| # Environment of the launchers. Copy, edit the paths, then `source env_template.sh` before running a stage. | |
| # Where stage outputs go: co-occurrence statistics, raw renders and audits, clean / strict / released trees. | |
| export WORK_ROOT="${WORK_ROOT:-$PWD/strange_bench_work}" | |
| # Stage 0: COCO 2017 train annotations (https://cocodataset.org). | |
| export COCO_TRAIN_ANN="${COCO_TRAIN_ANN:-/path/to/coco/annotations/instances_train2017.json}" | |
| # Stages 1, 2, 3, 5, 7: Hugging Face cache for Qwen/Qwen2.5-3B-Instruct, stabilityai/stable-diffusion-3.5-large | |
| # (gated: accept its license on the Hub and log in first), facebook/sam3, google/owlv2-large-patch14-ensemble and | |
| # bert-base-uncased (text encoder of Grounding DINO). | |
| export HF_HOME="${HF_HOME:-$HOME/.cache/huggingface}" | |
| # A TRANSFORMERS_CACHE or HF_HUB_CACHE left in the shell would override HF_HOME for transformers. | |
| unset TRANSFORMERS_CACHE HF_HUB_CACHE HUGGINGFACE_HUB_CACHE | |
| # Stages 2 and 5 (SAM 3): the BPE vocabulary shipped with the SAM 3 package (sam3/assets/bpe_simple_vocab_16e6.txt.gz). | |
| export SAM3_BPE="${SAM3_BPE:-/path/to/sam3/assets/bpe_simple_vocab_16e6.txt.gz}" | |
| # Stage 3 (lean audit, MMDetection 3.x environment): Grounding DINO Swin-B config of MMDetection and its checkpoint | |
| # https://download.openmmlab.com/mmdetection/v3.0/grounding_dino/grounding_dino_swin-b_finetune_16xb2_1x_coco/grounding_dino_swin-b_finetune_16xb2_1x_coco_20230921_153201-f219e0c0.pth | |
| export GDINO_CONFIG="${GDINO_CONFIG:-/path/to/mmdetection/configs/grounding_dino/grounding_dino_swin-b_finetune_16xb2_1x_coco.py}" | |
| export GDINO_CKPT="${GDINO_CKPT:-/path/to/grounding_dino_swin-b_finetune_16xb2_1x_coco_20230921_153201-f219e0c0.pth}" | |