Wandb - Selected runs are not logging media for the key objective/*, but instead are logging values of type list.

[gjmulder]

Getting this error in wandb when running either:

./examples/sentiment/scripts/gpt2-sentiment.py
./examples/sentiment/scripts/t5-sentiment.py

It looks like the data are lists of tensors. From the TRL Showcase it should be displaying histograms?

EDIT: I'm using accelerate across 4 GPUs, but the lists look to be of length 256, which at a guess is the batch size?

[lvwerra]

Yes, usually list of tensors are converted to histograms automatically (see rewards or ratios). Not sure what happened there but if you want to take a look that would be much appreciated.

[gjmulder]

I'm not too familiar with Wandb, so I had wondered if it could handle a list of tensors. env/reward_dist is rendering correctly as a histogram which is what made me suspect that tensors weren't supported.

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.