import sys import warnings from collections.abc import Callable, Iterator, Mapping from contextlib import contextmanager, suppress from typing import ( TYPE_CHECKING, Any, ClassVar, Union, cast, no_type_check, ) import pydantic from pydantic import PrivateAttr from ._group import SignalGroup from ._group_descriptor import _check_field_equality, _pick_equality_operator from ._signal import ReemissionMode, Signal if TYPE_CHECKING: from inspect import Signature from typing import TypeGuard from pydantic._internal import _model_construction as pydantic_main from pydantic._internal import _utils as utils from pydantic._internal._decorators import PydanticDescriptorProxy from typing_extensions import dataclass_transform as dataclass_transform # py311 from ._signal import SignalInstance EqOperator = Callable[[Any, Any], bool] else: from pydantic._internal import _model_construction as pydantic_main from pydantic._internal import _utils as utils try: # py311 from typing_extensions import dataclass_transform except ImportError: # pragma: no cover def dataclass_transform(*args, **kwargs): return lambda a: a NULL = object() ALLOW_PROPERTY_SETTERS = "allow_property_setters" FIELD_DEPENDENCIES = "field_dependencies" GUESS_PROPERTY_DEPENDENCIES = "guess_property_dependencies" REEMISSION = "reemission" @contextmanager def no_class_attributes() -> Iterator[None]: # pragma: no cover """Context in which pydantic_main.ClassAttribute just passes value 2. Due to a very annoying decision by PySide2, all class ``__signature__`` attributes may only be assigned **once**. (This seems to be regardless of whether the class has anything to do with PySide2 or not). Furthermore, the PySide2 ``__signature__`` attribute seems to break the python descriptor protocol, which means that class attributes that have a ``__get__`` method will not be able to successfully retrieve their value (instead, the descriptor object itself will be accessed). This plays terribly with Pydantic, which assigns a ``ClassAttribute`` object to the value of ``cls.__signature__`` in ``ModelMetaclass.__new__`` in order to avoid masking the call signature of object instances that have a ``__call__`` method (https://github.com/samuelcolvin/pydantic/pull/1466). So, because we only get to set the ``__signature__`` once, this context manager basically "opts-out" of pydantic's ``ClassAttribute`` strategy, thereby directly setting the ``cls.__signature__`` to an instance of ``inspect.Signature``. For additional context, see: - https://github.com/napari/napari/issues/2264 - https://github.com/napari/napari/pull/2265 - https://bugreports.qt.io/browse/PYSIDE-1004 - https://codereview.qt-project.org/c/pyside/pyside-setup/+/261411 """ if "PySide2" not in sys.modules: yield return # monkey patch the pydantic ClassAttribute object # the second argument to ClassAttribute is the inspect.Signature object def _return2(x: str, y: "Signature") -> "Signature": return y pydantic_main.ClassAttribute = _return2 # type: ignore try: yield finally: # undo our monkey patch pydantic_main.ClassAttribute = utils.ClassAttribute # type: ignore def _get_defaults( obj: pydantic.BaseModel | type[pydantic.BaseModel], ) -> dict[str, Any]: """Get possibly nested default values for a Model object.""" dflt = {} cls = obj if isinstance(obj, type) else type(obj) for k, v in cls.model_fields.items(): d = v.get_default() if ( d is None and isinstance(v.annotation, type) and issubclass(v.annotation, pydantic.BaseModel) ): d = _get_defaults(v.annotation) # pragma: no cover dflt[k] = d return dflt def _get_fields( cls: type[pydantic.BaseModel], ) -> dict[str, pydantic.fields.FieldInfo]: comp_fields = { name: pydantic.fields.FieldInfo(annotation=f.return_type, frozen=False) for name, f in cls.model_computed_fields.items() } return {**cls.model_fields, **comp_fields} def _is_pydantic_descriptor_proxy(obj: Any) -> "TypeGuard[PydanticDescriptorProxy]": if ( type(obj).__module__.startswith("pydantic") and type(obj).__name__ == "PydanticDescriptorProxy" and isinstance(getattr(obj, "wrapped", None), property) ): return True return False class ComparisonDelayer: """Context that delays before/after comparisons until exit.""" def __init__(self, target: "EventedModel") -> None: self._target = target def __enter__(self) -> None: self._target._delay_check_semaphore += 1 def __exit__(self, *_: Any, **__: Any) -> None: self._target._delay_check_semaphore -= 1 self._target._check_if_values_changed_and_emit_if_needed() class EventedMetaclass(pydantic_main.ModelMetaclass): """pydantic ModelMetaclass that preps "equality checking" operations. A metaclass is the thing that "constructs" a class, and ``ModelMetaclass`` is where pydantic puts a lot of it's type introspection and ``ModelField`` creation logic. Here, we simply tack on one more function, that builds a ``cls.