Control log_dict's on_epoch log name

[johanneskvamme]

Hey!

I tried to change to using the new way of logging through MetricCollection and self.log_dict instead of logging every metric through self.log on training step and test_epoch_end. However, each metric is then logged as [metric_name]_epoch_[epoch_number] which creates a new graph for every epoch instead of allowing me to use epoch on the x-axis of my graphs (on Comet, if that is relevant). Is there a way to control this behaviour of log_dict, or do I just have to keep logging "manually" to control log name?

[SkafteNicki]

Hi @FluidSense,

Could you try just updating the MetricCollection during the _step method and then log in _epoch_end method. Something like:

def training_step(self, batch, batch_idx):
    logits = self(x)
    self.train_metrics.update(logits, y)

def train_epoch_end(self, outputs):
    self.log_dict(self.train_metrics.compute())

[johanneskvamme]

@SkafteNicki Thanks for the suggestion! I'll try that first thing in the morning and come back to you here afterwards

[johanneskvamme]

This works great for us at least! It takes a bit more time to get the first indications of how things are going, but at least the metrics are gathered correctly.