File size: 2,364 Bytes
a8d795a | 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 | #!/usr/bin/env bash
# Score with the project's own OWL+SAM+DINO relevance eval.
#
# Each method writes under its own directory name, and the evaluator resolves
# runs as {profile}/{adapter}/{case}/{method}, so the method is derived from the
# profile tag rather than passed in: a mixed-method batch would otherwise need
# one invocation per method with the right name remembered by hand.
set -euo pipefail
root=/home/n2501108a/KV/experiments/rope_reindex_allmethods_20260824
repo=/home/n2501108a/KV/forcingkv_private_worktrees/Lingbot-World-2
frozen=/home/n2501108a/KV/experiments/lingbot_priority10_onlygen_allmethods_seed1_20260818
benchmark=$frozen/benchmark
metadata=$benchmark/metadata/benchmark_cases.jsonl
diff_python=/home/n2501108a/miniconda3/envs/diff/bin/python
export CONTEXT_MEMORY_MODEL_ROOT=/home/n2501108a/KV/forcingkv_private/models
export PATH="$(dirname "$diff_python"):$PATH"
gpu=${LINGBOT_EVAL_GPU:-0}
mkdir -p "$root/eval_dino" "$root/logs"
# The budget sweep writes to its own tree, so the runs directory is selectable
# rather than hardcoded; otherwise sweep profiles would be looked up under the
# main batch's runs/ and silently found missing.
runs_dir=${LINGBOT_RUNS_DIR:-runs}
if [[ $# -gt 0 ]]; then
tags=("$@")
else
mapfile -t tags < <(cd "$root/$runs_dir" && ls -d */ 2>/dev/null | tr -d '/')
fi
if [[ ${#tags[@]} -eq 0 ]]; then
echo "no run directories under $root/runs" >&2
exit 2
fi
for tag in "${tags[@]}"; do
if [[ -f "$root/eval_dino/$tag/.complete" ]]; then
echo "skip $tag (already scored)"
continue
fi
# Profiles are named <method>_seed<N>.
method=${tag%_seed*}
spec=$root/eval_dino/$tag.profiles.yaml
printf 'profiles:\n - {profile: %s, method: %s, label: %s}\n' \
"$tag" "$method" "$tag" >"$spec"
echo "scoring $tag (method=$method)"
(cd "$repo"; CUDA_VISIBLE_DEVICES="$gpu" ./cmctl evaluate --model lingbot \
--benchmark-root "$benchmark" --metadata "$metadata" \
--annotations "$metadata" \
--runs-roots "$root/$runs_dir/$tag" --eval-root "$root/eval_dino/$tag" \
--profile-spec "$spec" --methods "$method" --video-name continuation.mp4 \
--device cuda:0 --mask-mode sam --scoring-path owl_sam_dino \
--frame-stride 1 --no-overlays --no-comparison-videos) \
>"$root/logs/eval_$tag.log" 2>&1 && touch "$root/eval_dino/$tag/.complete"
done
echo "done"
|