[core] remove wandb dependency
#92
Conversation
|
The documentation is not available anymore as the PR was closed or merged. |
trl/trainer/ppo_trainer.py
Outdated
| wandb_logs.update(stats) | ||
| wandb_logs["env/reward_mean"] = torch.mean(rewards).cpu().numpy() | ||
| wandb_logs["env/reward_std"] = torch.std(rewards).cpu().numpy() | ||
| wandb_logs["env/reward_dist"] = rewards.cpu().numpy() | ||
| wandb.log(wandb_logs) | ||
| self.accelerator.log(wandb_logs) |
There was a problem hiding this comment.
since the user can select other frameworks (e.g. tensorboard) we should also log if we are not using wandb. maybe that whole block can be wandb agnostic, what do you think?
There was a problem hiding this comment.
Sounds good, based on your suggestion I proposed dc8cda2 , here is the corresponding log: https://wandb.ai/distill-bloom/trl/runs/baw84iyv?workspace=user-
|
Thanks! I will give it a try with |
|
I can confirm everything works fine (single & multi-GPU with wandb and with |
examples/scripts/ppo-sentiment.py
Outdated
| config = PPOConfig( | ||
| model_name="lvwerra/gpt2-imdb", | ||
| learning_rate=1.41e-5, | ||
| log_with="wandb", |
There was a problem hiding this comment.
if we remove wandb as a main dependency i think we should not use it as a default. otherwise it will fail if you use the ppo trainer out of the box. what do you think?
There was a problem hiding this comment.
yes makes sense!
There was a problem hiding this comment.
Should be now addressed in d3f9231 !
What does this PR do ?
This PR removes
wandbdependency fortrlwandb run of this branch: https://wandb.ai/distill-bloom/trl/runs/360f7cdc?workspace=user-
cc @lvwerra