facebookresearch / facebookresearch/sam2
Why not training on dynamic input sizes?
Open
- Dominant language
- Jupyter Notebook
- Stars
- 19.9k
- Forks
- 2.5k
- PR merge metrics
- No merged PRs in 30d
Description
Since many datasets, real world videos, might not have too large resolution. Why not consider traing the model with dynamic sizes? Like find the closest shape mutiple to 32? Since the Hiera can support rectangle and dynamic input.
Contributor guide
Assessment
This issue has not been assessed yet.