-
Notifications
You must be signed in to change notification settings - Fork 3.3k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
find alternative to gfile #3242
Comments
It is used for saving to remote locations. |
I agree it would be nice to remove a dependency on tensor-board but I felt it was ok because it already was a dependency of the project. I want to note that tensorboard can be installed without tensorflow and the gfile operations work in 'compatible' mode where only local and s3 is supported. This does not require full tensorflow to be installed, just the standalone tensorboard package. The problem is there are not really any good alternatives. Projects like If there is a package I don't know about I would be happy to use it. The other option if we want to remove tensorboard support is refactor all the IO further to check if a user does have tensorboard/tensorflow installed, and if so support remote directories, and otherwise warn that only local file operations are supported. |
sure, no one is blaming you, at the time it was a great solution, just later we found this limitation... :]
Well I also doe not say explicitly remove
that is pretty much what this issue is, just to have a look around and eventually circle back that there is nothing better then what we have 🐰 |
I'm very happy to help in this as well to make sure any new solution works for our use case. At the moment my org needs hdfs and s3 support. I think the main issue I see with the current implementation is nothing preventing someone from adding IO operations to lightning which would only work locally. In that case all tests would still pass but would then crash when trying to use with a remote dir. I like that currently things should all still work if all paths are local and a user doesn't have tensorflow installed. But a better way to notify and fail with a proper message if a remote directory is passed in and the user doesn't have tf. It currently does print some messages but not sure all possible cases are caught with a good message. |
Suggestion: fsspec |
Oh! Looks promising :) thanks for the suggestion let me give that a try. |
I played with fsspec a bit. I think this is the right solution. :) Very cool. I will start working on implementing this. |
Now that we don't depend on gfile anymore, can we remove tensorboard from dependencies and just make it optional? |
yes, it is in #3132 |
Nice, that looks good! But maybe rename the PR title :) |
Hi, I am receiving |
🚀 Feature
at this moment we are using GFile for saving checkpoints which are imported from tensorboard which underhood uses TF core...
Motivation
remove the hard dependency on tensorboard as it seems tidy linked to TensorFlow and for visualization user can install it independently
Pitch
Avoid tensorboard limitations
The text was updated successfully, but these errors were encountered: