Pytorch-lightning: Accessing measured time per epoch shown in progress bar.

[christabella]

What is your question?

How do I access the measured time per epoch shown in progress bar?

Epoch 8: : 150it [00:03, 46.84it/s, loss=1.625, v_num=0]

It clearly says 00:03 i.e. 3 seconds and I could parse the logs as a hack, but I was wondering if there was any clean way to access the measured time elapsed per training epoch, as I would like to write it to a file.

And if not, how would you recommend timing each epoch? I could put a time.time() in training_step but then would need to add it all up in training_epoch_end, would this work? Thank you!

What's your environment?

• OS: Mac OSX
• Packaging: Pip-installed packages in a conda environment
• Version: 0.7.1

[mpariente]

Just an idea, if you use tensorboard as the logger, it has timestamps for everything, you could open the log (in python) and find the stamps.

[williamFalcon]

I would use tensorboard logs.

Otherwise just build your own callback and overwrite epoch_start and epoch_end

https://pytorch-lightning.readthedocs.io/en/latest/callbacks.html