| # Codebase Rules |
|
|
| This repo has accumulated several near-duplicate implementations and a few |
| hidden shape/device assumptions. Use the rules below for any further cleanup or |
| feature work. |
|
|
| ## Ownership and layout |
|
|
| - Put shared GP-NeRF / Semantic-Frustum transformer pieces in |
| `gpnerf/transformer_blocks.py`. |
| - Keep `gpnerf/transformer_network.py` focused on GP-NeRF assembly logic. |
| - Keep `s_frustum/transformer.py` focused on frustum-specific logic only. |
| - Prefer extending existing modules over copy-pasting whole files. |
|
|
| ## Device and distributed rules |
|
|
| - Never hardcode `.cuda()` in model or dataloader utilities; use `.to(device)`. |
| - Never hardcode output directories; resolve them from `args.rootdir`. |
| - Convert to `SyncBatchNorm` only inside the distributed wrapping path, before |
| `DistributedDataParallel`. |
|
|
| ## Shape and indexing rules |
|
|
| - Do not hardcode image sizes such as `320x240` or feature sizes such as |
| `160x120`; infer them from tensors. |
| - Assume flattened pixel indices are zero-based unless a caller explicitly says |
| otherwise. |
| - When projecting semantic features, derive any downsampled index mapping from |
| the current feature map stride instead of dataset-specific constants. |
|
|
| ## Transformer rules |
|
|
| - Use the fused PyTorch SDPA path for standard self-attention whenever possible. |
| - If a block needs explicit attention weights for density sampling, keep that |
| path isolated from the main attention implementation. |
| - Avoid channel-wise hacks like `repeat_interleave(8)`; compute grouping from |
| the actual tensor shapes. |
|
|
| ## Validation rules |
|
|
| - Any change touching render paths should be covered by at least one synthetic |
| shape test. |
| - Any change touching sampling or semantic distillation should be checked for: |
| device placement, index bounds, and `N_importance == 0`. |
|
|
