google-research / google-research/pix2struct
More clarity on the warmup step before pretraining
Open
- Dominant language
- Python
- Stars
- 690
- Forks
- 61
- PR merge metrics
- No merged PRs in 30d
Description
I was unable to determine the specific number of distinct images utilized in the warmup phase from the paper. Also, although it's noted that training took place over 30k steps, the batch size for this phase isn't clearly stated. Can you offer any insights on this and any other hyperparameter that can be essential?
Contributor guide
Assessment
This issue has not been assessed yet.