R2MBench/run_config.example.conf at main · AMAP-ML/R2MBench · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
# Auto-generated by r2mbench. Standalone full_eval config.
# Consume via: bash scripts/run_full_eval.sh --config run_config.local.conf ...
#
# Third-party CODE is vendored under ${SCRIPT_DIR}/third_party (relocatable, internal).
# Model WEIGHTS live under ${SCRIPT_DIR}/checkpoints (populate by copy or symlink;
# see README §7). No external absolute paths are baked in.

METRICS_ARG=all
GPU_LIST=0,1,2,3
TASK_FILTER=
NUM_VIDEO_SHARDS=10

# Runtime: single fresh conda env for both GPU metrics and the VLM family.
# (auto-filled by `bash r2mbench setup`; edit if you point at an existing env)
WORLD_SCORE_ENV=/path/to/conda/envs/r2mbench
GEMINI_PYTHON=/path/to/conda/envs/r2mbench/bin/python3

TORCH_HOME=${SCRIPT_DIR}/checkpoints/torch_hub
HF_ENDPOINT=https://hf-mirror.com
USE_TF=0
USE_TORCH=1
TRANSFORMERS_NO_TF=1

# Vendored third-party code roots (internal).
MEM_BENCH_THIRDPARTY_ROOT=${SCRIPT_DIR}/third_party
BOQ_ROOT=${SCRIPT_DIR}/third_party/Bag-of-Queries
MUTUALVPR_ROOT=${SCRIPT_DIR}/third_party/MutualVPR
LIGHTGLUE_ROOT=${SCRIPT_DIR}/third_party/LightGlue
GROUNDINGDINO_ROOT=${SCRIPT_DIR}/third_party/GroundingDINO
SAM2_ROOT=${SCRIPT_DIR}/third_party/sam2

# Model weights (internal paths under checkpoints/; bytes may be symlinks).
CHECKPOINT_ROOT=${SCRIPT_DIR}/checkpoints
CLIP_CKPT=${CHECKPOINT_ROOT}/ViT-B-32.pt
BOQ_CHECKPOINT=${CHECKPOINT_ROOT}/dinov2_12288.pth
MUTUALVPR_CHECKPOINT=${CHECKPOINT_ROOT}/MutualVPR_model_512.pth
GROUNDINGDINO_CONFIG=${GROUNDINGDINO_ROOT}/groundingdino/config/GroundingDINO_SwinT_OGC.py
GROUNDINGDINO_CKPT=${CHECKPOINT_ROOT}/groundingdino_swint_ogc.pth
SAM2_CONFIG=${SAM2_ROOT}/sam2/configs/sam2.1/sam2.1_hiera_l.yaml
SAM2_CKPT=${CHECKPOINT_ROOT}/sam2.1_hiera_large.pt

# VLM API (persistent_state family only).
API_KEY=
API_URL=

NMR_RESULT_BASE=${SCRIPT_DIR}/results/nmr

# ---- Tasks --------------------------------------------------------------
# One TASK per model x trajectory:  TASK=<name>|<video_dir>|<pose_json>
# EDIT the <video_dir> paths below to point at YOUR generated videos (README
# Step 1). The pose_json uses the reference trajectories shipped in this repo;
# prefer the exact pose you fed to the model. Task names are only used for the
# results folder and the --tasks filter.
MODEL_NAME=mymodel
TRAJECTORY_DIR=${SCRIPT_DIR}/trajectories

TASK=${MODEL_NAME}_DDDAAA_30s|/path/to/your_videos/DDDAAA|${TRAJECTORY_DIR}/trajectory_DDDAAA_30s.json
TASK=${MODEL_NAME}_WSLRRLL_30s|/path/to/your_videos/WSLRRLL|${TRAJECTORY_DIR}/trajectory_WSLRRLL_30s.json
TASK=${MODEL_NAME}_revisit_loop_30s|/path/to/your_videos/revisit_loop|${TRAJECTORY_DIR}/trajectory_revisit_loop_30s.json