__eq_operators__`` dict which is mapping of field name to a function that can be called to check equality of the value of that field with some other object. (used in ``EventedModel.__eq__``) This happens only once, when an ``EventedModel`` class is created (and not when each instance of an ``EventedModel`` is instantiated). """ __property_setters__: dict[str, property] @no_type_check def __new__( mcs: type, name: str, bases: tuple, namespace: dict, **kwargs: Any ) -> "EventedMetaclass": """Create new EventedModel class.""" with no_class_attributes(): cls = super().__new__(mcs, name, bases, namespace, **kwargs) cls.__eq_operators__ = {} signals = {} model_fields = _get_fields(cls) model_config = cls.model_config emission_cfg = model_config.get(REEMISSION, {}) default_strategy: ReemissionMode = ReemissionMode.LATEST emission_map: Mapping[str, ReemissionMode] = {} if isinstance(emission_cfg, (str, ReemissionMode)): default_strategy = ReemissionMode.validate(emission_cfg) else: try: emission_map = { k: ReemissionMode.validate(v) for k, v in emission_cfg.items() } except (ValueError, TypeError) as e: valid = ", ".join(repr(x) for x in ReemissionMode._members()) raise ValueError( f"Invalid reemission value {emission_cfg!r}. Must be a mapping " f"of field names to one of {valid}." ) from e for n, f in model_fields.items(): cls.__eq_operators__[n] = _pick_equality_operator(f.annotation) if not f.frozen: recursion = emission_map.get(n, default_strategy) signals[n] = Signal(f.annotation, reemission=recursion) allow_props = model_config.get(ALLOW_PROPERTY_SETTERS, False) # check for @_.setters defined on the class, so we can allow them # in EventedModel.__setattr__ cls.__property_setters__ = {} if allow_props: # inherit property setters from base classes for b in reversed(cls.__bases__): if hasattr(b, "__property_setters__"): cls.__property_setters__.update(b.__property_setters__) # add property setters from this class for key, attr in namespace.items(): if _is_pydantic_descriptor_proxy(attr): attr = attr.wrapped if isinstance(attr, property) and attr.fset is not None: cls.__property_setters__[key] = attr recursion = emission_map.get(key, default_strategy) signals[key] = Signal(object, reemission=recursion) else: for b in cls.__bases__: with suppress(AttributeError): conf = b.model_config if conf and conf.get(ALLOW_PROPERTY_SETTERS, False): raise ValueError( "Cannot set 'allow_property_setters' to 'False' when base " f"class {b} sets it to True" ) cls.__field_dependents__ = _get_field_dependents( cls, model_config, model_fields ) cls.__signal_group__ = type(f"{name}SignalGroup", (SignalGroup,), signals) if not cls.__field_dependents__ and hasattr(cls, "_setattr_no_dependants"): cls._setattr_default = cls._setattr_no_dependants elif hasattr(cls, "_setattr_with_dependents"): cls._setattr_default = cls._setattr_with_dependents return cls def _get_field_dependents( cls: "EventedMetaclass", model_config: dict, model_fields: dict ) -> dict[str, set[str]]: """Return mapping of field name -> dependent set of property names. Dependencies may be declared in the Model Config to emit an event for a computed property when a model field that it depends on changes e.g. (@property 'c' depends on model fields 'a' and 'b') Examples -------- class MyModel(EventedModel): a: int = 1 b: int = 1 @property def c(self) -> List[int]: return [self.a, self.b] @c.setter def c(self, val: Sequence[int]): self.a, self.b = val class Config: field_dependencies={'c': ['a', 'b']} """ deps: dict[str, set[str]] = {} cfg_deps = model_config.get(FIELD_DEPENDENCIES, {}) # sourcery skip if cfg_deps: if not isinstance(cfg_deps, dict): # pragma: no cover raise TypeError( f"Config field_dependencies must be a dict, not {cfg_deps!r}" ) for prop, fields in cfg_deps.items(): if prop not in {*model_fields, *cls.