congrads 1.1.2__py3-none-any.whl → 1.2.0__py3-none-any.whl

congrads/__init__.py

@@ -9,20 +9,3 @@ try:
     version = get_version("congrads")  # Replace with your package name
 except Exception:
     version = "0.0.0"  # Fallback if the package isn't installed
-
-# Only expose the submodules, not individual classes
-from . import constraints, core, datasets, descriptor, metrics, networks, utils
-
-# Define __all__ to specify that the submodules are accessible,
-# but not classes directly.
-__all__ = [
-    "checkpoints",
-    "constraints",
-    "core",
-    "datasets",
-    "descriptor",
-    "metrics",
-    "networks",
-    "transformations",
-    "utils",
-]

congrads/callbacks/base.py

@@ -0,0 +1,357 @@
+"""Callback and Operation Framework for Modular Training Pipelines.
+
+This module provides a structured system for defining and executing
+callbacks and operations at different stages of a training lifecycle.
+It is designed to support:
+
+- Stateless, reusable operations that produce outputs merged into
+  the event-local data.
+- Callbacks that group operations and/or custom logic for specific
+  stages of training, epochs, batches, and steps.
+- A central CallbackManager to orchestrate multiple callbacks,
+  maintain shared context, and execute stage-specific pipelines
+  in deterministic order.
+
+Stages supported:
+    - on_train_start
+    - on_train_end
+    - on_epoch_start
+    - on_epoch_end
+    - on_batch_start
+    - on_batch_end
+    - on_test_start
+    - on_test_end
+    - on_train_batch_start
+    - on_train_batch_end
+    - on_valid_batch_start
+    - on_valid_batch_end
+    - on_test_batch_start
+    - on_test_batch_end
+    - after_train_forward
+    - after_valid_forward
+    - after_test_forward
+
+Usage:
+    1. Define Operations by subclassing `Operation` and implementing
+       the `compute` method.
+    2. Create a Callback subclass or instance and register Operations
+       to stages via `add(stage, operation)`.
+    3. Register callbacks with `CallbackManager`.
+    4. Invoke `CallbackManager.run(stage, data)` at appropriate points
+       in the training loop, passing in event-local data.
+"""
+
+from abc import ABC, abstractmethod
+from collections.abc import Iterable
+from typing import Any, Literal, Self
+
+Stage = Literal[
+    "on_train_start",
+    "on_train_end",
+    "on_epoch_start",
+    "on_epoch_end",
+    "on_test_start",
+    "on_test_end",
+    "on_batch_start",
+    "on_batch_end",
+    "on_train_batch_start",
+    "on_train_batch_end",
+    "on_valid_batch_start",
+    "on_valid_batch_end",
+    "on_test_batch_start",
+    "on_test_batch_end",
+    "after_train_forward",
+    "after_valid_forward",
+    "after_test_forward",
+]
+
+STAGES: tuple[Stage, ...] = (
+    "on_train_start",
+    "on_train_end",
+    "on_epoch_start",
+    "on_epoch_end",
+    "on_test_start",
+    "on_test_end",
+    "on_batch_start",
+    "on_batch_end",
+    "on_train_batch_start",
+    "on_train_batch_end",
+    "on_valid_batch_start",
+    "on_valid_batch_end",
+    "on_test_batch_start",
+    "on_test_batch_end",
+    "after_train_forward",
+    "after_valid_forward",
+    "after_test_forward",
+)
+
+
+class Operation(ABC):
+    """Abstract base class representing a stateless unit of work executed inside a callback stage.
+
+    Subclasses should implement the `compute` method which returns
+    a dictionary of outputs to merge into the running event data.
+    """
+
+    def __repr__(self) -> str:
+        """Return a concise string representation of the operation."""
+        return f"<{self.__class__.__name__}>"
+
+    def __call__(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute the operation with the given event-local data and shared context.
+
+        Args:
+            data (dict[str, Any]): Event-local dictionary containing data for this stage.
+            ctx (dict[str, Any]): Shared context dictionary accessible by all operations and callbacks.
+
+        Returns:
+            dict[str, Any]: Outputs produced by the operation to merge into the running data.
+                            Returns an empty dict if `compute` returns None.
+        """
+        out = self.compute(data, ctx)
+        if out is None:
+            return {}
+        if not isinstance(out, dict):
+            raise TypeError(f"{self.__class__.__name__}.compute must return dict or None")
+        return out
+
+    @abstractmethod
+    def compute(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any] | None:
+        """Perform the operation's computation.
+
+        Args:
+            data (dict[str, Any]): Event-local dictionary containing the current data.
