Add ignore_metrics field to the MLflow logger (mosaicml#2869)

* Add ignore_metrics field to MLflow logger

* Improve comment

* Address pyright issues

* Fix comment

* Use regex instead

* Use glob patterns instead

* Improve test case

---------

Co-authored-by: Daniel King <[email protected]>

composer/loggers/mlflow_logger.py

@@ -5,6 +5,7 @@
 
 from __future__ import annotations
 
+import fnmatch
 import os
 import pathlib
 import textwrap
@@ -54,6 +55,7 @@ class MLFlowLogger(LoggerDestination):
             synchronously to the MLflow backend. If ``False``, Mlflow will log asynchronously. (default: ``False``)
         log_system_metrics (bool, optional): Whether to log system metrics. If ``True``, Mlflow will
             log system metrics (CPU/GPU/memory/network usage) during training. (default: ``True``)
+        ignore_metrics (List[str], optional): A list of glob patterns for metrics to ignore when logging. (default: ``None``)
     """
 
     def __init__(
@@ -68,6 +70,7 @@ def __init__(
         model_registry_uri: Optional[str] = None,
         synchronous: bool = False,
         log_system_metrics: bool = True,
+        ignore_metrics: Optional[List[str]] = None,
     ) -> None:
         try:
             import mlflow
@@ -85,6 +88,7 @@ def __init__(
         self.model_registry_uri = model_registry_uri
         self.synchronous = synchronous
         self.log_system_metrics = log_system_metrics
+        self.ignore_metrics = [] if ignore_metrics is None else ignore_metrics
         if self.model_registry_uri == 'databricks-uc':
             if len(self.model_registry_prefix.split('.')) != 2:
                 raise ValueError(f'When registering to Unity Catalog, model_registry_prefix must be in the format ' +
@@ -198,7 +202,11 @@ def log_metrics(self, metrics: Dict[str, Any], step: Optional[int] = None) -> No
         from mlflow import log_metrics
         if self._enabled:
             # Convert all metrics to floats to placate mlflow.
-            metrics = {k: float(v) for k, v in metrics.items()}
+            metrics = {
+                k: float(v)
+                for k, v in metrics.items()
+                if not any(fnmatch.fnmatch(k, pattern) for pattern in self.ignore_metrics)
+            }
             log_metrics(
                 metrics=metrics,
                 step=step,

tests/loggers/test_mlflow_logger.py

@@ -576,3 +576,69 @@ def before_forward(self, state: State, logger: Logger):
     run_file_path = mlflow_uri / Path(experiment_id) / Path(run_id)
     im_dir = run_file_path / Path('artifacts')
     assert len(os.listdir(im_dir)) == expected_num_ims
+
+
+@device('cpu')
+def test_mlflow_ignore_metrics(tmp_path, device):
+    mlflow = pytest.importorskip('mlflow')
+
+    mlflow_uri = tmp_path / Path('my-test-mlflow-uri')
+    experiment_name = 'mlflow_logging_test'
+    test_mlflow_logger = MLFlowLogger(
+        tracking_uri=mlflow_uri,
+        experiment_name=experiment_name,
+        log_system_metrics=False,
+        ignore_metrics=['metrics/eval/*', 'nothing/should/match', 'metrics/train/CrossEntropy'],
+    )
+    # Reduce the system metrics sampling interval to speed up the test.
+    mlflow.set_system_metrics_sampling_interval(1)
+
+    dataset_size = 64
+    batch_size = 4
+    num_batches = 4
+    eval_interval = '1ba'
+
+    trainer = Trainer(model=SimpleConvModel(),
+                      loggers=test_mlflow_logger,
+                      train_dataloader=DataLoader(RandomImageDataset(size=dataset_size), batch_size),
+                      eval_dataloader=DataLoader(RandomImageDataset(size=dataset_size), batch_size),
+                      max_duration=f'{num_batches}ba',
+                      eval_interval=eval_interval,
+                      device=device)
+    trainer.fit()
+    # Allow async logging to finish.
+    time.sleep(3)
+    test_mlflow_logger.post_close()
+
+    run = _get_latest_mlflow_run(
+        experiment_name=experiment_name,
+        tracking_uri=mlflow_uri,
+    )
+    run_id = run.info.run_id
+    experiment_id = run.info.experiment_id
+
+    run_file_path = mlflow_uri / Path(experiment_id) / Path(run_id)
+
+    # Test metrics logged.
+    for metric_name in [
+            'metrics/train/MulticlassAccuracy',
+            'loss/train/total',
+    ]:
+        metric_file = run_file_path / Path('metrics') / Path(metric_name)
+        with open(metric_file) as f:
+            csv_reader = csv.reader(f, delimiter=' ')
+            lines = list(csv_reader)
+
+        assert len(lines) == num_batches
+
+    # Test metrics are not logged.
+    for metric_name in ['metrics/eval/MulticlassAccuracy', 'metrics/eval/CrossEntropy', 'metrics/train/CrossEntropy']:
+        metric_file = run_file_path / Path('metrics') / Path(metric_name)
+        assert not os.path.exists(metric_file)
+
+    # Test system metrics are not logged.
+    metric_file = run_file_path / Path('metrics') / Path('system/cpu_utilization_percentage')
+    assert not os.path.exists(metric_file)
+
+    # Undo the setup to avoid affecting other test cases.
+    mlflow.set_system_metrics_sampling_interval(None)