TestRandomization / code /case1 /runner /run_campaign.sh
harryzhang1018's picture
Add reproduction code and README
0c3dd1d verified
Raw History Blame Contribute Delete
3.04 kB
#!/bin/bash
# Launch one long-lived, network-isolated container that runs campaign.py with
# a fixed worker pool. Usage (host):
# bash run_campaign.sh EPISODES.json OUTDIR WORKERS [NAME_SUFFIX]
# EPISODES.json and OUTDIR must be inside the directory that contains runner/ (code/case1).
# Any extra arguments after the suffix are passed to campaign.py.
# tini (--init) reaps orphaned physics servers and forwards SIGTERM; TR_HOST_NAME enters the host fingerprint
# written into every manifest and summary.json.
# The exit code of campaign.py is the exit code of this script (0 complete, 3 stalled, 4 infrastructure errors
# left after retries, 5 worker start failure, 130/143 interrupted).
# Every path is resolved from this file's location (root = runner/..). The pinned image is run by its
# content-addressed image Id, which survives `docker save | docker load` (a RepoDigest does not) and never triggers
# a pull; the launcher refuses to run if that image is absent. TR_IMAGE_DIGEST (recorded in the code digest) keeps
# the RepoDigest label. TR_MAX_WORKERS (default 12) is forwarded to campaign.py's worker cap.
set -euo pipefail
HERE="$(cd "$(dirname "$0")" && pwd)"
C="$(cd "$HERE/.." && pwd)"
EPISODES="$(realpath "$1")"; OUT="$(realpath -m "$2")"; WORKERS="${3:-1}"; SUFFIX="${4:-$$}"
shift $(( $# < 4 ? $# : 4 ))
IMAGE="uwsbel/ros@sha256:60118a64d8e751c1adafef1abebebbc504a5814d63eae5bb46120de5f7b4554c"
IMAGE_ID="sha256:73882e33184a5e9a586b5ce5167731ce0c569337104b273b05709a607730fa29"
if ! docker image inspect "$IMAGE_ID" > /dev/null 2>&1; then
echo "pinned image $IMAGE_ID ($IMAGE) is not present on $(hostname); load it first (docker load) - no pull is attempted" >&2; exit 2
fi
case "$EPISODES" in "$C"/*) ;; *) echo "episodes file must be under $C" >&2; exit 2;; esac
case "$OUT" in "$C"/*) ;; *) echo "output dir must be under $C" >&2; exit 2;; esac
mkdir -p "$OUT"
docker run --rm --init --network none --name "tr-c1-runner-$SUFFIX" \
--user "$(id -u):$(id -g)" -e HOME=/tmp -e TR_IMAGE_DIGEST="$IMAGE" -e TR_IMAGE_ID="$IMAGE_ID" \
-e TR_HOST_NAME="$(hostname)" -e TR_MAX_WORKERS="${TR_MAX_WORKERS:-12}" \
--cpus "$WORKERS" -v "$C:$C" -w "$HERE" --entrypoint /bin/bash "$IMAGE_ID" -c "
set -e
source /opt/ros/humble/setup.bash
source /chrono_ros_interfaces/install/setup.bash
source /sbel/Desktop/ros_ws/install/setup.bash
export KERAS_BACKEND=torch CUDA_VISIBLE_DEVICES='' OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1
export PYTHONWARNINGS=ignore PYTHONDONTWRITEBYTECODE=1
# CPU-dispatch pins. Verified bitwise no-ops on the host that ran the batches (AMD Ryzen 7 5800X). With
# MKL_CBWR=COMPATIBLE, an Intel host (i9-14900KF) reproduced a 168-episode reference set bytewise.
export ATEN_CPU_CAPABILITY=avx2 MKL_CBWR=COMPATIBLE OPENBLAS_CORETYPE=Haswell ONEDNN_MAX_CPU_ISA=AVX2
export NPY_DISABLE_CPU_FEATURES='AVX512F AVX512CD AVX512_KNL AVX512_KNM AVX512_SKX AVX512_CLX AVX512_CNL AVX512_ICL AVX512_SPR'
exec python3 campaign.py --episodes '$EPISODES' --out '$OUT' --workers '$WORKERS' $*
"