How to print acc while training?

[fire717]

🚀 Feature

Hi, I wanna print acc(or other metrics) while training in the traing bar,(just like Keras).

And for pytorch, I can compute cumulative acc for every training batch and print it,
However, In lightning, maybe I can compute cumulative acc in training_step(), but it seems like I can only print at the val time? How to print it in the training bar?

I am new to pytorch-lightning, so maybe there is something I missed in the docs?

Motivation

To know the training status at every batch, not every epoch (like training_epoch_end() ).

possibly related issue: 100

[github-actions]

Hi! thanks for your contribution!, great first issue!

[fire717]

Like in Keras:

We set:
model.compile(loss="categorical_crossentropy", optimizer=opt, metrics=['acc'])

Then can get:
25/5405 [..............................] - ETA: 1:13:15 - loss: 0.7136 - acc: 0.6556

And default in lightning:
Epoch 0: 5%|██▉ | 63/1353 [00:22<07:50, 2.74it/s, loss=0.71, v_num=20

Only have loss.

[justusschock]

Hey thanks for your question.

In lightning we have the Metrics Package that also describes it's usage and if you log them with prog_bar=True they will also appear on your progress bar as described here.

Hope this answers your question :)

[fire717]

Thx, it seems like useful! I am going to have a try!

[fire717]

It is a little strange, when I run in version 1.0.8, the bar cannot show acc and loss is nan,
while I change to v1.0.0, the bar can show loss value and acc .

[justusschock]

cc @teddykoker @ananyahjha93 any idea on this?

[ananyahjha93]

I'll have to see what changes have been added in between, nothing that I am aware of should cause this issue.