+            ctx (dict[str, Any]): Shared context dictionary.
+
+        Returns:
+            dict[str, Any] or None: Outputs to merge into the running data.
+                                     Returning None is equivalent to {}.
+        """
+        raise NotImplementedError
+
+
+class Callback(ABC):  # noqa: B024
+    """Abstract base class representing a callback that can have multiple operations registered to different stages of the training lifecycle.
+
+    Each stage method executes all operations registered for that stage
+    in insertion order. Operations can modify the event-local data dictionary.
+    """
+
+    def __init__(self):
+        """Initialize the callback with empty operation lists for all stages."""
+        self._ops_by_stage: dict[Stage, list[Operation]] = {s: [] for s in STAGES}
+
+    def __repr__(self) -> str:
+        """Return a concise string showing number of operations per stage."""
+        ops_summary = {stage: len(ops) for stage, ops in self._ops_by_stage.items() if ops}
+        return f"<{self.__class__.__name__} ops={ops_summary}>"
+
+    def add(self, stage: Stage, op: Operation) -> Self:
+        """Register an operation to execute at the given stage.
+
+        Args:
+            stage (Stage): Lifecycle stage at which to run the operation.
+            op (Operation): Operation instance to add.
+
+        Returns:
+            Self: Returns self for method chaining.
+        """
+        self._ops_by_stage[stage].append(op)
+        return self
+
+    def _run_ops(self, stage: Stage, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute all operations registered for a specific stage.
+
+        Args:
+            stage (Stage): Lifecycle stage to execute.
+            data (dict[str, Any]): Event-local data to pass to operations.
+            ctx (dict[str, Any]): Shared context across callbacks and operations.
+
+        Returns:
+            dict[str, Any]: Merged data dictionary after executing all operations.
+
+        Notes:
+            - Operations are executed in insertion order.
+            - If an operation overwrites existing keys, a warning is issued.
+        """
+        out = dict(data)
+
+        for operation in self._ops_by_stage[stage]:
+            try:
+                produced = operation(out, ctx) or {}
+            except Exception as e:
+                raise RuntimeError(f"Error in operation {operation} at stage {stage}") from e
+
+            collisions = set(produced.keys()) & set(out.keys())
+            if collisions:
+                import warnings
+
+                warnings.warn(
+                    f"Operation {operation} at stage '{stage}' is overwriting keys: {collisions}",
+                    stacklevel=2,
+                )
+
+            out.update(produced)
+
+        return out
+
+    # --- training ---
+    def on_train_start(self, data: dict[str, Any], ctx: dict[str, Any]):
+        """Execute operations registered for the 'on_train_start' stage."""
+        self._run_ops("on_train_start", data, ctx)
+
+    def on_train_end(self, data: dict[str, Any], ctx: dict[str, Any]):
+        """Execute operations registered for the 'on_train_end' stage."""
+        self._run_ops("on_train_end", data, ctx)
+
+    # --- epoch ---
+    def on_epoch_start(self, data: dict[str, Any], ctx: dict[str, Any]):
+        """Execute operations registered for the 'on_epoch_start' stage."""
+        self._run_ops("on_epoch_start", data, ctx)
+
+    def on_epoch_end(self, data: dict[str, Any], ctx: dict[str, Any]):
+        """Execute operations registered for the 'on_epoch_end' stage."""
+        self._run_ops("on_epoch_end", data, ctx)
+
+    # --- test ---
+    def on_test_start(self, data: dict[str, Any], ctx: dict[str, Any]):
+        """Execute operations registered for the 'on_test_start' stage."""
+        self._run_ops("on_test_start", data, ctx)
+
+    def on_test_end(self, data: dict[str, Any], ctx: dict[str, Any]):
+        """Execute operations registered for the 'on_test_end' stage."""
+        self._run_ops("on_test_end", data, ctx)
+
+    # --- batch ---
+    def on_batch_start(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'on_batch_start' stage."""
+        return self._run_ops("on_batch_start", data, ctx)
+
+    def on_batch_end(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'on_batch_end' stage."""
+        return self._run_ops("on_batch_end", data, ctx)
+
+    def on_train_batch_start(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'on_train_batch_start' stage."""
+        return self._run_ops("on_train_batch_start", data, ctx)
+
+    def on_train_batch_end(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'on_train_batch_end' stage."""
