyco import pyon
from artiq.language import units
from artiq.language.core import rpc
__all__ = ["NoDefault", "DefaultMissing",
"PYONValue", "BooleanValue", "EnumerationValue",
"NumberValue", "StringValue",
"HasEnvironment", "Experiment", "EnvExperiment"]
class NoDefault:
"""Represents the absence of a default value."""
pass
class DefaultMissing(Exception):
"""Raised by the ``default`` method of argument processors when no default
value is available."""
pass
class _SimpleArgProcessor:
def __init__(self, default=NoDefault):
# If default is a list, it means multiple defaults are specified, with
# decreasing priority.
if isinstance(default, list):
raise NotImplementedError
if default is not NoDefault:
self.default_value = self.process(default)
def default(self):
if not hasattr(self, "default_value"):
raise DefaultMissing
return self.default_value
def process(self, x):
return x
def describe(self):
d = {"ty": self.__class__.__name__}
if hasattr(self, "default_value"):
d["default"] = self.default_value
return d
class PYONValue(_SimpleArgProcessor):
"""An argument that can be any PYON-serializable value."""
def __init__(self, default=NoDefault):
# Override the _SimpleArgProcessor init, as list defaults are valid
# PYON values
if default is not NoDefault:
self.default_value = default
def process(self, x):
return pyon.decode(x)
def describe(self):
d = {"ty": self.__class__.__name__}
if hasattr(self, "default_value"):
d["default"] = pyon.encode(self.default_value)
return d
class BooleanValue(_SimpleArgProcessor):
"""A boolean argument."""
def process(self, x):
if type(x) != bool:
raise ValueError("Invalid BooleanValue value")
return x
class EnumerationValue(_SimpleArgProcessor):
"""An argument that can take a string value among a predefined set of
values.
:param choices: A list of string representing the possible values of the
argument.
"""
def __init__(self, choices, default=NoDefault):
self.choices = choices
super().__init__(default)
def process(self, x):
if x not in self.choices:
raise ValueError("Invalid EnumerationValue value")
return x
def describe(self):
d = _SimpleArgProcessor.describe(self)
d["choices"] = self.choices
return d
class NumberValue(_SimpleArgProcessor):
"""An argument that can take a numerical value.
If ``type=="auto"``, the result will be a ``float`` unless
ndecimals = 0, scale = 1 and step is an integer. Setting ``type`` to
``int`` will also result in an error unless these conditions are met.
When ``scale`` is not specified, and the unit is a common one (i.e.
defined in ``artiq.language.units``), then the scale is obtained from
the unit using a simple string match. For example, milliseconds (``"ms"``)
units set the scale to 0.001. No unit (default) corresponds to a scale of
1.0.
For arguments with uncommon or complex units, use both the unit parameter
(a string for display) and the scale parameter (a numerical scale for
experiments).
For example, ``NumberValue(1, unit="xyz", scale=0.001)`` will display as
1 xyz in the GUI window because of the unit setting, and appear as the
numerical value 0.001 in the code because of the scale setting.
:param unit: A string representing the unit of the value.
:param scale: A numerical scaling factor by which the displayed value is
multiplied when referenced in the experiment.
:param step: The step with which the value should be modified by up/down
buttons in a UI. The default is the scale divided by 10.
:param min: The minimum value of the argument.
:param max: The maximum value of the argument.
:param ndecimals: The number of decimals a UI should use.
:param type: Type of this number. Accepts ``"float"``, ``"int"`` or
``"auto"``. Defaults to ``"auto"``.
"""
valid_types = ["auto", "float", "int"]
def __init__(self, default=NoDefault, unit="", scale=None,
step=None, min=None, max=None, ndecimals=2, type="auto"):
if scale is None:
if unit == "":
scale = 1.0
else:
try:
scale = getattr(units, unit)
except AttributeError:
raise KeyError("Unit {} is unknown, you must specify "
"the scale manually".format(unit))
if step is None:
step = scale/10.0
self.unit = unit
self.scale = scale
self.step = step
self.min = min
self.max = max
self.ndecimals = ndecimals
self.type = type
if self.type not in NumberValue.valid_types:
raise TypeError("type must be 'float', 'int' or 'auto'")
if self.type == "int" and not self._is_int_compatible():
raise ValueError(("Value marked as integer but settings are "
"not compatible. Please set ndecimals = 0, "
"scale = 1 and step to an integer"))
super().__init__(default)
def _is_int_compatible(self):
'''
Are the settings other than `type` compatible with this being
an integer?
'''
return (self.ndecimals == 0
and int(self.step) == self.step
and self.scale == 1)
def _is_int(self):
'''
Will this argument return an integer?
