learning-loop-node 0.27.2__tar.gz → 0.28.0__tar.gz

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: learning_loop_node
-Version: 0.27.2
+Version: 0.28.0
 Summary: Python Library for Nodes which connect to the Zauberzeug Learning Loop
 Author-email: Zauberzeug GmbH <info@zauberzeug.com>
 License: MIT

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/learning_loop_node/data_classes/__init__.py

@@ -30,6 +30,7 @@ from .image_metadata import ImageMetadata, ImagesMetadata
 from .socket_response import SocketResponse
 from .training import (
     Errors,
+    Hyperparameter,
     PretrainedModel,
     Training,
     TrainingError,
@@ -45,6 +46,6 @@ __all__ = [
     'AnnotationNodeStatus', 'Category', 'Context', 'DetectorStatus', 'ErrorConfiguration',
     'ModelInformation', 'NodeState', 'NodeStatus', 'ModelVersionResponse', 'ImagesMetadata',
     'SocketResponse',
-    'Errors', 'PretrainedModel', 'Training',
+    'Errors', 'Hyperparameter', 'PretrainedModel', 'Training',
     'TrainingError', 'TrainingOut', 'TrainingStateData', 'TrainingStatus',
 ]

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/learning_loop_node/data_classes/training.py

@@ -3,7 +3,7 @@ import sys
 import time
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Literal, Optional
 from uuid import uuid4
 
 from ..enums import TrainerState
@@ -22,6 +22,28 @@ class PretrainedModel():
     description: str
 
 
+HyperparameterType = Literal['int', 'float', 'bool', 'str', 'pairs']
+
+
+@dataclass(**KWONLY_SLOTS)
+class Hyperparameter:
+    """One hyperparameter a trainer accepts, reported to the loop so it can offer it in the project settings.
+
+    Plain data only, so it reaches the loop as it is. ``minimum`` and ``maximum`` apply to ``int`` and ``float``,
+    ``choices`` to ``str``. A ``pairs`` value is text of the form ``"key:value,..."``, which ``parse_pairs`` reads.
+    ``notes`` help whoever sets the value. ``expert`` ones are offered behind an expansion.
+    Whether a value can be changed when a training is started is the project's choice, not the trainer's.
+    """
+    name: str
+    type: HyperparameterType
+    default: Any
+    minimum: float | None = None
+    maximum: float | None = None
+    choices: list[str] | None = None
+    notes: str = ''
+    expert: bool = False
+
+
 @dataclass(**KWONLY_SLOTS)
 class TrainingStatus():
     id: str  # NOTE this must not be changed, but tests wont detect a change -> update tests!
@@ -33,6 +55,7 @@ class TrainingStatus():
     progress: Optional[float]
 
     pretrained_models: List[PretrainedModel] = field(default_factory=list)
+    hyperparameters: list[Hyperparameter] = field(default_factory=list)
     architecture: Optional[str] = None
     context: Optional[Context] = None
 

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/learning_loop_node/trainer/exceptions.py

@@ -15,6 +15,13 @@ class TrainingRefusedError(Exception):
     """Raised when a trainer cannot take on a training it was asked to begin; nothing was set up."""
 
 
+class HyperparameterMismatchError(Exception):
+    """Raised when a training's hyperparameters do not fit the trainer's declaration; nothing was set up.
+
+    Unlike a refusal, the training would fail on any trainer of the same declaration.
+    """
+
+
 class TrainingSlotTakenError(TrainingRefusedError):
     """Raised when a sibling node holds the training slot."""
 

