Spaces:
Configuration error
Configuration error
Commit History
-
fix(RCLane) : fix option parallel mode
2206380 -
feat(RCLane): archive parallel BEV projection modes
287b358 -
feat(RCLane): switch to C++ for runtime modeling
3624d0b
-
fix(RCLane): stabilize lane decode and ONNX inference
73a2d91
-
fix(RCLane): persist GT cache in HF bucket
fa5c5c2 -
perf(RCLane): bound and cache F1 lane decoding
452dc3a -
perf(RCLane): saturate dual H200 train and eval
71e78c0 -
perf(RCLane): speed up CARLA F1 eval ~14x (downscaled IoU + cached masks)
Root cause: _match_count rasterized each pred/gt lane at full original resolution (CARLA 1920x1080 ~2M px) and ran logical_or/and per pred*gt pair; with an untrained model decode emits many spurious lanes, so P*G full-res mask ops dominated eval time (~130s/500 imgs just for IoU). Fix: rasterize each lane once on a downscaled canvas (--f1-eval-scale, default 0.25 -> 16x fewer px) and compute IoU as count_nonzero(a & b) with precomputed areas (union = area_a + area_b - inter), dropping logical_or entirely. Lane-IoU is a ratio so downscaling pred+gt+width together leaves the metric ~unchanged.8932089 -
chore(RCLane): add HF Jobs runner for CARLA (Space code + F1 eval + per-epoch upload)
14c6142 -
feat(RCLane): add resume, F1 eval, full logging & per-epoch HF upload for CARLA
Ports the full-featured training loop (previously only on the CULane work) to the CARLA branch: resumable checkpoints (--resume), CULane-style lane-IoU F1 validation (--eval-list label_val.json --eval-f1), per-epoch + best/last checkpoints, detailed step/epoch logging (data/step time, img/s, per-term losses), AMP, and best-effort upload of every epoch's checkpoint to an HF model repo (--push-to). CARLA eval reads a separate split file (label_val.json) instead of reusing the training label.9911299
-
feat(RCLane): add CARLA dataset
ecf4d58 -
feat(RCLane): add PyTorch port of RCLane
5d3daa9 -
Initial commit
Lương Ngọc HuyUnverified signature7c84f5f
