TensorBoardLogger creates another tfevents file.

[S-aiueo32]

🐛 Bug

TensorBoardLogger creates another tfevents file when fit() is running.
It seems that no metrics are logged in the redundant file, but it will be shown in TensorBoard as a run.
I don't do anything about loggers in my LightningModules.

Expected file structure:

|
|- save_dir
|    |- name
|         |- version_0
|               |- events.out.tfevents.1579833025.ip-xxx-xxx-xxx-xxx.17584.0
|               |- meta_tags.csv
|- train.py

Observed file structure:

|
|- save_dir
|    |- name
|         |- version_0
|               |- 1579833032
|                     |- events.out.tfevents.1579833032.ip-xxx-xxx-xxx-xxx.17584.1
|               |- events.out.tfevents.1579833025.ip-xxx-xxx-xxx-xxx.17584.0
|               |- meta_tags.csv
|- train.py

To Reproduce / Code sample

Basic training step of PyTorch Lightning:

# define a logger
logger = TensorBoardLogger(
    save_dir='runs',
    name=args.model
)

# define callbacks
ckpt_path = Path(logger.experiment.log_dir) / 'ckpts'
checkpoint_callback = ModelCheckpoint(filepath=ckpt_path)

# instantiate trainer
trainer = Trainer(
    logger=logger,
    checkpoint_callback=checkpoint_callback,
    gpus=args.gpus
)

# define a model
model = CoolModel(args)

# start training!
trainer.fit(model)

Environment

• PyTorch Lightning Version (e.g., 1.0): 0.6.0
• PyTorch Version (e.g., 1.0): 1.3.1
• OS (e.g., Linux): Ubuntu 16.04 LTS
• How you installed PyTorch (conda, pip, source): pip
• Build command you used (if compiling from source):
• Python version: 3.7.4
• CUDA/cuDNN version: 10.1