Ablation testing of architecture
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 18
- Forks
- 28
- Avg merge
- 2d 2h
- Merged PRs (30d)
- 7
Description
The regression branch introduces several changes over the base architecture: a dilated convolutional heatmap predictor after the final residual layer; reflect padding on all Conv1D layers to avoid edge effects; CoordConv (replacing the first Conv1D) to allow the model to learn translation dependence; CoordConv stride=1 to reduce downsampling to 16× for heatmap fidelity; and a temporary switch to ResNet18 for memory. Determine which of these are actually necessary for merger time accuracy, and whether simpler alternatives (e.g. predicting a point estimate with Gaussian NLL loss rather than a full heatmap) are competitive.
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 comparing the regression branch with the base architecture, focusing on the dilated heatmap predictor, reflect padding, CoordConv changes, stride, and ResNet18 switch. Run controlled ablations for each change and compare merger-time accuracy with the Gaussian-NLL point-estimate alternative. Done means identifying which changes are necessary and documenting competitive simpler options.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Refactor
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Needs clarification
- Newbie friendliness
- 35/100