learning_loop_node-0.28.0/learning_loop_node/trainer/hyperparameters.py

@@ -0,0 +1,180 @@
+"""Resolving the hyperparameters a training was sent against the ones its trainer declares.
+
+The loop sends only what the project sets, plus what the job overrides. Everything else comes from
+the trainer's declared defaults, so after :func:`resolve` a training holds the full set it runs
+with, and that is what gets reported back to the loop.
+"""
+
+import contextlib
+import logging
+import math
+from collections.abc import Mapping, Sequence
+from typing import Any, get_args
+
+from ..data_classes import Hyperparameter
+from ..data_classes.training import HyperparameterType
+from .batch_size import REQUESTED_BATCH_SIZE
+from .exceptions import CriticalError
+from .validation_set import COMPUTE_VALIDATION_SET
+
+logger = logging.getLogger(__name__)
+
+LIBRARY_NAMES = (REQUESTED_BATCH_SIZE, COMPUTE_VALIDATION_SET)
+"""Hyperparameters the library reads itself, so a trainer that does not declare them still uses them."""
+
+BOOL_STRINGS = {'true': True, '1': True, 'yes': True, 'on': True,
+                'false': False, '0': False, 'no': False, 'off': False}
+
+
+def resolve(declared: Sequence[Hyperparameter], received: Mapping[str, Any]) -> dict[str, Any]:
+    """Coerce and check ``received`` against ``declared`` and fill in the declared defaults.
+
+    ``declared`` is trusted to have passed :func:`check_declaration`. Names the trainer does not declare are kept
+    unchanged: old project configurations still carry them, and a resolved training carries the values its trainer
+    reported. An unset hyperparameter with a default of ``None`` is left out, so the trainer's own fallback applies.
+    Resolving twice gives the same result.
+
+    :raises CriticalError: If a value cannot be coerced to its type, is out of range or not one of its choices.
+    """
+    resolved = dict(received)
+    for hyperparameter in declared:
+        value = received.get(hyperparameter.name)
+        if _is_unset(hyperparameter, value):
+            if hyperparameter.default is None:
+                resolved.pop(hyperparameter.name, None)
+            else:
+                resolved[hyperparameter.name] = _coerce(hyperparameter, hyperparameter.default)
+        else:
+            resolved[hyperparameter.name] = _check(hyperparameter, _coerce(hyperparameter, value))
+    return resolved
+
+
+def check_declaration(declared: Sequence[Hyperparameter]) -> None:
+    """Check that ``declared`` is consistent, so that :func:`resolve` can rely on it.
+
+    A trainer's declaration does not change while it runs, so the node checks it once, when it is built. A default of
+    ``None`` stands for "the trainer decides" and is not checked.
+
+    :raises ValueError: If a name is declared twice, a hyperparameter contradicts itself or its default does not fit.
+    """
+    names = [h.name for h in declared]
+    if duplicates := sorted({name for name in names if names.count(name) > 1}):
+        raise ValueError(f'the trainer declares these hyperparameters twice: {", ".join(duplicates)}')
+    for hyperparameter in declared:
+        _check_one_declaration(hyperparameter)
+
+
+def parse_pairs(text: str) -> list[tuple[str, str]]:
+    """The ``key:value`` items of a ``pairs`` hyperparameter such as ``"uuid:uuid,uuid:uuid"``.
+
+    Empty items and the spaces around keys and values are skipped, so ``''`` holds no pairs.
+
+    :raises ValueError: If an item is not a key and a value separated by a single colon.
+    """
+    pairs = []
+    for item in (i.strip() for i in text.split(',')):
+        if not item:
+            continue
+        key, _, value = (part.strip() for part in item.partition(':'))
+        if not key or not value or ':' in value:
+            raise ValueError(f'"{item}" is not of the form "key:value"')
+        pairs.append((key, value))
+    return pairs
+
+
+def warn_about_undeclared(declared: Sequence[Hyperparameter], received: Mapping[str, Any]) -> None:
+    """Log the names in ``received`` that neither the trainer nor the library knows; they are not rejected."""
+    unknown = sorted(set(received) - {h.name for h in declared} - set(LIBRARY_NAMES))
+    if unknown:
+        logger.warning('these hyperparameters are not declared by this trainer: %s', ', '.join(unknown))
+
+
+def _check_one_declaration(hyperparameter: Hyperparameter) -> None:
+    def contradiction(reason: str) -> ValueError:
