[FYI] Bug in R2+1D implementation
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 17.9k
- Forks
- 7.3k
- Avg merge
- 1d 15h
- Merged PRs (30d)
- 13
Description
I found that there is a lack of clarity in the original R2+1D paper and official code implementation for models utilizing BottleNeck layers, which makes it impossible to transfer weights from large pretrained models in C2 to the models implemented in torchvision.
For background info see the question in their repo.
The fix is very straightforward (change bottleneck midplanes computation), but the question is whether we should do it, which I suspect should be based on author's answer. I'm leaving this here just so that people are aware of it.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reading the issue and the linked VMZ question, then compare the R2+1D paper and official implementation with the torchvision models. Resolve whether the bottleneck midplanes computation should change before attempting a fix. Done means the intended behavior is decided and the weight-transfer incompatibility is addressed or clearly documented.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- computer-vision, machine-learning
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100