Using tensorboard.dev

Wanted to mention about tensorboard.dev - I had been meaning to explore this as an option for sharing training run details, but didn’t get round to trying it until this evening.

It lets you upload your tensorboard logs, so people can view them and collaborate.

There is a limit to it, in that it currently doesn’t support images or audio, so for TTS it’s only able to show the Scalars data. In spite of that, it could still be handy to give people an idea how their own training compares to that of others.

Here’s an example from a recent training run of mine using voice data from M-AILABS:

It’s easy to use (instructions here). You need a Google account to be able to authorise the posting, although no account is needed to view data. Only significant point to check, so far as I can see, is that the log data is then public, so be sure you are indeed okay with sharing it, but if your dataset it public then this seems unlikely to be a concern and it’s not surfacing the actual dataset anyway, it’s the training log info - but how you feel about that is obviously a decision to make yourself.

3 Likes

Thinking about it, for the WaveGrad logs I believe it would show the Histograms and Distributions that are produced as well (but still no audio or the spectrograms)

Hey @nmstoker.
Thanks for sharing this. I tried it with my current wavegrad training.

1 Like