Skip to content

eval.md

Latest commit

 

History

History
27 lines (18 loc) · 1.52 KB

eval.md

File metadata and controls

27 lines (18 loc) · 1.52 KB

Evaluation instructions

Test configs

  • Folder: config/natural_motions/leveltestD/
  • Sample config: config/natural_motions/leveltestD/swan/ABASTCVX_556_765/prompt_0000/motion_0001.yaml
    (Same structure as other leveltestD configs: .../leveltestD/<object>/<id>/prompt_<n>/motion_<m>.yaml.)

Running evaluation

Run box optimization over the leveltestD set using the command below (adjust --output-path / --validate_dirname as needed, or try a few samples by adding --val_start 0 --val_stop 2):

optim_timestamp="2026-03-01/00-00-00"
python3 -m bin.Fwd_CmdTrailBlazer --config config/natural_motions/leveltestD --run_config config/run_opt_fwd.yaml --validate --validate_dirname evalD --val_model_name optim --shared_config config/common_shared.yaml --timestamp ${optim_timestamp} --lr 1e-2 --n_opt_iterations 5 --bb_deviate_lambda 0.1 --outside_bbox_loss_scale 10 --width 320 --height 320 --sigma_strength 0.03 --off_normalize_mask --set_global_deterministic

Run box trailblazer over the leveltestD set

tbl_timestamp="2026-03-01/00-00-10"
python3 -m bin.Fwd_CmdTrailBlazer --config config/natural_motions/leveltestD --run_config '' --validate --validate_dirname evalD --val_model_name trailblazer_origin --shared_config config/common_shared.yaml --timestamp ${tbl_timestamp}  --width 320 --height 320

Outputs are written under output/<validate_dirname>/<model_type>/natural_motions/leveltestD/... (videos, gifs, etc.).

TODO: Metrics / aggregation scripts can be documented here when added.