+        return ValueError(f"the trainer declares hyperparameter '{hyperparameter.name}' {reason}")
+
+    if hyperparameter.type not in get_args(HyperparameterType):
+        raise contradiction(f'with the unknown type {hyperparameter.type!r}')
+    if hyperparameter.type not in ('int', 'float') and (hyperparameter.minimum is not None or
+                                                        hyperparameter.maximum is not None):
+        raise contradiction(f'as {hyperparameter.type} with a minimum or maximum')
+    if hyperparameter.type != 'str' and hyperparameter.choices:
+        raise contradiction(f'as {hyperparameter.type} with choices')
+    for bound in (hyperparameter.minimum, hyperparameter.maximum):
+        if bound is not None and (isinstance(bound, bool) or not isinstance(bound, (int, float))):
+            raise contradiction(f'with the bound {bound!r}, which is not a number')
+    if (hyperparameter.minimum is not None and hyperparameter.maximum is not None and
+            hyperparameter.minimum > hyperparameter.maximum):
+        raise contradiction('with a minimum above its maximum')
+    if hyperparameter.default is not None:
+        try:
+            _check(hyperparameter, _coerce(hyperparameter, hyperparameter.default))
+        except CriticalError as e:
+            raise contradiction(f'with a default that does not fit: {e}') from None
+
+
+def _is_unset(hyperparameter: Hyperparameter, value: Any) -> bool:
+    """``None``, or the empty value the loop sends for a field left blank, unless it fits as a value of its own."""
+    if value is None:
+        return True
+    if value != '':
+        return False
+    try:
+        _check(hyperparameter, _coerce(hyperparameter, value))
+    except CriticalError:
+        return True
+    return False
+
+
+def _invalid(hyperparameter: Hyperparameter, requirement: str, value: Any) -> CriticalError:
+    return CriticalError(f"hyperparameter '{hyperparameter.name}' {requirement}, got {value!r}")
+
+
+def _coerce(hyperparameter: Hyperparameter, value: Any) -> Any:
+    if hyperparameter.type == 'pairs':
+        if not isinstance(value, str):
+            raise _invalid(hyperparameter, 'must be text of key:value pairs', value)
+        try:
+            return ','.join(f'{key}:{item}' for key, item in parse_pairs(value))
+        except ValueError as e:
+            raise CriticalError(f"hyperparameter '{hyperparameter.name}': {e}") from None
+    if hyperparameter.type == 'str':
+        if isinstance(value, bool) or not isinstance(value, (str, int, float)):
+            raise _invalid(hyperparameter, 'must be text', value)
+        return str(value)
+    if hyperparameter.type == 'bool':
+        if isinstance(value, bool):
+            return value
+        if isinstance(value, str) and (word := value.strip().lower()) in BOOL_STRINGS:
+            return BOOL_STRINGS[word]
+        with contextlib.suppress(TypeError, ValueError):
+            number = float(value)
+            if math.isfinite(number):
+                return number != 0
+        raise _invalid(hyperparameter, 'must be true or false', value)
+    if isinstance(value, bool):
+        raise _invalid(hyperparameter, 'must be a number', value)
+    if hyperparameter.type == 'int' and isinstance(value, (int, str)):
+        with contextlib.suppress(ValueError):
+            return int(value)
+    try:
+        number = float(value)
+    except (TypeError, ValueError):
+        raise _invalid(hyperparameter, 'must be a number', value) from None
+    if not math.isfinite(number):
+        raise _invalid(hyperparameter, 'must be finite', value)
+    if hyperparameter.type == 'float':
+        return number
+    if not number.is_integer():
+        raise _invalid(hyperparameter, 'must be a whole number', value)
+    return int(number)
+
+
+def _check(hyperparameter: Hyperparameter, value: Any) -> Any:
+    if hyperparameter.minimum is not None and value < hyperparameter.minimum:
+        raise _invalid(hyperparameter, f'must be at least {hyperparameter.minimum}', value)
+    if hyperparameter.maximum is not None and value > hyperparameter.maximum:
+        raise _invalid(hyperparameter, f'must be at most {hyperparameter.maximum}', value)
+    if hyperparameter.choices and value not in hyperparameter.choices:
+        raise _invalid(hyperparameter, f"must be one of {', '.join(hyperparameter.choices)}", value)
+    return value

