wandb init hasn't been called

• Weights and Biases version: 0.8.10
• Python version: 3.7.4
• Operating System: Windows

Description

Trying to launch training code in pytorch. Launching code as recommended with wandb run python train.py I have a wandb.init(...) at the begining of my code.

This code works perfectly fine on linux.

What I Did

Put all my code in a if __name__ == "__main__" because windows needs it to launch processes : No success

wandb run python train.py 

wandb: ERROR wandb.init hasn't been called, can't configure run
Traceback (most recent call last):
  File "train.py", line 164, in <module>
    wandb.save(PATH_WEIGHTS_EMBEDDER)
  File "C:\Users\pisne\AppData\Local\Continuum\anaconda3\lib\site-packages\wandb\__init__.py", line 493, in save
    os.symlink(abs_path, wandb_path)
OSError: symbolic link privilege not held
wandb: Program failed with code 1. Press ctrl-c to abort syncing.

[noklam]

@noklam Please do not use wandb run any more (This command will likely be removed in future releases)

Example python:

import wandb
def train():
    wandb.init()
    wandb.log(dict(metric=value))
if __name__ == "__main__":
  train()

Then run with: python train.py

ops sorry, I came in this issue as I hit the os.symbolic link issue on Window.

in my case, it is due to wandb.save()

[ierezell]

@alberduris I know it’s only a patch but you can use Popen from subprocesses in python to automate launching commands with arguments…

@raubitsj Thanks a lot for the precisions. I’m a linux power user and the unix wandb is amazing but unfornutaltely my lab granted me acces to a really powerfull Gpu… on windows…

I wanted the name of my run to be linked to parameters (like Arch_8_0.0001 etc…) which is good under linux but I will try on windows with os.environ and tell here my successes.

Thanks again for the fast and precise feedback !