Examples¶
This page is a recipe collection for SpaRRTa.
Use it when you already have the project set up and want short, runnable command patterns. If you still need installation or first-run help, start with Getting Started. If you want the reasoning behind the workflows, see User Guide.
Quick Recipes¶
First egocentric run¶
python train.py \
backbone=dino_b16 \
dataset=unreal_position \
probe=classifier probe._target_=sparrta.models.probes.EfficientProbing \
dataset.perspective=camera \
environment=forest
First allocentric run¶
python train.py \
backbone=dino_b16 \
dataset=unreal_position \
probe=classifier probe._target_=sparrta.models.probes.EfficientProbing \
dataset.perspective=human \
environment=forest
Switch to a simpler probe head¶
python train.py \
backbone=dino_b16 \
dataset=unreal_position \
probe=classifier probe._target_=sparrta.models.probes.ClassificationHead \
dataset.perspective=camera \
environment=forest
Swap the backbone¶
python train.py \
backbone=mae_b16 \
dataset=unreal_position \
probe=classifier probe._target_=sparrta.models.probes.EfficientProbing \
dataset.perspective=camera \
environment=forest
Inspect the resolved Hydra config before running¶
Data Setup Recipes¶
Synthetic benchmark data¶
Lego sim-to-real data¶
If you want to download the lego split directly:
huggingface-cli download turhancan97/SpaRRTa-Lego --repo-type dataset --local-dir ./hf_SpaRRTa-Lego
export SPARRTA_LEGO_ROOT=$(pwd)/hf_SpaRRTa-Lego/train
Attention-analysis data¶
Cache and model directories¶
Advanced Workflow Recipes¶
Leave-one-environment-out / few-shot transfer¶
Use this workflow when you want to test cross-environment generalization and adaptation.
Expected output: result CSVs for holdout-environment transfer and few-shot adaptation, plus summary tables/plots after post-processing.
Summarize leave-one-environment-out / few-shot results¶
Lego sim-to-real evaluation¶
Use this workflow when you want to evaluate synthetic-to-real transfer on the lego split.
Expected output: result CSVs comparing performance on the real-world lego evaluation.
Summarize lego results¶
Attention analysis for one environment¶
Use this workflow when you want to inspect where a frozen backbone attends and how attention flows between objects.
Expected output: attention CSVs and plots, usually under result/attention/<environment>/.
Inspect the attention-analysis config¶
Custom Data Checks¶
If you generated data locally with unreal-scene-gen, sanity-check it before trying evaluation.
Check the folder contains paired images and metadata¶
output/<environment>/<triple_id>/
├── img_0000.jpg
├── params_0000.json
├── img_0001.jpg
├── params_0001.json
└── ...
Check the files you care about¶
img_*.jpg: rendered scene imagesparams_*.json: camera and actor geometry
Quick validation checklist¶
- image files and
params_*.jsonfiles use matching indices - each environment/triple folder contains multiple valid pairs
params_*.jsonincludes camera and actor entries- your final layout matches what the evaluation repo expects, or you have a clear adaptation step
Important compatibility note¶
Custom Unreal outputs provide the core ingredients needed by the benchmark, but this page does not assume they are automatically drop-in compatible with the evaluation repo’s expected dataset layout.
Failure Examples¶
Missing SPARRTA_DATA_ROOT¶
If the main benchmark run fails immediately, verify:
Missing external backbone repo¶
If a repo-backed model fails, verify the corresponding environment variable, for example:
If you want the safest path, prefer backbones that work out of the box:
dino_b16dinov2_b14dinov2_b14_regdinov2_l14_regdinov3_timmmae_b16clip_b16_laion
Unexpected output location¶
If results appear to be missing, inspect the run config first:
Then check:
result/for experiment outputsSPARRTA_CACHE_DIRfor cached features
References¶
- Getting Started
- User Guide
- Evaluation of VFMs
- Unreal Scene Generation
- Upstream evaluation repo: turhancan97/SpaRRTa