Fastconform SSL pretrained recipe failed potentially due to PTL 2.0

[XuesongYang]

Describe the bug

the examples/asr/speech_pretraining/speech_pre_training.py is not working properly on the lates main branch, but working in 1.20. At first glance, the error could be related to the upgrade of Pytorch lightning 2.0.

Error Brief

lightning_fabric.utilities.exceptions.MisconfigurationException: `ModelCheckpoint(monitor='val_loss')` could not find the monitored key in the returned metrics: ['learning_rate', 'global_step', 'train_contrastive', 'train_backward_timing in s', 'train_step_timing in s', 'epoch', 'step']. HINT: Did you call `log('val_loss', value)` in the `LightningModule`?

Error Details

error_logs.zip

[XuesongYang]

It seems this PR will fix the error: #7505

[titu1994]

Good catch, for some reason PR was closed though.. @KunalDhawan ?

[KunalDhawan]

I've opened a new PR for the issue to keep a consistent format of logging across RNNT, CTC, and Hybrid models - #7531. Once this is approved we can make the same changes for SSL, SLU and other models

[nithinraok]

@XuesongYang are you still seeing these errors?

[XuesongYang]

I switched back to r1.20.0 release, and it worked well. I haven't tried it on the latest r1.21.0 yet. Do we have a nightly test for the minimalist fastconformer SSL training recipe?

[nithinraok]

We have SSL training notebook, and it was fixed for PTL upgrade.

[XuesongYang]

Let's close this issue then. Thanks!