+        return self._run_ops("on_train_batch_end", data, ctx)
+
+    def on_valid_batch_start(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'on_valid_batch_start' stage."""
+        return self._run_ops("on_valid_batch_start", data, ctx)
+
+    def on_valid_batch_end(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'on_valid_batch_end' stage."""
+        return self._run_ops("on_valid_batch_end", data, ctx)
+
+    def on_test_batch_start(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'on_test_batch_start' stage."""
+        return self._run_ops("on_test_batch_start", data, ctx)
+
+    def on_test_batch_end(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'on_test_batch_end' stage."""
+        return self._run_ops("on_test_batch_end", data, ctx)
+
+    def after_train_forward(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'after_train_forward' stage."""
+        return self._run_ops("after_train_forward", data, ctx)
+
+    def after_valid_forward(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'after_valid_forward' stage."""
+        return self._run_ops("after_valid_forward", data, ctx)
+
+    def after_test_forward(self, data: dict[str, Any], ctx: dict[str, Any]) -> dict[str, Any]:
+        """Execute operations registered for the 'after_test_forward' stage."""
+        return self._run_ops("after_test_forward", data, ctx)
+
+
+class CallbackManager:
+    """Orchestrates multiple callbacks and executes them at specific lifecycle stages.
+
+    - Callbacks are executed in registration order.
+    - Event-local data flows through all callbacks.
+    - Shared context is available for cross-callback communication.
+    """
+
+    def __init__(self, callbacks: Iterable[Callback] | None = None):
+        """Initialize a CallbackManager instance.
+
+        Args:
+        callbacks (Iterable[Callback] | None): Optional initial callbacks to register.
+            If None, starts with an empty callback list.
+
+        Attributes:
+        _callbacks (list[Callback]): Internal list of registered callbacks.
+        ctx (dict[str, Any]): Shared context dictionary accessible to all callbacks
+            and operations for cross-event communication.
+        """
+        self._callbacks: list[Callback] = list(callbacks) if callbacks else []
+        self.ctx: dict[str, Any] = {}
+
+    def __repr__(self) -> str:
+        """Return a concise representation showing registered callbacks and ctx keys."""
+        names = [cb.__class__.__name__ for cb in self._callbacks]
+        return f"<CallbackManager callbacks={names} ctx_keys={list(self.ctx.keys())}>"
+
+    def add(self, callback: Callback) -> Self:
+        """Register a single callback.
+
+        Args:
+            callback (Callback): Callback instance to add.
+
+        Returns:
+            Self: Returns self for fluent chaining.
+        """
+        self._callbacks.append(callback)
+        return self
+
+    def extend(self, callbacks: Iterable[Callback]) -> None:
+        """Register multiple callbacks at once.
+
+        Args:
+            callbacks (Iterable[Callback]): Iterable of callbacks to add.
+        """
+        self._callbacks.extend(callbacks)
+
+    def run(self, stage: Stage, data: dict[str, Any]) -> dict[str, Any]:
+        """Execute all registered callbacks for a specific stage.
+
+        Args:
+            stage (Stage): Lifecycle stage to run (e.g., "on_batch_start").
+            data (dict[str, Any]): Event-local data dictionary to pass through callbacks.
+
+        Returns:
+            dict[str, Any]: The final merged data dictionary after executing all callbacks.
+
+        Raises:
+            ValueError: If a callback does not implement the requested stage.
+            RuntimeError: If any callback raises an exception during execution.
+        """
+        for cb in self._callbacks:
+            if not hasattr(cb, stage):
+                raise ValueError(
+                    f"Callback {cb.__class__.__name__} has no handler for stage {stage}"
+                )
+            handler = getattr(cb, stage)
+
+            try:
+                new_data = handler(data, self.ctx)
+                if new_data is not None:
+                    data = new_data
+
+            except Exception as e:
+                raise RuntimeError(f"Error in callback {cb.__class__.__name__}.{stage}") from e
+
+        return data
+
+    @property
+    def callbacks(self) -> tuple[Callback, ...]:
+        """Return a read-only tuple of registered callbacks.
+
+        Returns:
+            tuple[Callback, ...]: Registered callbacks.
+        """
+        return tuple(self._callbacks)