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/learning_loop_node/trainer/trainer_logic_generic.py

@@ -13,6 +13,7 @@ from fastapi.encoders import jsonable_encoder
 from ..data_classes import (
     Context,
     Errors,
+    Hyperparameter,
     PretrainedModel,
     Training,
     TrainingOut,
@@ -26,7 +27,8 @@ from ..helpers.misc import (
     is_valid_uuid4,
 )
 from .downloader import TrainingsDownloader
-from .exceptions import CriticalError, NodeNeedsRestartError, TrainingRefusedError
+from .exceptions import CriticalError, HyperparameterMismatchError, NodeNeedsRestartError, TrainingRefusedError
+from .hyperparameters import resolve, warn_about_undeclared
 from .io_helpers import ActiveTrainingIO, EnvironmentVars, LastTrainingIO
 from .training_slot import TrainingSlot, training_slot_from_env
 from .validation_set import assign_validation_set
@@ -186,6 +188,12 @@ class TrainerLogicGeneric(ABC):
         """
         raise NotImplementedError
 
+    @property
+    @abstractmethod
+    def provided_hyperparameters(self) -> list[Hyperparameter]:
+        """Returns the hyperparameters this trainer accepts; an empty list leaves them unchecked."""
+        raise NotImplementedError
+
     # ---------------------------------------- METHODS ----------------------------------------
 
     # NOTE: Trainings are started by the Learning Loop via the begin_training event
@@ -204,6 +212,7 @@ class TrainerLogicGeneric(ABC):
                                 progress=self.general_progress)
 
         status.pretrained_models = self.provided_pretrained_models
+        status.hyperparameters = self.provided_hyperparameters
         status.architecture = self.model_architecture
 
         if self._training:
@@ -224,9 +233,18 @@ class TrainerLogicGeneric(ABC):
 
     def _init_from_last_training(self) -> None:
         """Initializes a new training object from the last training saved on disc via last_training_io.
+
+        The hyperparameters are resolved again, since the training may have been saved before this trainer declared
+        them. If one no longer fits, the training continues with the values it was saved with, since it already runs
+        with them.
         """
         self._training = self.last_training_io.load()
         assert self._training is not None and self._training.training_folder is not None, 'could not restore training folder'
+        try:
+            self._training.hyperparameters = resolve(self.provided_hyperparameters, self._training.hyperparameters)
+        except CriticalError as e:
+            logger.warning('restored training does not fit the declared hyperparameters, continuing with the saved '
+                           'values: %s', e)
         logger.info('restored training: \n%s', self._training)
         self._active_training_io = ActiveTrainingIO(
             self._training.training_folder, self.node.loop_communicator, self._training.context)
@@ -234,13 +252,15 @@ class TrainerLogicGeneric(ABC):
     async def begin_training(self, organization: str, project: str, training_config: Dict) -> None:
         """Called on `begin_training` event from the Learning Loop.
 
-        The training slot is taken before anything is set up, so a refused training leaves the
-        previous training folders in place.
+        The hyperparameters are resolved and the training slot is taken before anything is set up, so a refused
+        training leaves the previous training folders in place.
 
         :raise TrainingRefusedError: if a training is active or a sibling node holds the training slot.
+        :raise HyperparameterMismatchError: if a hyperparameter does not fit its declaration.
         """
         if self.training_active:
             raise TrainingRefusedError(f'training {self.training.id} is still active')
+        training_config = self._with_resolved_hyperparameters(training_config)
         context = Context(organization=organization, project=project)
         self.training_slot.claim(self._slot_claim(context, training_config.get('training_number')))
         try:
@@ -250,6 +270,24 @@ class TrainerLogicGeneric(ABC):
             raise
         self._begin_training_task()
 
+    def _with_resolved_hyperparameters(self, training_config: dict) -> dict:
+        """Returns ``training_config`` with its hyperparameters resolved against the declared ones.
+
+        A config without a hyperparameters dict is returned as it is, for ``Training.generate_training`` to reject.
+
+        :raise HyperparameterMismatchError: if a hyperparameter does not fit its declaration.
+        """
+        declared = self.provided_hyperparameters
+        received = training_config.get('hyperparameters')
+        if not declared or not isinstance(received, dict):
+            return training_config
+        try:
+            resolved = resolve(declared, received)
+        except CriticalError as e:
+            raise HyperparameterMismatchError(str(e)) from e
+        warn_about_undeclared(declared, received)
+        return {**training_config, 'hyperparameters': resolved}
+
     def _begin_training_task(self) -> None:
         # NOTE: Task object is used to potentially cancel the task
         self.training_task = asyncio.get_event_loop().create_task(self._run())

