Learning rate increasing with Noam scheduler over time?

[RamakrishnaChaitanya]

As far as i know, the learning rate gradually decreases over time. However, with the Noam scheduler, the learning rate is always increasing in my case (~100k steps). From the documentation, i understood that there will be a warmup stage in which learning rate gradually increases and a decay stage in which the learning rate gradually decreases. However, from the resultant "lr" graph on the tensorboard, i observed that the learning rate is always increasing.

I have mentioned the number of warmup steps as 4000. Can anyone please let me know if this is the expected behavior with Noam scheduler? I have observed a similar trend for "lr" in some of the queries that are already raised. Moreover, I think there is no documentation on how to interpret the multiple graphs which includes alignment on the tensorboard.

I'm also attaching the config file, that I'm using.
config.json

Also, I'm trying to train a fastpitch model on custom datasets (using phonemes). The dataset contains 2 languages (2 speakers per language) with each speaker having ~6 hours of data. Can i know the approximate number of steps for which i need to train the model?