__property_setters__}: raise ValueError( "Fields with dependencies must be fields or property.setters. " f"{prop!r} is not." ) for field in fields: if field not in model_fields and not hasattr(cls, field): warnings.warn( f"property {prop!r} cannot depend on unrecognized attribute " f"name: {field!r}", stacklevel=2, ) deps.setdefault(field, set()).add(prop) if model_config.get(GUESS_PROPERTY_DEPENDENCIES, False): # if field_dependencies haven't been explicitly defined, we can glean # them from the property.fget code object: # SKIP THIS MAGIC FOR NOW? for prop, setter in cls.__property_setters__.items(): if setter.fget is not None: for name in setter.fget.__code__.co_names: if name in model_fields: deps.setdefault(name, set()).add(prop) return deps @dataclass_transform(kw_only_default=True, field_specifiers=(pydantic.Field,)) class EventedModel(pydantic.BaseModel, metaclass=EventedMetaclass): """A pydantic BaseModel that emits a signal whenever a field value is changed. !!! important This class requires `pydantic` to be installed. You can install directly (`pip install pydantic`) or by using the psygnal extra: `pip install psygnal[pydantic]` In addition to standard pydantic `BaseModel` properties (see [pydantic docs](https://pydantic-docs.helpmanual.io/usage/models/)), this class adds the following: 1. Gains an `events` attribute that is an instance of [`psygnal.SignalGroup`][]. This group will have a signal for each field in the model (excluding private attributes and non-mutable fields). Whenever a field in the model is mutated, the corresponding signal will emit with the new value (see example below). 2. Gains support for properties and property.setters (not supported in pydantic's BaseModel). Enable by adding `allow_property_setters = True` to your model `Config`. 3. If you would like properties (i.e. "computed fields") to emit an event when one of the model fields it depends on is mutated you must set one of the following options in the `Config`: - `field_dependencies` may be a `Dict[str, List[str]]`, where the keys are the names of properties, and the values are a list of field names (strings) that the property depends on for its value - `guess_property_dependencies` may be set to `True` to "guess" property dependencies by inspecting the source code of the property getter for. 4. If you would like to allow custom fields to provide their own json_encoders, you can either: 1. use the [standard pydantic method](https://pydantic-docs.helpmanual.io/usage/exporting_models) of adding json_encoders to your model, for each field type you'd like to support: 1. This `EventedModel` class will additionally look for a `_json_encode` method on any field types in the model. If a field type declares a `_json_encode` method, it will be added to the [`json_encoders`](https://pydantic-docs.helpmanual.io/usage/exporting_models/#json_encoders) dict in the model `Config`. (Prefer using the standard pydantic method) Examples -------- Standard EventedModel example: ```python class MyModel(EventedModel): x: int = 1 m = MyModel() m.events.x.connect(lambda v: print(f"new value is {v}")) m.x = 3 # prints 'new value is 3' ``` An example of using property_setters and emitting signals when a field dependency is mutated. ```python class MyModel(EventedModel): a: int = 1 b: int = 1 @property def c(self) -> List[int]: return [self.a, self.b] @c.setter def c(self, val: Sequence[int]) -> None: self.a, self.b = val class Config: allow_property_setters = True field_dependencies = {"c": ["a", "b"]} m = MyModel() assert m.c == [1, 1] m.events.c.connect(lambda v: print(f"c updated to {v}")) m.a = 2 # prints 'c updated to [2, 1]' ``` """ # add private attributes for event emission _events: ClassVar[SignalGroup] = PrivateAttr() # mapping of name -> property obj for methods that are property setters __property_setters__: ClassVar[dict[str, property]] # mapping of field name -> dependent set of property names # when field is changed, an event for dependent properties will be emitted. __field_dependents__: ClassVar[dict[str, set[str]]] __eq_operators__: ClassVar[dict[str, "EqOperator"]] __slots__ = {"__weakref__"} __signal_group__: ClassVar[type[SignalGroup]] _changes_queue: dict[str, Any] = PrivateAttr(default_factory=dict) _primary_changes: set[str] = PrivateAttr(default_factory=set) _delay_check_semaphore: int = PrivateAttr(0) def __init__(_model_self_, **data: Any) -> None: super().__init__(**data) Group = _model_self_.