Is there a way to turn off Weights and Biases without code changes?

[mathemakitten]

• Weights and Biases version: 0.8.19
• Python version: 3.7
• Operating System: Linux

Is there a way to totally turn off W&B without having to build if-logic? I know I can do WANDB_MODE="dryrun" but I want to turn it completely off (i.e. ignore Wandb.init(), ignore wand.save, don't even write files).

[issue-label-bot]

Issue-Label Bot is automatically applying the label question to this issue, with a confidence of 0.78. Please mark this comment with 👍 or 👎 to give our bot feedback!

Links: app homepage, dashboard and code for this bot.

[mathemakitten]

Also, is there a way to turn off W&B warnings? I'm running in dryrun mode as a workaround but still getting wandb warnings; ideally I'd like to ignore them all. (Best case scenario would be to fully turn off wandb.)

[vanpelt]

Hey @mathemakitten currently you can't completely turn off wandb.init but it's a feature we should add. You can turn off all messages from wandb by setting WANDB_SILENT=true

[jarednielsen]

@vanpelt Is there a timeline for adding this feature? I have a use case where wandb is cluttering up my filesystem even if WANDB_SILENT=true; I would like an environment variable that makes wandb act as a no-op. If there's no way to fully disable it, then I'm wary of adding wandb to my code.

[raubitsj]

@jarednielsen : This is a feature of a new experimental client library:
https://github.com/wandb/client-ng

We are working out the timeline for getting these changes into the released library. But starting next week we will have a noop mode.

The trick to the noop mode is that all the wandb methods and objects have to behave sanely so that your code execution still works like:
run = wandb.init()
run.config.value = 3
print(run.config["value"])

When the experimental client has validated and documented how to use this feature we will comment here and mark this issue with the label "c:cling-flixed". when it is released in the main library this issue will be marked closed

[JDE65]

Hey @mathemakitten currently you can't completely turn off wandb.init but it's a feature we should add. You can turn off all messages from wandb by setting WANDB_SILENT=true

I tried both directly and with os.environ... it doesn't work :-(
Is there no other way ?
THX

[vanpelt]

We just released wandb 0.10.7 which has support for the WANDB_SILENT environment variable. You can set it in code before calling wandb.init with os.environ['WANDB_SILENT']="true".

[JDE65]

Great, I'll immediately update it :-)

[JDE65]

I tried and I stll get all warnings and unuseful messages :-(

And I run the same code several times in a row, I also get trouble with the conflict between init and login

[vanpelt]

Can you share a few examples of message you're seeing? WANDB_SILENT should disable any lines beginning with **wandb: **. You need to set this env variable before any calls to wandb api's in your script.

[glenn-jocher]

I'm going to +1 this thread. I think wandb off should do what it says, and shut it off completely. My current workaround is pip uninstall wandb when I want to be sure it's not logging, since I'm confused by the variety of other options, none of which seem to simply turn it off easily.

I think this is applicable to a variety of situations, i.e. debugging small issues that may cause you to run short trainings dozens of times in a day that you don't want logged anywhere, but then easily turn it on afterward, ideally with $ wandb on.

[vanpelt]

Hey @glenn-jocher we're going to address this in the next release hopefully tomorrow. We're deprecating wandb on and wandb off, there will now be 4 commands:

• wandb online, WANDB_MODE=online or wandb.init(mode="online") - runs in online mode, the default
• wandb offline, WANDB_MODE=offline or wandb.init(mode="offline") - runs in offline mode, writes all data to disk for later syncing to a server
• wandb disabled, WANDB_MODE=disabled or wandb.init(mode="disabled") - makes all calls to wandb api's noop's, while maintaining core functionality such as wandb.config and wandb.summary in case you have logic that reads from these dicts.
• wandb enabled - sets the mode to back online

[glenn-jocher]

@vanpelt ah that sounds like a good improvement, I think this will help a lot!

[drozzy]

Yes, this was rally a pita for me for one-off tweaks/experiments, so I wrote my own if-else statements around wandb code.
The mode="disabled" would be awesome!

[glenn-jocher]

Can confirm this is working better now.

wandb enabled and wandb disabled seem to behave well and as expected with YOLOv5 :)

[vanpelt]

Awesome! We just deployed this feature with wandb==0.10.11, anyone finding this can upgrade with pip install wandb --upgrade. This version now supports the wandb enabled and wandb disabled CLI commands as well as passing mode="disabled" or mode="enabled" to wandb.init. (You can also just set the WANDB_MODE environment variable)

[tom-andersson]

@vanpelt If running a sweep, does mode in wandb.init() do anything? In my case I have mode='disabled' by default (controlled by a command line input to my train script), and I'm wondering if that will affect my sweeps.

[raubitsj]

@tom-andersson disabled mode is not compatible with sweeps as we need to know what configurations have run. The wandb.init() call is what signals our hyperparameter algorithm that a parameter set was received by an agent.

[tom-andersson]

Thank you @raubitsj, I've since edited my sweep.yaml to pass in the --wandb command to my train script which triggers mode='online' in wandb.init():

command:      
  - ${env}    
  - python3   
  - ${program}
  - "--wandb" 
  - ${args}   

I'm was previously getting no information sent to the web app when running sweeps, but now everything's running nicely!

[WilliamAshbee]

WANDB_MODE=offline
seems to make output go to stderr

[glenn-jocher]

pip uninstall wandb or wandb disabled