Logging Validation Loss and Test Loss during Epoch

[severin-magel]

I want to log validation and test losses also during my epochs because they are veeeeery long and I would like to have some insides into this already during the epochs. I’m using the PyTorch Lightning WandbLogger and I tried the following:
construct the Trainer:

Trainer(logger=WandbLogger, val_check_interval=0.1,  log_every_n_steps=1)

log in the validation_step() and training_step() with

self.log(
        "train_loss",  # or 'val_loss' in the validation case
        loss,
        batch_size=self._get_batch_size(train_batch),
        prog_bar = True,
        on_epoch = True,
        on_step   = True,
        sync_dist = True,
)

The outcome I get in W&B looks like this:

with the dashed line being the val_loss. This was during one epoch with batchsize 10 and 100 datasamples in both validation and training datasets. I thought that global_step is a good metric to plot against. But global step seems to have a weird behaviour.