When using torchmetrics does pytorch-lightning averages the previous metrics

[yashrathi-git]

By default lightning logs every n steps. So in the nth step does the metrics displayed the average of previous n minibatches or just the accuracy of this specific minibatch?

[rohitgr7]

accuracy of the specific batch. the metric is calculated at every step. log_every_n_steps is just a barrier that allows that value at that step to pass through the loggers.

[yashrathi-git]

why do the logger not avg the prev n metrics, that here seems most sensible thing to do? What is a good way to implement this?

[rohitgr7]

you can manually compute it at every steps using some state variables and log it. to create prev n-step logic, you can reset the state after certain steps.

[yashrathi-git]

Alright thanks

[gisilvs]

Shouldn't this be a feature of Lightning? When I train using iterations and not epochs, I am interested in the average loss value over the last n batches since the last logs, not on the epoch average or current step value.

[adosar]

According to the docs:

• reduce_fx: Reduction function over step values for end of epoch. Uses torch.mean() by default and is not applied when a torchmetrics.Metric is logged.

Does this mean that torchmetrics.Metric is not averaged during the end of the epoch?