Buckets:

AsianPlayer
/

TriDrive

Files

xet

AsianPlayer/TriDrive / code /eval /run_joint_evals_final_ablations.sh

AsianPlayer

16 days ago

download

raw

2.79 kB

#!/usr/bin/env bash
# cwm3/eval/run_joint_evals_final_ablations.sh — E3 (no P) and E5 (naive_fresh) pairs of the final assembly (2026-09-11), CPU; layout guards on.
#   bash eval/run_joint_evals_final_ablations.sh <main_variant: all_minus_road2drv|all> "0 1 2" [render=1]
# Tags carry the EXPLICIT variant short name (allminus | all), never "main", so a rule change cannot alias two different B runs under one tag.
# Per seed k (identical rows; driver populations from the v2 causal kin targets), M = allminus | all:
#   finalc_s<k>_<M>noP_vs_indepnoP     I1_<main>_s<k>_noP_final vs I1_indep_s<k>_noP_final   — does I still help without the P chains on E? (E3)
#   finalc_s<k>_<M>_vs_<M>noP          I1_<main>_s<k>_final vs I1_<main>_s<k>_noP_final       — what the P head buys / costs the shared endpoints (E3)
#   finalc_s<k>_indep_vs_indepnoP      I1_indep_s<k>_final vs I1_indep_s<k>_noP_final         — pooled_P vs pooled base, assembled (no training)
#   finalc_s<k>_naivefresh_vs_indep    I1_naive_fresh_s<k>_final vs I1_indep_s<k>_final       — E5: from-scratch fusion vs the pretrained independent branches
#   finalc_s<k>_naivefresh_vs_<M>      I1_naive_fresh_s<k>_final vs I1_<main>_s<k>_final      — E5 vs the main model
set -u
cd "$(dirname "$0")/.."
PY=python
export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4
MAIN="$1"; SEEDS="${2:-0 1 2}"
SH=all; [[ "$MAIN" == all_minus_road2drv ]] && SH=allminus
have_tag() { grep -q ",$1,$2,test," results/paired_compare.csv 2>/dev/null; }
done_run() { [[ -f results/runs/$1/PRED_DONE && -f results/runs/$1/TRAIN_DONE ]]; }
pc() { local tag="$1" a="$2" b="$3"; shift 3
  done_run "$a" || { echo "skip $tag ($a not finished)"; return; }; done_run "$b" || { echo "skip $tag ($b not finished)"; return; }
  for reg in manual engaged; do have_tag "$tag" "$reg" && continue; echo "== $tag $reg"
    $PY eval/paired_compare_joint.py --regime "$reg" --a "results/runs/$a" --b "results/runs/$b" --tag "$tag" --kin-targets causal_routes --kin-version v2 "$@" 2>&1 | grep -E "appended|Error|Traceback|File \"|LAYOUT GUARD"; done; }
for s in $SEEDS; do
  M=I1_${MAIN}_s${s}_final; I=I1_indep_s${s}_final; MN=I1_${MAIN}_s${s}_noP_final; IN=I1_indep_s${s}_noP_final; NF=I1_naive_fresh_s${s}_final
  pc "finalc_s${s}_naivefresh_vs_indep" $NF $I
  pc "finalc_s${s}_indep_vs_indepnoP" $I $IN
  pc "finalc_s${s}_${SH}noP_vs_indepnoP" $MN $IN
  pc "finalc_s${s}_${SH}_vs_${SH}noP" $M $MN
  pc "finalc_s${s}_naivefresh_vs_${SH}" $NF $M
done
if [[ "${3:-1}" == "1" ]]; then
  $PY eval/report_joint.py --stage finalc --a ${SH}noP --b indepnoP --claim B2 --seeds ${SEEDS// /,} --extra ${SH}:${SH}noP indep:indepnoP naivefresh:indep naivefresh:${SH} --out results/tables/stage_final_ablations.md 2>&1 | grep -E "written|verdict"
fi
echo FINAL_ABL_EVALS_DONE

Xet Storage Details

Size:: 2.79 kB
Xet hash:: 66de8505406f430d9f030ab1e0580b8bfe4e58c9dda640857b83b91cc26b2ab1

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.

