Add base AnomalyEvaluator class (#850)

Summary:
Pull Request resolved: #850

### This Stack

Based on [this RFC](https://docs.google.com/document/d/1K1KQ886dynMRejR0ySH1fctOjS7gxaCS8AB1L_PHxU4/edit?usp=sharing), we are adding a new logger that warns about anomalous values in metrics, and optionally executes a callback function with potential side effects. This could be useful for users to realize sooner that something has gone wrong during training.

### This Diff

To provide flexibility when detecting anomalous metric values, instead of assuming and hardcoding a predefined check (like a threshold), let's create an interface that can be overriden to implement custom checks.

Reviewed By: JKSenthil

Differential Revision: D58564201

fbshipit-source-id: 662ae38dd866aa979e84e76f6364ec1d57c1e090

Author: diego-urgell

torchtnt/utils/anomaly_evaluation.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+
+
+import logging
+from abc import ABC, abstractmethod
+
+_logger: logging.Logger = logging.getLogger(__name__)
+
+
+class MetricAnomalyEvaluator(ABC):
+    """
+    Abstract base class for metric anomaly evaluators. An evaluator specifies the logic to determine that
+    a particular metric value is anomalous. To implement a custom method, create a subclass and implement
+    the following methods:
+    - :py:meth:`~torchtnt.utils.loggers.metric_anomaly_logger.MetricAnomalyEvaluator.update` should receive
+        the metric value and update the internal state. This is specially useful for algorithms that require
+        storing some previous values, moving averages, etc.
+    - :py:meth:`~torchtnt.utils.loggers.metric_anomaly_logger.MetricAnomalyEvaluator.is_anomaly` determines
+        whether the current metric state is anomalous.
+
+    Likely there are some warm-up steps before the metric is stable and can be checked against anomalies, so
+    the separation of state update and actual detection provides this flexibility.
+    """
+
+    @abstractmethod
+    def update(self, value: float) -> None:
+        """
+        Update the internal state with the given metric value. This should not determine anomalies itself, but
+        only aggregate the current value according to the anomaly detection algorithm.
+
+        Note:: If no aggregation is required, this method can store the value directly, to be used in `is_anomaly`.
+
+        Args:
+            value: Metric value
+        """
+        pass
+
+    @abstractmethod
+    def is_anomaly(self) -> bool:
+        """
+        Determine whether the current metric state is anomalous. This should be overridden with custom logic related to
+        an anomaly detection algorithm.
+        """
+        pass