__signal_group__ # the type error is "cannot assign to a class variable" ... # but if we don't use `ClassVar`, then the `dataclass_transform` decorator # will add _events: SignalGroup to the __init__ signature, for *all* user models _model_self_._events = Group(_model_self_) # type: ignore [misc] # expose the private SignalGroup publicly @property def events(self) -> SignalGroup: """Return the `SignalGroup` containing all events for this model.""" return self._events @property def _defaults(self) -> dict[str, Any]: return _get_defaults(self) def __eq__(self, other: Any) -> bool: """Check equality with another object. We override the pydantic approach (which just checks ``self.model_dump() == other.model_dump()``) to accommodate more complicated types like arrays, whose truth value is often ambiguous. ``__eq_operators__`` is constructed in ``EqualityMetaclass.__new__`` """ if not isinstance(other, EventedModel): return bool(self.model_dump() == other) for f_name, _ in self.__eq_operators__.items(): if not hasattr(self, f_name) or not hasattr(other, f_name): return False # pragma: no cover a = getattr(self, f_name) b = getattr(other, f_name) if not _check_field_equality(type(self), f_name, a, b): return False return True def update(self, values: Union["EventedModel", dict], recurse: bool = True) -> None: """Update a model in place. Parameters ---------- values : Union[dict, EventedModel] Values to update the model with. If an EventedModel is passed it is first converted to a dictionary. The keys of this dictionary must be found as attributes on the current model. recurse : bool If True, recursively update fields that are EventedModels. Otherwise, just update the immediate fields of this EventedModel, which is useful when the declared field type (e.g. ``Union``) can have different realized types with different fields. """ if isinstance(values, pydantic.BaseModel): values = values.model_dump() if not isinstance(values, dict): # pragma: no cover raise TypeError(f"values must be a dict or BaseModel. got {type(values)}") with self.events._psygnal_relay.paused(): # TODO: reduce? for key, value in values.items(): field = getattr(self, key) if isinstance(field, EventedModel) and recurse: field.update(value, recurse=recurse) else: setattr(self, key, value) def reset(self) -> None: """Reset the state of the model to default values.""" model_config = self.model_config model_fields = _get_fields(type(self)) for name, value in self._defaults.items(): if isinstance(value, EventedModel): cast("EventedModel", getattr(self, name)).reset() elif not model_config.get("frozen") and not model_fields[name].frozen: setattr(self, name, value) def _check_if_values_changed_and_emit_if_needed(self) -> None: """Check if field values changed and emit events if needed. This method is called when exiting a ComparisonDelayer context. It processes all queued changes, compares old vs new values, and emits signals for fields that actually changed. The advantage of moving this to the end of all modifications is that comparisons are performed only once for every potential change. """ # Early exit if we're still delaying comparisons or have no changes to check if self._delay_check_semaphore > 0 or len(self._changes_queue) == 0: return # ----------------- Process primary changes ------------------- # "Primary changes" are fields that were directly assigned to (as opposed # to dependent properties that might have changed as a side effect). # `_primary_changes` get added in the `_setattr_with_dependents_impl` method to_emit: list[tuple[str, Any]] = [] # list of (field name, new value) primary_changes_occurred = False for name in self._primary_changes: old_value = self._changes_queue[name] new_value = getattr(self, name) if not _check_field_equality(type(self), name, new_value, old_value): # This field actually changed value if name in self._events: # Field has a signal, queue it for emission to_emit.append((name, new_value)) else: # Field doesn't have a signal but might have dependents # that need checking primary_changes_occurred |= name in self.