'''
return (self.type == "int"
or (self.type == "auto" and self._is_int_compatible()))
def process(self, x):
if self._is_int():
return int(x)
else:
return float(x)
def describe(self):
d = {"ty": self.__class__.__name__}
if hasattr(self, "default_value"):
d["default"] = self.default_value
d["unit"] = self.unit
d["scale"] = self.scale
d["step"] = self.step
d["min"] = self.min
d["max"] = self.max
d["ndecimals"] = self.ndecimals
d["type"] = self.type
return d
class StringValue(_SimpleArgProcessor):
"""A string argument."""
pass
class TraceArgumentManager:
def __init__(self):
self.requested_args = OrderedDict()
def get(self, key, processor, group, tooltip):
self.requested_args[key] = processor, group, tooltip
return None
def check_unprocessed_arguments(self):
pass
class ProcessArgumentManager:
def __init__(self, unprocessed_arguments):
self.unprocessed_arguments = unprocessed_arguments
self._processed_arguments = set()
def get(self, key, processor, group, tooltip):
if key in self.unprocessed_arguments:
r = processor.process(self.unprocessed_arguments[key])
self._processed_arguments.add(key)
else:
r = processor.default()
return r
def check_unprocessed_arguments(self):
unprocessed = set(self.unprocessed_arguments.keys()) -\
self._processed_arguments
if unprocessed:
raise AttributeError("Invalid argument(s): " +
", ".join(unprocessed))
class HasEnvironment:
"""Provides methods to manage the environment of an experiment (arguments,
devices, datasets)."""
def __init__(self, managers_or_parent, *args, **kwargs):
self.children = []
if isinstance(managers_or_parent, tuple):
self.__device_mgr = managers_or_parent[0]
self.__dataset_mgr = managers_or_parent[1]
self.__argument_mgr = managers_or_parent[2]
self.__scheduler_defaults = managers_or_parent[3]
else:
self.__device_mgr = managers_or_parent.__device_mgr
self.__dataset_mgr = managers_or_parent.__dataset_mgr
self.__argument_mgr = managers_or_parent.__argument_mgr
self.__scheduler_defaults = {}
managers_or_parent.register_child(self)
self.__in_build = True
self.build(*args, **kwargs)
self.__in_build = False
if self.__argument_mgr is not None:
self.__argument_mgr.check_unprocessed_arguments()
def register_child(self, child):
self.children.append(child)
def call_child_method(self, method, *args, **kwargs):
"""Calls the named method for each child, if it exists for that child,
in the order of registration.
:param method: Name of the method to call
:type method: str
:param args: Tuple of positional arguments to pass to all children
:param kwargs: Dict of keyword arguments to pass to all children
"""
for child in self.children:
try:
child_method = getattr(child, method)
except AttributeError:
pass
else:
child_method(*args, **kwargs)
def build(self):
"""Should be implemented by the user to request arguments.
Other initialization steps such as requesting devices may also be
performed here.
There are two situations where the requested devices are replaced by
``DummyDevice()`` and arguments are set to their defaults (or ``None``)
instead: when the repository is scanned to build the list of
available experiments and when the dataset browser ``artiq_browser``
is used to open or run the analysis stage of an experiment. Do not
rely on being able to operate on devices or arguments in :meth:`build`.
Datasets are read-only in this method.
Leftover positional and keyword arguments from the constructor are
forwarded to this method. This is intended for experiments that are
only meant to be executed programmatically (not from the GUI)."""
pass
def get_argument(self, key, processor, group=None, tooltip=None):
"""Retrieves and returns the value of an argument.
This function should only be called from ``build``.
:param key: Name of the argument.
:param processor: A description of how to process the argument, such
as instances of ``BooleanValue`` and ``NumberValue``.
:param group: An optional string that defines what group the argument
belongs to, for user interface purposes.
:param tooltip: An optional string to describe the argument in more
detail, applied as a tooltip to the argument name in the user
interface.
"""
if not self.__in_build:
raise TypeError("get_argument() should only "
"be called from build()")
return self.__argument_mgr.get(key, processor, group, tooltip)
def setattr_argument(self, key, processor=None, group=None, tooltip=None):
"""Sets an argument as attribute. The names of the argument and of the
attribute are the same.
The key is added to the instance's kernel invariants."""
setattr(self, key, self.get_argument(key, processor, group, tooltip))
kernel_invariants = getattr(self, "kernel_invariants", set())
self.kernel_invariants = kernel_invariants | {key}
def get_device_db(self):
"""Returns the full contents of the device database."""
return self.__device_mgr.get_device_db()
def get_device(self, key):
"""Creates and returns a device driver."""
return self.__device_mgr.get(key)
def setattr_device(self, key):
"""Sets a device driver as attribute. The names of the device driver
and of the attribute are the same.
The key is added to the instance's kernel invariants."""
setattr(self, key, self.get_device(key))
kernel_invariants = getattr(self, "kernel_invariants", set())
self.kernel_invariants = kernel_invariants | {key}
@rpc(flags={"async"})
def set_dataset(self, key, value,
broadcast=False, persist=False, archive=True):
"""Sets the contents and handling modes of a dataset.
