Custom Logger does not create output file with DDP

[yukkyo]

I tried the following source code with ddp.
(reference: https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/)
.ckpt file is created, but a .csv file (by metrics_df.to_csv(self.df_path, index=False)) is not created.
And print(f'metrics_df: \n {metrics_df}') is worked.
Do you know anything? 😢

class MyLogger(LightningLoggerBase):
    def __init__(self, logger_df_path: Path):
        super(MyLogger, self).__init__()
        self.all_metrics = defaultdict(list)
        self.df_path = logger_df_path

    @rank_zero_only
    def log_hyperparams(self, params):
        pass

    @rank_zero_only
    def log_metrics(self, metrics, step_num):
        if self.rank > 0 or len(metrics) < 2:
            return

        if step_num is None:
            step_num = len(self.all_metrics['step_num'])
        self.all_metrics['step_num'].append(step_num)

        if 'created_at' not in metrics:
            metrics['created_at'] = str(datetime.utcnow())

        for k, v in metrics.items():
            self.all_metrics[k].append(v)

        metrics_df = pd.DataFrame(self.all_metrics)
        print(f'metrics_df: \n {metrics_df}')
        print(f'save to: {self.df_path}')
        metrics_df.to_csv(self.df_path, index=False)


class MyModelCheckpoint(ModelCheckpoint):
    def __init__(self, kfold, **args):
        super(MyModelCheckpoint, self).__init__(**args)
        self.monitor = 'val_loss'
        self.kfold = kfold
        self.latest_path = f'{self.filepath}/kfold_{kfold}_latest.ckpt'

    def on_epoch_end(self, epoch, logs=None):
        if self.verbose > 0:
            print(f'\nSaving latest model to {self.latest_path}')
        self.save_model(self.latest_path, overwrite=True)


def train_a_kfold(cfg, output_path):
    model = LightningModuleSeg(cfg)
    debug = True

    kfold = 1
    logger_name = f'kfold_{str(kfold).zfill(2)}.csv'
    mylogger = MyLogger(logger_df_path=output_path / logger_name)

    checkpoint_callback = MyModelCheckpoint(
        kfold=kfold,
        filepath=str(output_path),
        verbose=True,  # print save result, not must
    )

    trainer = MyTrainer(
        logger=mylogger,
        max_nb_epochs=3,
        checkpoint_callback=checkpoint_callback,
        train_percent_check=0.01 if debug else 1.0,
        val_percent_check=0.01 if debug else 1.0,
        gpus=[0, 1, 2, 3],
        use_amp=True,
        amp_level='O2',
        distributed_backend='ddp',
        log_save_interval=5 if debug else 200,
        row_log_interval=1e8
    )
    trainer.fit(model)

[yukkyo]

This was due to overwrite = True in self.save_model(self.latest_path, overwrite=True).
I solved by overwrite = False.