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/learning_loop_node/trainer/trainer_node.py

@@ -10,7 +10,8 @@ from socketio import AsyncClient, exceptions
 from ..data_classes import SocketResponse
 from ..enums import TrainerState
 from ..node import Node
-from .exceptions import TrainingRefusedError
+from .exceptions import HyperparameterMismatchError, TrainingRefusedError
+from .hyperparameters import check_declaration
 from .io_helpers import LastTrainingIO
 from .rest import backdoor_controls
 from .trainer_logic_generic import TrainerLogicGeneric
@@ -19,6 +20,8 @@ from .trainer_logic_generic import TrainerLogicGeneric
 class TrainerNode(Node):
 
     def __init__(self, name: str, trainer_logic: TrainerLogicGeneric, uuid: Optional[str] = None, use_backdoor_controls: bool = False):
+        """:raise ValueError: if the trainer declares hyperparameters that contradict themselves."""
+        check_declaration(trainer_logic.provided_hyperparameters)
         super().__init__(name, uuid=uuid, node_type='trainer')
         trainer_logic._node = self
         self.trainer_logic = trainer_logic
@@ -72,6 +75,10 @@ class TrainerNode(Node):
             except TrainingRefusedError as e:
                 self.log.warning('refused begin_training: %s', e)
                 return SocketResponse(success=False, error_msg=str(e), payload='refused')
+            except HyperparameterMismatchError as e:
+                # NOTE: not a refusal, which the loop answers by offering the same job again
+                self.log.warning('rejected begin_training: %s', e)
+                return SocketResponse.for_failure(str(e))
             return True
 
         @sio_client.event

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/learning_loop_node/trainer/validation_set.py

@@ -21,6 +21,7 @@ import logging
 from collections.abc import Mapping
 from typing import Any
 
+from ..data_classes import Hyperparameter
 from .exceptions import CriticalError
 
 logger = logging.getLogger(__name__)
@@ -28,6 +29,12 @@ logger = logging.getLogger(__name__)
 COMPUTE_VALIDATION_SET = 'compute_validation_set'
 """The hyperparameter that asks to split a validation set off the train images when the loop sent none."""
 
+COMPUTE_VALIDATION_SET_HYPERPARAMETER = Hyperparameter(
+    name=COMPUTE_VALIDATION_SET, type='bool', default=False,
+    notes='Split about a tenth of the train images off to select the models on, so the test set is only reported on; '
+          'a validation set sent by the loop is used either way')
+"""The declaration of `compute_validation_set`, for a trainer to list among its `provided_hyperparameters`."""
+
 VALIDATION_FRACTION = 0.1
 
 

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/learning_loop_node.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: learning_loop_node
-Version: 0.27.2
+Version: 0.28.0
 Summary: Python Library for Nodes which connect to the Zauberzeug Learning Loop
 Author-email: Zauberzeug GmbH <info@zauberzeug.com>
 License: MIT

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/learning_loop_node.egg-info/SOURCES.txt

@@ -61,6 +61,7 @@ learning_loop_node/trainer/cuda.py
 learning_loop_node/trainer/downloader.py
 learning_loop_node/trainer/exceptions.py
 learning_loop_node/trainer/executor.py
+learning_loop_node/trainer/hyperparameters.py
 learning_loop_node/trainer/io_helpers.py
 learning_loop_node/trainer/metrics.py
 learning_loop_node/trainer/subprocess.py

{learning_loop_node-0.27.2 → learning_loop_node-0.28.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "learning_loop_node"
-version = "0.27.2"
+version = "0.28.0"
 description = "Python Library for Nodes which connect to the Zauberzeug Learning Loop"
 authors = [{ name = "Zauberzeug GmbH", email = "info@zauberzeug.com" }]
 license = { text = "MIT" }