	#!/usr/bin/env bash
	# cwm3/eval/run_joint_evals_final_ablations.sh — E3 (no P) and E5 (naive_fresh) pairs of the final assembly (2026-09-11), CPU; layout guards on.
	# bash eval/run_joint_evals_final_ablations.sh <main_variant: all_minus_road2drv\|all> "0 1 2" [render=1]
	# Tags carry the EXPLICIT variant short name (allminus \| all), never "main", so a rule change cannot alias two different B runs under one tag.
	# Per seed k (identical rows; driver populations from the v2 causal kin targets), M = allminus \| all:
	# finalc_s<k>_<M>noP_vs_indepnoP I1_<main>_s<k>_noP_final vs I1_indep_s<k>_noP_final — does I still help without the P chains on E? (E3)
	# finalc_s<k>_<M>_vs_<M>noP I1_<main>_s<k>_final vs I1_<main>_s<k>_noP_final — what the P head buys / costs the shared endpoints (E3)
	# finalc_s<k>_indep_vs_indepnoP I1_indep_s<k>_final vs I1_indep_s<k>_noP_final — pooled_P vs pooled base, assembled (no training)
	# finalc_s<k>_naivefresh_vs_indep I1_naive_fresh_s<k>_final vs I1_indep_s<k>_final — E5: from-scratch fusion vs the pretrained independent branches
	# finalc_s<k>_naivefresh_vs_<M> I1_naive_fresh_s<k>_final vs I1_<main>_s<k>_final — E5 vs the main model
	set -u
	cd "$(dirname "$0")/.."
	PY=python
	export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4
	MAIN="$1"; SEEDS="${2:-0 1 2}"
	SH=all; [[ "$MAIN" == all_minus_road2drv ]] && SH=allminus
	have_tag() { grep -q ",$1,$2,test," results/paired_compare.csv 2>/dev/null; }
	done_run() { [[ -f results/runs/$1/PRED_DONE && -f results/runs/$1/TRAIN_DONE ]]; }
	pc() { local tag="$1" a="$2" b="$3"; shift 3
	done_run "$a" \|\| { echo "skip $tag ($a not finished)"; return; }; done_run "$b" \|\| { echo "skip $tag ($b not finished)"; return; }
	for reg in manual engaged; do have_tag "$tag" "$reg" && continue; echo "== $tag $reg"
	$PY eval/paired_compare_joint.py --regime "$reg" --a "results/runs/$a" --b "results/runs/$b" --tag "$tag" --kin-targets causal_routes --kin-version v2 "$@" 2>&1 \| grep -E "appended\|Error\|Traceback\|File \"\|LAYOUT GUARD"; done; }
	for s in $SEEDS; do
	M=I1_${MAIN}_s${s}_final; I=I1_indep_s${s}_final; MN=I1_${MAIN}_s${s}_noP_final; IN=I1_indep_s${s}_noP_final; NF=I1_naive_fresh_s${s}_final
	pc "finalc_s${s}_naivefresh_vs_indep" $NF $I
	pc "finalc_s${s}_indep_vs_indepnoP" $I $IN
	pc "finalc_s${s}_${SH}noP_vs_indepnoP" $MN $IN
	pc "finalc_s${s}_${SH}_vs_${SH}noP" $M $MN
	pc "finalc_s${s}_naivefresh_vs_${SH}" $NF $M
	done
	if [[ "${3:-1}" == "1" ]]; then
	$PY eval/report_joint.py --stage finalc --a ${SH}noP --b indepnoP --claim B2 --seeds ${SEEDS// /,} --extra ${SH}:${SH}noP indep:indepnoP naivefresh:indep naivefresh:${SH} --out results/tables/stage_final_ablations.md 2>&1 \| grep -E "written\|verdict"
	fi
	echo FINAL_ABL_EVALS_DONE