As main.py has many command line arguments, retyping it on terminal every time becomes cumbersome.
We prepared inference_base.sh, which can be used as a template for running experiments.
Here are some samples of how it can be used:
# Our IROS submission configuration
./scripts/inference_base.sh 0 821 0 valid_unseen testrun mlm mlmscore_equal "aggregate_sum sem_search_all spatial_norm temperature_annealing new_obstacle_fn no_slice_replay" 1 240 1 high local_adjustment 0.5 9
# Random collocation map
./scripts/inference_base.sh 0 821 0 valid_unseen random none mlmscore_equal "aggregate_sum spatial_norm sem_search_all new_obstacle_fn no_slice_replay" 1 240 1 high
# FILM
./scripts/inference_base.sh 0 821 0 valid_unseen cnn cnn mlmscore_equal "aggregate_sum spatial_norm sem_search_all new_obstacle_fn no_slice_replay" 1 240 1 high
# Ground truth language instruction for ablation studies
./scripts/inference_base.sh 0 821 0 valid_unseen testrun mlm mlmscore_equal "aggregate_sum sem_search_all spatial_norm temperature_annealing new_obstacle_fn no_slice_replay" 1 240 1 gt local_adjustment 0.5 9
# Low level instruction for ablation studies
./scripts/inference_base.sh 0 821 0 valid_unseen testrun mlm mlmscore_equal "aggregate_sum sem_search_all spatial_norm temperature_annealing new_obstacle_fn no_slice_replay" 1 240 1 low local_adjustment 0.5 9We prepared convenience scripts other than inference_base.sh for following cases.
| name | Purpose |
|---|---|
inference_debug.sh |
For debugging. Can set set_trace and debug interactively. |
inference_gtSemDepth.sh |
For ablation study. Use ground truth depth and instance segmentation. |
inference_manual.sh |
For debugging. Manually run the agent. |
inference_pics.sh |
Output various information (agent view, segmentation/depth estimation, etc.) during inference. |