You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently datadreamer seems to rely on occasional messages to indicate the current progress/amount of time left such as:
Step '********' progress: 90% 🔄 (Estimated time left: 28 secs)
However, it would be nice to have the option of using a progress bar instead such as tqdm (which can also display other information such as loss/error rates during finetuning) that other libraries like HuggingFace Trainers use. Right now, there isn't even a way to reenable model's default progress bar if desired. For instance, the vLLM wrapper hardcodes use_tqdm=False.
The text was updated successfully, but these errors were encountered:
Thanks for the suggestion. I think I'd like to keep the progress printing the way it is. The library has options for the interval at which it prints also (default is 60 seconds).
I mostly chose this design because tqdm, while nice for interactive use on a Terminal, is awful in research settings when submitting a script to a cluster and having it write the output to a log file, where tqdm tends to output way too many lines and creates a mess in the logs due to the way it writes tot he terminal.
That being said, I have made it so you can re-enable the original logs of libraries. For example, you can see here that DataDreamer takes in hf_log which you can set hf_log=True to re-enable all Hugging Face library logs. You can also set verbose=True to re-enable logs from other libraries like VLLM. I just enabled that in the latest version of DataDreamer.
Currently datadreamer seems to rely on occasional messages to indicate the current progress/amount of time left such as:
However, it would be nice to have the option of using a progress bar instead such as tqdm (which can also display other information such as loss/error rates during finetuning) that other libraries like HuggingFace Trainers use. Right now, there isn't even a way to reenable model's default progress bar if desired. For instance, the vLLM wrapper hardcodes
use_tqdm=False
.The text was updated successfully, but these errors were encountered: