Skip to content

Add _metrics dict to Trainer for custom metric logging #43636

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
abigailtech wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Add _metrics dict to Trainer for custom metric logging #43636

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1478c0c
Add _metrics dict to Trainer for custom metric logging
abigailtech Jan 30, 2026
57d1f56
Merge branch 'main' into add-trainer-metrics-dict
abigailtech Jan 30, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 11 additions & 0 deletions src/transformers/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,7 @@
import tempfile
import time
import warnings
from collections import defaultdict
from collections.abc import Callable, Iterator, Mapping
from functools import partial
from pathlib import Path
Expand Down Expand Up @@ -435,6 +436,8 @@ def __init__(
self._memory_tracker = TrainerMemoryTracker(self.args.skip_memory_metrics)
self._memory_tracker.start()

self._metrics = {"train": defaultdict(list), "eval": defaultdict(list)}
Copy link
Copy Markdown
Member

@qgallouedec qgallouedec Feb 3, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm wondering if we can leverage this self._metrics to log the loss

Copy link
Copy Markdown
Author

@abigailtech abigailtech Mar 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

though about it, but loss uses tensor accumulation to avoid TPU sync overhead. metrics is more for auxiliary custom metrics that don't need that optimization


# set the correct log level depending on the node
log_level = args.get_process_log_level()
logging.set_verbosity(log_level)
Expand Down Expand Up @@ -3567,6 +3570,14 @@ def log(self, logs: dict[str, float], start_time: float | None = None) -> None:
start_time (`Optional[float]`):
The start of training.
"""
mode = "train" if self.model.training else "eval"
if self._metrics[mode]:
metrics = {key: sum(val) / len(val) for key, val in self._metrics[mode].items()}
if mode == "eval":
metrics = {f"eval_{key}": val for key, val in metrics.items()}
logs = {**logs, **metrics}
self._metrics[mode].clear()

if self.state.epoch is not None:
logs["epoch"] = self.state.epoch
if self.args.include_num_input_tokens_seen != "no":
Expand Down