lanpa/tensorboardX

[for Pytorch users] Why should one use TensorboardX if torch.utils.tensorboard is available?

taras-sereda opened this issue · 2 comments

taras-sereda commented

Hey guys, great work on developing this repo! It's quite popular and apparently trusted by many people given the number of stars.

I have one important question. Why should one use TensorboardX instead of torch.utils.tensorboard?
I see that TensorboardX is tested with pytorch 1.8.1, but right now pytorch 1.10.1 is already available.

It's clear to me that this repo can be used with other DL frameworks which don't have tensorboard integration. That some code bases are already using TensorboardX. But when it comes to development of a new project in pytorch, would you use TensorboardX? If so then why exactly?

Thanks!

shouldsee commented

Is there a hidden relation between these two modules?

Seems like torch.utils.tensorboard was based on tensorboardX code as indicated by its early commits

  • ❤️1
fellhorn commented

I can tell you one reason, why people are using tensorboardX instead of torch.utils.tensorboard is the direct support for writing the tfevent files to Google Cloud Storage. In torch.utils.tensorboard relies on tensorboard for this feature, which comes with a large storage overhead if your only goal is to write the files to GCS.