congrads/callbacks/registry.py

@@ -0,0 +1,106 @@
+"""Holds all callback implementations for use in the training workflow.
+
+This module acts as a central registry for defining and storing different
+callback classes, such as logging, checkpointing, or custom behaviors
+triggered during training, validation, or testing. It is intended to
+collect all callback implementations in one place for easy reference
+and import, and can be extended as new callbacks are added.
+"""
+
+from torch.utils.tensorboard import SummaryWriter
+
+from ..callbacks.base import Callback
+from ..metrics import MetricManager
+from ..utils.utility import CSVLogger
+
+
+class LoggerCallback(Callback):
+    """Callback to log metrics to TensorBoard and CSV after each epoch or test.
+
+    This callback queries a MetricManager for aggregated metrics, writes them
+    to TensorBoard using SummaryWriter, and logs them to a CSV file via CSVLogger.
+    It also flushes loggers and resets metrics after logging.
+
+    Methods implemented:
+    - on_epoch_end: Logs metrics at the end of a training epoch.
+    - on_test_end: Logs metrics at the end of testing.
+    """
+
+    def __init__(
+        self,
+        metric_manager: MetricManager,
+        tensorboard_logger: SummaryWriter,
+        csv_logger: CSVLogger,
+    ):
+        """Initialize the LoggerCallback.
+
+        Args:
+            metric_manager: Instance of MetricManager used to collect metrics.
+            tensorboard_logger: TensorBoard SummaryWriter instance for logging scalars.
+            csv_logger: CSVLogger instance for logging metrics to CSV files.
+        """
+        super().__init__()
+        self.metric_manager = metric_manager
+        self.tensorboard_logger = tensorboard_logger
+        self.csv_logger = csv_logger
+
+    def on_epoch_end(self, data: dict[str, any], ctx: dict[str, any]):
+        """Log training metrics at the end of an epoch.
+
+        Aggregates metrics from the MetricManager under the 'during_training' category,
+        writes them to TensorBoard and CSV, flushes the loggers, and resets the metrics
+        for the next epoch.
+
+        Args:
+            data: Dictionary containing batch/epoch context (must include 'epoch').
+            ctx: Additional context dictionary (unused in this implementation).
+
+        Returns:
+            data: The same input dictionary, unmodified.
+        """
+        epoch = data["epoch"]
+
+        # Log training metrics
+        metrics = self.metric_manager.aggregate("during_training")
+        for name, value in metrics.items():
+            self.tensorboard_logger.add_scalar(name, value.item(), epoch)
+            self.csv_logger.add_value(name, value.item(), epoch)
+
+        # Flush/save
+        self.tensorboard_logger.flush()
+        self.csv_logger.save()
+
+        # Reset metric manager for training
+        self.metric_manager.reset("during_training")
+
+        return data
+
+    def on_test_end(self, data: dict[str, any], ctx: dict[str, any]):
+        """Log test metrics at the end of testing.
+
+        Aggregates metrics from the MetricManager under the 'after_training' category,
+        writes them to TensorBoard and CSV, flushes the loggers, and resets the metrics.
+
+        Args:
+            data: Dictionary containing test context (must include 'epoch').
+            ctx: Additional context dictionary (unused in this implementation).
+
+        Returns:
+            data: The same input dictionary, unmodified.
+        """
+        epoch = data["epoch"]
+
+        # Log test metrics
+        metrics = self.metric_manager.aggregate("after_training")
+        for name, value in metrics.items():
+            self.tensorboard_logger.add_scalar(name, value.item(), epoch)
+            self.csv_logger.add_value(name, value.item(), epoch)
+
+        # Flush/save
+        self.tensorboard_logger.flush()
+        self.csv_logger.save()
+
+        # Reset metric manager for test
+        self.metric_manager.reset("after_training")
+
+        return data

