congrads 0.2.0__py3-none-any.whl → 0.3.1__py3-none-any.whl

congrads/__init__.py

@@ -1,22 +1,11 @@
-# __init__.py
-version = "0.2.0"
+try:  # noqa: D104
+    from importlib.metadata import version as get_version  # Python 3.8+
+except ImportError:
+    from pkg_resources import (
+        get_distribution as get_version,
+    )  # Fallback for older versions
 
-# Only expose the submodules, not individual classes
-from . import constraints
-from . import core
-from . import datasets
-from . import descriptor
-from . import metrics
-from . import networks
-from . import utils
-
-# Define __all__ to specify that the submodules are accessible, but not classes directly.
-__all__ = [
-    "constraints",
-    "core",
-    "datasets",
-    "descriptor",
-    "metrics",
-    "networks",
-    "utils",
-]
+try:
+    version = get_version("congrads")  # Replace with your package name
+except Exception:
+    version = "0.0.0"  # Fallback if the package isn't installed

congrads/callbacks/base.py

@@ -0,0 +1,360 @@
+"""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
+
+__all__ = ["Callback", "CallbackManager", "Operation"]
+
+
+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 +1,165 @@
+"""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 import Tensor
+from torch.utils.tensorboard import SummaryWriter
+
+from ..callbacks.base import Callback
+from ..metrics import MetricManager
+from ..utils.utility import CSVLogger
+
+__all__ = ["LoggerCallback"]
+
+
+class LoggerCallback(Callback):
+    """Callback to periodically aggregate and store metrics during training and testing.
+
+    This callback works in conjunction with a MetricManager that accumulates metrics
+    internally (e.g. per batch). Metrics are:
+
+    - Aggregated at a configurable epoch interval (`aggregate_interval`)
+    - Cached in memory (GPU-resident tensors)
+    - Written to TensorBoard and CSV at a separate interval (`store_interval`)
+
+    Aggregation and storage are decoupled to avoid unnecessary GPU-to-CPU
+    synchronization. Any remaining cached metrics are flushed at the end of training.
+
+    Methods implemented:
+    - on_epoch_end: Periodically aggregates and stores training metrics.
+    - on_train_end: Flushes any remaining cached training metrics.
+    - on_test_end: Aggregates and stores test metrics immediately.
+    """
+
+    def __init__(
+        self,
+        metric_manager: MetricManager,
+        tensorboard_logger: SummaryWriter,
+        csv_logger: CSVLogger,
+        *,
+        aggregate_interval: int = 1,
+        store_interval: int = 1,
+    ):
+        """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.
+            aggregate_interval: Number of epochs between metric aggregation.
+            store_interval: Number of epochs between metric storage.
+        """
+        super().__init__()
+
+        # Input validation
+        if aggregate_interval <= 0 or store_interval <= 0:
+            raise ValueError("Intervals must be positive integers")
+
+        if store_interval % aggregate_interval != 0:
+            raise ValueError("store_interval must be a multiple of aggregate_interval")
+
+        # Store references
+        self.metric_manager = metric_manager
+        self.tensorboard_logger = tensorboard_logger
+        self.csv_logger = csv_logger
+        self.aggregate_interval = aggregate_interval
+        self.store_interval = store_interval
+
+        # Cached metrics on GPU by epoch
+        self._accumulated_metrics: dict[int, dict[str, Tensor]] = {}
+
+    def on_epoch_end(self, data: dict[str, any], ctx: dict[str, any]):
+        """Handle end-of-epoch training logic.
+
+        At the end of each epoch, this method may:
+        - Aggregate training metrics from the MetricManager (every `aggregate_interval` epochs)
+        - Cache aggregated metrics keyed by epoch
+        - Store cached metrics to disk (every `store_interval` epochs)
+
+        Metric aggregation resets the MetricManager accumulation state.
+        Metric storage triggers GPU-to-CPU synchronization and writes to loggers.
+
+        Args:
+            data: Dictionary containing epoch context (must include 'epoch').
+            ctx: Additional context dictionary (unused).
+
+        Returns:
+            data: The same input dictionary, unmodified.
+        """
+        epoch = data["epoch"]
+
+        # Cache training metrics
+        if epoch % self.aggregate_interval == 0:
+            metrics = self.metric_manager.aggregate("during_training")
+            self._accumulated_metrics[epoch] = metrics
+            self.metric_manager.reset("during_training")
+
+        # Store metrics to disk
+        if epoch % self.store_interval == 0:
+            self._save(self._accumulated_metrics)
+            self._accumulated_metrics.clear()
+
+        return data
+
+    def on_train_end(self, data, ctx):
+        """Flush any remaining cached training metrics at the end of training.
+
+        This ensures that aggregated metrics that were not yet written due to
+        `store_interval` alignment are persisted before training terminates.
+
+        Args:
+            data: Dictionary containing training context (unused).
+            ctx: Additional context dictionary (unused).
+
+        Returns:
+            data: The same input dictionary, unmodified.
+        """
+        if self._accumulated_metrics:
+            self._save(self._accumulated_metrics)
+            self._accumulated_metrics.clear()
+
+        return data
+
+    def on_test_end(self, data: dict[str, any], ctx: dict[str, any]):
+        """Aggregate and store test metrics at the end of testing.
+
+        Test metrics are aggregated once and written immediately to disk.
+        Interval-based aggregation and caching are not applied to testing.
+
+        Args:
+            data: Dictionary containing test context (must include 'epoch').
+            ctx: Additional context dictionary (unused).
+
+        Returns:
+            data: The same input dictionary, unmodified.
+        """
+        epoch = data["epoch"]
+
+        # Save test metrics
+        metrics = self.metric_manager.aggregate("after_training")
+        self._save({epoch: metrics})
+        self.metric_manager.reset("after_training")
+
+        return data
+
+    def _save(self, metrics: dict[int, dict[str, Tensor]]):
+        """Write aggregated metrics to TensorBoard and CSV loggers.
+
+        Args:
+            metrics: Mapping from epoch to a dictionary of metric name to scalar tensor.
+                Tensors are expected to be detached and graph-free.
+        """
+        for epoch, metrics_by_name in metrics.items():
+            for name, value in metrics_by_name.items():
+                cpu_value = value.item()
+                self.tensorboard_logger.add_scalar(name, cpu_value, epoch)
+                self.csv_logger.add_value(name, cpu_value, epoch)
+
+        # Flush/save
+        self.tensorboard_logger.flush()
+        self.csv_logger.save()