[Bug] MLflowVisBackend.add_image fails when image name has no extension

[kpysanyi-raylytic]

Prerequisite

• I have searched Issues and Discussions but cannot get the expected help.
• The bug has not been fixed in the latest version(https://github.com/open-mmlab/mmengine).

Environment

OrderedDict([
  ('sys.platform', 'linux'),
  ('Python', '3.11.14'),
  ('CUDA available', True),
  ('GPU 0,1', 'NVIDIA GeForce RTX 3090'),
  ('CUDA_HOME', '/usr/local/cuda'),
  ('NVCC', 'Cuda compilation tools, release 12.6, V12.6.85'),
  ('GCC', '9.4.0'),
  ('PyTorch', '2.1.0+cu121'),
  ('TorchVision', '0.16.0+cu121'),
  ('OpenCV', '4.11.0'),
  ('MMEngine', '0.10.7'),
])

Reproduces the problem - code sample

🧩 Minimal config to reproduce

vis_backends = [
    dict(type='LocalVisBackend'),
    dict(
        type='MLflowVisBackend',
        tracking_uri="http://localhost:2222",
        exp_name='vitdet_coco',
        run_name=None,
    ),
]

visualizer = dict(
    type='DetLocalVisualizer',
    vis_backends=vis_backends,
    name='visualizer',
)

default_hooks = dict(
    logger=dict(type='LoggerHook', interval=50),
    visualization=dict(
        type='DetVisualizationHook',
        draw=True,
        interval=1,
    ),
)

During validation, DetVisualizationHook calls:

visualizer.add_image('val_img', image, step)

• LocalVisBackend → saves val_img_{step}.png ✅
• MLflowVisBackend → passes val_img to mlflow.log_image() ❌

---

Below is a minimal end-to-end reproduction using a tiny COCO subset so it runs quickly, but still triggers a validation visualization step and crashes in MLflowVisBackend.add_image() due to an extension-less image name.

1) Prerequisites

• MMDetection checkout with COCO available at data/coco/:

  • data/coco/annotations/instances_train2017.json
  • data/coco/annotations/instances_val2017.json
  • data/coco/train2017/*.jpg
  • data/coco/val2017/*.jpg

• MLflow tracking server running locally (example):

  mlflow server --host 0.0.0.0 --port 2222

2) Create a minimal debug config

Save this as configs/_debug/mlflow_visbackend_no_ext_repro.py:

# Repro config: triggers MLflowVisBackend crash when name has no extension

_base_ = ['./vitdet_mask-rcnn_vit-b-dinov3.py']  # any base detector config that runs on COCO

# Make it fast: 10 train iters then run val once.
train_cfg = dict(type="IterBasedTrainLoop", max_iters=10, val_interval=10)

# Ensure validation is very small (2 iters) but still calls visualization hook.
train_dataloader = dict(
    dataset=dict(
        indices=64,
        # avoid empty dataset when taking first N
        filter_cfg=dict(filter_empty_gt=False, min_size=0),
    )
)
val_dataloader = dict(dataset=dict(indices=2))
test_dataloader = val_dataloader

# Enable visualization every iter.
default_hooks = dict(
    logger=dict(type='LoggerHook', interval=1),
    checkpoint=dict(type='CheckpointHook', by_epoch=False, interval=10, save_last=True),
    visualization=dict(
        type='DetVisualizationHook',
        draw=True,
        interval=1,
        # show=False is default; when show=False, DetVisualizationHook uses an extension-less name
    ),
)

# Use both LocalVisBackend (works) and MLflowVisBackend (crashes).
vis_backends = [
    dict(type='LocalVisBackend'),
    dict(
        type='MLflowVisBackend',
        tracking_uri="http://localhost:2222",
        exp_name='vitdet_coco',
        run_name=None,
    ),
]
visualizer = dict(type='DetLocalVisualizer', vis_backends=vis_backends, name='visualizer')

Reproduces the problem - command or script

3) Run training

python tools/train.py configs/_debug/mlflow_visbackend_no_ext_repro.py

Reproduces the problem - error message

Traceback (most recent call last):
File "/home/kpysanyi/xraivision-backbone/.venv/lib/python3.11/site-packages/PIL/Image.py", line 2526, in save
format = EXTENSION[ext]
~~~~~~~~~^^^^^
KeyError: ''

The above exception was the direct cause of the following exception:
...
File "/home/kpysanyi/xraivision-backbone/.venv/lib/python3.11/site-packages/mmengine/visualization/vis_backend.py", line 784, in add_image
self._mlflow.log_image(image, name)
File "/home/kpysanyi/xraivision-backbone/.venv/lib/python3.11/site-packages/mlflow/tracking/fluent.py", line 1473, in log_image
MlflowClient().log_image(run_id, image, artifact_file, key, step, timestamp, synchronous)
File "/home/kpysanyi/xraivision-backbone/.venv/lib/python3.11/site-packages/mlflow/tracking/client.py", line 2797, in log_image
image.save(tmp_path)
File "/home/kpysanyi/xraivision-backbone/.venv/lib/python3.11/site-packages/PIL/Image.py", line 2529, in save
raise ValueError(msg) from e
ValueError: unknown file extension:

Additional information

🔍 Root cause

Inconsistency between backends:

• LocalVisBackend.add_image() forces a valid filename
• MLflowVisBackend.add_image() assumes name already includes an extension
• step argument is ignored in MLflow backend

This makes MLflowVisBackend fragile and incompatible with existing hooks.

---

✅ Proposed solutions

Either of the following would fix the issue cleanly:

Option A (configurable)

Add an argument to MLflowVisBackend, e.g.

auto_append_ext=True

which would automatically transform names like val_img → val_img_{step}.png.

Option B (default behavior)

Make MLflowVisBackend.add_image() mirror LocalVisBackend behavior:

• If name has no extension, append _{step}.png by default.

Example logic:

if '.' not in os.path.basename(name):
    name = f'{name}_{step}.png'
self._mlflow.log_image(image, name)

This would:

• Make behavior consistent across backends
• Respect the existing step argument
• Avoid hard-to-debug runtime crashes

I’d be happy to submit a PR implementing this fix (either as a default behavior or a configurable option), if that aligns with the maintainers’ preferences.