congrads/checkpoints.py

@@ -14,7 +14,7 @@ from torch.nn import Module
 from torch.optim import Optimizer
 
 from .metrics import MetricManager
-from .utils import validate_callable, validate_type
+from .utils.validation import validate_callable, validate_type
 
 
 class CheckpointManager:

congrads/constraints/base.py

@@ -0,0 +1,174 @@
+"""Defines the abstract base class `Constraint` for specifying constraints on neural network outputs.
+
+A `Constraint` monitors whether the network predictions satisfy certain
+conditions during training, validation, and testing. It can optionally
+adjust the loss to enforce constraints, and logs the relevant metrics.
+
+Responsibilities:
+- Track which network layers/tags the constraint applies to
+- Check constraint satisfaction for a batch of predictions
+- Compute adjustment directions to enforce the constraint
+- Provide a rescale factor and enforcement flag to influence loss adjustment
+
+Subclasses must implement the abstract methods:
+- `check_constraint(data)`: Evaluate constraint satisfaction for a batch
+- `calculate_direction(data)`: Compute directions to adjust predictions
+"""
+
+import random
+import string
+import warnings
+from abc import ABC, abstractmethod
+from numbers import Number
+
+from torch import Tensor
+
+from congrads.descriptor import Descriptor
+from congrads.utils.validation import validate_iterable, validate_type
+
+
+class Constraint(ABC):
+    """Abstract base class for defining constraints applied to neural networks.
+
+    A `Constraint` specifies conditions that the neural network outputs
+    should satisfy. It supports monitoring constraint satisfaction
+    during training and can adjust loss to enforce constraints. Subclasses
+    must implement the `check_constraint` and `calculate_direction` methods.
+
+    Args:
+        tags (set[str]): Tags referencing parts of the network where this constraint applies to.
+        name (str, optional): A unique name for the constraint. If not provided,
+            a name is generated based on the class name and a random suffix.
+        enforce (bool, optional): If False, only monitor the constraint
+            without adjusting the loss. Defaults to True.
+        rescale_factor (Number, optional): Factor to scale the
+            constraint-adjusted loss. Defaults to 1.5. Should be greater
+            than 1 to give weight to the constraint.
+
+    Raises:
+        TypeError: If a provided attribute has an incompatible type.
+        ValueError: If any tag in `tags` is not
+            defined in the `descriptor`.
+
+    Note:
+        - If `rescale_factor <= 1`, a warning is issued.
+        - If `name` is not provided, a name is auto-generated,
+          and a warning is logged.
+
+    """
+
+    descriptor: Descriptor = None
+    device = None
+
+    def __init__(
+        self, tags: set[str], name: str = None, enforce: bool = True, rescale_factor: Number = 1.5
+    ) -> None:
+        """Initializes a new Constraint instance.
+
+        Args:
+            tags (set[str]): Tags referencing parts of the network where this constraint applies to.
+            name (str, optional): A unique name for the constraint. If not
+                provided, a name is generated based on the class name and a
+                random suffix.
+            enforce (bool, optional): If False, only monitor the constraint
+                without adjusting the loss. Defaults to True.
+            rescale_factor (Number, optional): Factor to scale the
+                constraint-adjusted loss. Defaults to 1.5. Should be greater
+                than 1 to give weight to the constraint.
+
+        Raises:
+            TypeError: If a provided attribute has an incompatible type.
+            ValueError: If any tag in `tags` is not defined in the `descriptor`.
