Don't include NaN metric values in ckpt paths (#896)

Summary: Pull Request resolved: #896

Reviewed By: JKSenthil

Differential Revision: D62469085

fbshipit-source-id: 746ba7d16390e2cc7fa513961f500317c73bcf06

Author: diego-urgell

tests/framework/callbacks/test_base_checkpointer.py

@@ -919,6 +919,23 @@ def test_get_tracked_metric_value(self) -> None:
         ):
             val_loss = val_loss_ckpt_cb._get_tracked_metric_value(val_loss_unit)
 
+        val_loss_unit.val_loss = float("nan")  # Test nan metric value
+        error_container = []
+        with patch(
+            "torchtnt.framework.callbacks.base_checkpointer.logging.Logger.error",
+            side_effect=error_container.append,
+        ):
+            val_loss = val_loss_ckpt_cb._get_tracked_metric_value(val_loss_unit)
+
+        self.assertEqual(
+            [
+                "Monitored metric 'val_loss' is NaN. Will not be included in checkpoint path, nor tracked for optimality."
+            ],
+            error_container,
+        )
+        self.assertIsNone(val_loss)
+
+        # test with mismatched monitored metric
         train_loss_ckpt_cb = BaseCheckpointSaver(
             dirpath="checkpoint",
             best_checkpoint_config=BestCheckpointConfig("train_loss", "max"),

tests/utils/test_checkpoint.py

@@ -47,6 +47,37 @@
 
 
 class CheckpointPathTest(unittest.TestCase):
+
+    def test_create_checkpoint_path(self) -> None:
+        # phase-naive and metric-naive
+        ckpt = CheckpointPath("foo", epoch=0, step=1)
+        self.assertEqual(ckpt.path, "foo/epoch_0_step_1")
+
+        # phase-aware and metric-naive
+        ckpt = CheckpointPath("foo", epoch=0, step={Phase.TRAIN: 1})
+        self.assertEqual(ckpt.path, "foo/epoch_0_train_step_1")
+
+        # phase-aware and metric-aware
+        ckpt = CheckpointPath(
+            "foo",
+            epoch=0,
+            step={Phase.TRAIN: 1, Phase.EVALUATE: 1},
+            metric_data=MetricData("foo", 1.0),
+        )
+        self.assertEqual(ckpt.path, "foo/epoch_0_train_step_1_eval_step_1_foo=1.0")
+
+        # nan metric value
+        with self.assertRaisesRegex(
+            ValueError,
+            "Value of monitored metric 'foo' can't be NaN in CheckpointPath.",
+        ):
+            CheckpointPath(
+                "foo",
+                epoch=0,
+                step={Phase.TRAIN: 1, Phase.EVALUATE: 1},
+                metric_data=MetricData("foo", float("nan")),
+            )
+
     def test_from_str(self) -> None:
         # invalid paths
         malformed_paths = [

torchtnt/framework/callbacks/base_checkpointer.py

@@ -8,6 +8,7 @@
 
 import abc
 import logging
+import math
 from datetime import timedelta
 from typing import Any, cast, Iterable, List, Literal, Optional, Union
 
@@ -256,6 +257,12 @@ def _get_tracked_metric_value(
                     "can be converted to float and is not a multi-element tensor value."
                 ) from e
 
+        if metric_value_f and math.isnan(metric_value_f):
+            logger.error(
+                f"Monitored metric '{monitored_metric_name}' is NaN. Will not be included in checkpoint path, nor tracked for optimality."
+            )
+            return None
+
         return metric_value_f
 
     def on_train_start(self, state: State, unit: TTrainUnit) -> None:

torchtnt/utils/checkpoint.py

@@ -7,6 +7,7 @@
 # pyre-strict
 import bisect
 import logging
+import math
 import os
 import re
 from dataclasses import dataclass
@@ -105,6 +106,11 @@ def __init__(
             step if isinstance(step, dict) else {Phase.NONE: step}
         )
 
+        if metric_data and math.isnan(metric_data.value):
+            raise ValueError(
+                f"Value of monitored metric '{metric_data.name}' can't be NaN in CheckpointPath."
+            )
+
     @classmethod
     def from_str(cls, checkpoint_path: str) -> "CheckpointPath":
         """