Datasets must be scalars (``bool``, ``int``, ``float`` or NumPy scalar)
or NumPy arrays.
:param broadcast: the data is sent in real-time to the master, which
dispatches it.
:param persist: the master should store the data on-disk. Implies
broadcast.
:param archive: the data is saved into the local storage of the current
run (archived as a HDF5 file).
"""
self.__dataset_mgr.set(key, value, broadcast, persist, archive)
@rpc(flags={"async"})
def mutate_dataset(self, key, index, value):
"""Mutate an existing dataset at the given index (e.g. set a value at
a given position in a NumPy array)
If the dataset was created in broadcast mode, the modification is
immediately transmitted.
If the index is a tuple of integers, it is interpreted as
``slice(*index)``.
If the index is a tuple of tuples, each sub-tuple is interpreted
as ``slice(*sub_tuple)`` (multi-dimensional slicing)."""
self.__dataset_mgr.mutate(key, index, value)
@rpc(flags={"async"})
def append_to_dataset(self, key, value):
"""Append a value to a dataset.
The target dataset must be a list (i.e. support ``append()``), and must
have previously been set from this experiment.
The broadcast/persist/archive mode of the given key remains unchanged
from when the dataset was last set. Appended values are transmitted
efficiently as incremental modifications in broadcast mode."""
self.__dataset_mgr.append_to(key, value)
def get_dataset(self, key, default=NoDefault, archive=True):
"""Returns the contents of a dataset.
The local storage is searched first, followed by the master storage
(which contains the broadcasted datasets from all experiments) if the
key was not found initially.
If the dataset does not exist, returns the default value. If no default
is provided, raises ``KeyError``.
By default, datasets obtained by this method are archived into the output
HDF5 file of the experiment. If an archived dataset is requested more
than one time or is modified, only the value at the time of the first call
is archived. This may impact reproducibility of experiments.
:param archive: Set to ``False`` to prevent archival together with the run's results.
Default is ``True``.
"""
try:
return self.__dataset_mgr.get(key, archive)
except KeyError:
if default is NoDefault:
raise
else:
return default
def setattr_dataset(self, key, default=NoDefault, archive=True):
"""Sets the contents of a dataset as attribute. The names of the
dataset and of the attribute are the same."""
setattr(self, key, self.get_dataset(key, default, archive))
def set_default_scheduling(self, priority=None, pipeline_name=None, flush=None):
"""Sets the default scheduling options.
This function should only be called from ``build``."""
if not self.__in_build:
raise TypeError("set_default_scheduling() should only "
"be called from build()")
if priority is not None:
self.__scheduler_defaults["priority"] = int(priority)
if pipeline_name is not None:
self.__scheduler_defaults["pipeline_name"] = pipeline_name
if flush is not None:
self.__scheduler_defaults["flush"] = flush
class Experiment:
"""Base class for top-level experiments.
Deriving from this class enables automatic experiment discovery in
Python modules.
"""
def prepare(self):
"""Entry point for pre-computing data necessary for running the
experiment.
Doing such computations outside of :meth:`run` enables more efficient
scheduling of multiple experiments that need to access the shared
hardware during part of their execution.
This method must not interact with the hardware.
"""
pass
def run(self):
"""The main entry point of the experiment.
This method must be overloaded by the user to implement the main
control flow of the experiment.
This method may interact with the hardware.
The experiment may call the scheduler's :meth:`pause` method while in
:meth:`run`.
"""
raise NotImplementedError
def analyze(self):
"""Entry point for analyzing the results of the experiment.
This method may be overloaded by the user to implement the analysis
phase of the experiment, for example fitting curves.
Splitting this phase from :meth:`run` enables tweaking the analysis
algorithm on pre-existing data, and CPU-bound analyses to be run
overlapped with the next experiment in a pipelined manner.
This method must not interact with the hardware.
"""
pass
class EnvExperiment(Experiment, HasEnvironment):
"""Base class for top-level experiments that use the
:class:`~artiq.language.environment.HasEnvironment` environment manager.
Most experiments should derive from this class."""
def prepare(self):
"""This default prepare method calls :meth:`~artiq.language.environment.Experiment.prepare`
for all children, in the order of registration, if the child has a
:meth:`~artiq.language.environment.Experiment.prepare` method."""
self.call_child_method("prepare")
def is_experiment(o):
"""Checks if a Python object is a top-level experiment class."""
return (isclass(o)
and issubclass(o, Experiment)
and o is not Experiment
and o is not EnvExperiment)
def is_public_experiment(o):
"""Checks if a Python object is a top-level,
non underscore-prefixed, experiment class.
"""
return is_experiment(o) and not o.__name__.startswith("_")