__field_dependents__ # Remove from queue since we've processed this primary change self._changes_queue.pop(name) # -------------------------------------------------------------- # If no primary changes occurred and no signals need emitting, # we can skip checking dependents (optimization) if not to_emit and not primary_changes_occurred: self._changes_queue.clear() self._primary_changes.clear() return # ---------- Process dependent property changes ---------- # Any remaining items in the changes queue are now # dependent properties that were queued for checking. for name, old_value in self._changes_queue.items(): new_value = getattr(self, name) if not _check_field_equality(type(self), name, new_value, old_value): to_emit.append((name, new_value)) # Clean up tracking state self._changes_queue.clear() self._primary_changes.clear() # Emit all the signals that need emitting # Use ComparisonDelayer to prevent re-entrancy issues when callbacks # modify the model again if to_emit: with ComparisonDelayer(self): for name, new_value in to_emit: getattr(self._events, name)(new_value) def __setattr__(self, name: str, value: Any) -> None: # can happen on init if name == "_events" or not hasattr(self, "_events"): # fallback to default behavior for special fields and during init return self._super_setattr_(name, value) # Check if this is a property setter first - property setters should # always go through _super_setattr_ regardless of signals/dependencies if name in self.__property_setters__: return self._super_setattr_(name, value) # Check if this field needs special handling (has signal or dependencies) is_signal_field = name in self._events has_dependents = self.__field_dependents__ and name in self.__field_dependents__ if is_signal_field or has_dependents: # For signal fields with no dependents, use the faster path if available if ( is_signal_field and not has_dependents and hasattr(self, "_setattr_no_dependants") ): self._setattr_no_dependants(name, value) else: # Use the full setattr method for fields with dependents self._setattr_default(name, value) else: # Field doesn't have signals or dependents, use fast path self._super_setattr_(name, value) def _super_setattr_(self, name: str, value: Any) -> None: # pydantic will raise a ValueError if extra fields are not allowed # so we first check to see if this field has a property.setter. # if so, we use it instead. if name in self.__property_setters__: # Wrap property setter calls in ComparisonDelayer to batch field changes with ComparisonDelayer(self): self.__property_setters__[name].fset(self, value) # type: ignore[misc] elif name == "_events": # pydantic v2 prohibits shadowing class vars, on instances object.__setattr__(self, name, value) else: super().__setattr__(name, value) def _setattr_default(self, name: str, value: Any) -> None: """Will be overwritten by metaclass __new__. It will become either `_setattr_no_dependants` (if the class has neither properties nor `__field_dependents__`), or `_setattr_with_dependents` if it does. """ def _setattr_no_dependants(self, name: str, value: Any) -> None: """Simple __setattr__ behavior when the class has no properties.""" group = self._events signal_instance: SignalInstance = group[name] if len(signal_instance) < 1: return self._super_setattr_(name, value) old_value = getattr(self, name, object()) self._super_setattr_(name, value) if not _check_field_equality(type(self), name, value, old_value): getattr(self._events, name)(value) def _setattr_with_dependents(self, name: str, value: Any) -> None: """__setattr__ behavior when the class has properties.""" with ComparisonDelayer(self): self._setattr_with_dependents_impl(name, value) def _setattr_with_dependents_impl(self, name: str, value: Any) -> None: """The "real" __setattr__ implementation inside of the comparison delayer.""" # if there are no listeners, we can just set the value without emitting # so first check if there are any listeners for this field or any of its # dependent properties. # note that ALL signals will have sat least one listener simply by nature of # being in the `self._events` SignalGroup. signal_group = self._events if name in signal_group: signal_instance: SignalInstance = signal_group[name] deps_with_callbacks = { dep_name for dep_name in self.__field_dependents__.get(name, ()) if len(signal_group[dep_name]) } if ( len(signal_instance) < 1 # the signal itself has no listeners and not deps_with_callbacks # no dependent properties with listeners and not len(signal_group._psygnal_relay) # no listeners on the group ): return self._super_setattr_(name, value) elif name in self.__field_dependents__: deps_with_callbacks = self.__field_dependents__[name] else: return self._super_setattr_(name, value) # pragma: no cover self._primary_changes.add(name) if name not in self._changes_queue: self._changes_queue[name] = getattr(self, name, object()) for dep in deps_with_callbacks: if dep not in self._changes_queue: self._changes_queue[dep] = getattr(self, dep, object()) self._super_setattr_(name, value)