+
+        Note:
+            - If `rescale_factor <= 1`, a warning is issued.
+            - If `name` is not provided, a name is auto-generated, and a
+            warning is logged.
+        """
+        # Init parent class
+        super().__init__()
+
+        # Type checking
+        validate_iterable("tags", tags, str)
+        validate_type("name", name, str, allow_none=True)
+        validate_type("enforce", enforce, bool)
+        validate_type("rescale_factor", rescale_factor, Number)
+
+        # Init object variables
+        self.tags = tags
+        self.rescale_factor = rescale_factor
+        self.initial_rescale_factor = rescale_factor
+        self.enforce = enforce
+
+        # Perform checks
+        if rescale_factor <= 1:
+            warnings.warn(
+                f"Rescale factor for constraint {name} is <= 1. The network "
+                "will favor general loss over the constraint-adjusted loss. "
+                "Is this intended behavior? Normally, the rescale factor "
+                "should always be larger than 1.",
+                stacklevel=2,
+            )
+
+        # If no constraint_name is set, generate one based
+        # on the class name and a random suffix
+        if name:
+            self.name = name
+        else:
+            random_suffix = "".join(random.choices(string.ascii_uppercase + string.digits, k=6))
+            self.name = f"{self.__class__.__name__}_{random_suffix}"
+            warnings.warn(f"Name for constraint is not set. Using {self.name}.", stacklevel=2)
+
+        # Infer layers from descriptor and tags
+        self.layers = set()
+        for tag in self.tags:
+            if not self.descriptor.exists(tag):
+                raise ValueError(
+                    f"The tag {tag} used with constraint "
+                    f"{self.name} is not defined in the descriptor. Please "
+                    "add it to the correct layer using "
+                    "descriptor.add('layer', ...)."
+                )
+
+            layer, _ = self.descriptor.location(tag)
+            self.layers.add(layer)
+
+    @abstractmethod
+    def check_constraint(self, data: dict[str, Tensor]) -> tuple[Tensor, Tensor]:
+        """Evaluates whether the given model predictions satisfy the constraint.
+
+        1 IS SATISFIED, 0 IS NOT SATISFIED
+
+        Args:
+            data (dict[str, Tensor]): Dictionary that holds batch data, model predictions and context.
+
+        Returns:
+            tuple[Tensor, Tensor]: A tuple where the first element is a tensor of floats
+            indicating whether the constraint is satisfied (with value 1.0
+            for satisfaction, and 0.0 for non-satisfaction, and the second element is a tensor
+            mask that indicates the relevance of each sample (`True` for relevant
+            samples and `False` for irrelevant ones).
+
+        Raises:
+            NotImplementedError: If not implemented in a subclass.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def calculate_direction(self, data: dict[str, Tensor]) -> dict[str, Tensor]:
+        """Compute adjustment directions to better satisfy the constraint.
+
+        Given the model predictions, input batch, and context, this method calculates the direction
+        in which the predictions referenced by a tag should be adjusted to satisfy the constraint.
+
+        Args:
+            data (dict[str, Tensor]): Dictionary that holds batch data, model predictions and context.
+
+        Returns:
+            dict[str, Tensor]: Dictionary mapping network layers to tensors that
+                specify the adjustment direction for each tag.
+
+        Raises:
+            NotImplementedError: Must be implemented by subclasses.
+        """
+        raise NotImplementedError