scripts/run_dense_cotracker.sh

#!/bin/bash
# Dense CoTracker launcher for LIBERO.
#
# Modes:
# 1) Single shard (backward compatible):
#    ./scripts/run_dense_cotracker.sh <CUDA_DEVICE> <START_EP> <END_EP>
#
# 2) 8-GPU (or N-GPU) parallel over full dataset:
#    ./scripts/run_dense_cotracker.sh --all
#
# Optional env overrides for --all:
#   GPU_IDS="0,1,2,3,4,5,6,7"
#   SHARD_TARGET_PREFIX=/mnt/kevin/data/physical-intelligence/libero_cotracker_densegrid
#   SHARD_LOG_DIR=/mnt/kevin/code/wmrl/howard-branch/openpi/results/dense_cotracker_logs
#
# Notes:
# - --all writes one dataset per GPU shard:
#     ${SHARD_TARGET_PREFIX}_gpu{ID}_{START}_{END}
# - This avoids multi-process write races on a single target dataset.

set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
# shellcheck source=openpi_cache_env.sh
source "${SCRIPT_DIR}/openpi_cache_env.sh"

SOURCE_REPO_ID=${SOURCE_REPO_ID:-/mnt/kevin/data/physical-intelligence/libero}
TARGET_REPO_ID=${TARGET_REPO_ID:-/mnt/kevin/data/physical-intelligence/libero_with_tracks_mesh_cotracker}
SHARD_TARGET_PREFIX=${SHARD_TARGET_PREFIX:-/mnt/kevin/data/physical-intelligence/libero_cotracker_densegrid}
COTRACKER_CHECKPOINT=${COTRACKER_CHECKPOINT:-/mnt/kevin/vlm_models/cotracker/scaled_offline.pth}
HORIZON=${HORIZON:-16}
MESH_SOURCE=${MESH_SOURCE:-dataset}
MESH_REPO_ID=${MESH_REPO_ID:-/mnt/kevin/data/physical-intelligence/libero_mesh_dynamic_plus_wrist25}
GPU_IDS=${GPU_IDS:-0,1,2,3,4,5,6,7}
SHARD_LOG_DIR=${SHARD_LOG_DIR:-/mnt/kevin/code/wmrl/howard-branch/openpi/results/dense_cotracker_logs}
LIBERO_CODE_ROOT=${LIBERO_CODE_ROOT:-/mnt/kevin/code/wmrl/howard-branch/openpi/third_party/libero}
EXTRA_LIBERO_PATH=${EXTRA_LIBERO_PATH:-/mnt/kevin/code/wmrl/howard-branch/openpi/third_party/libero/libero}
ROBOSUITE_ROOT=${ROBOSUITE_ROOT:-/mnt/kevin/code/ATM/third_party/robosuite}
OPENPI_ROOT=/mnt/kevin/code/wmrl/howard-branch/openpi

cd "$OPENPI_ROOT"

if [ "${UV_NO_SYNC:-0}" = "1" ]; then
  export UV_NO_SYNC=1
else
  unset UV_NO_SYNC || true
fi

# robosuite warning is non-fatal; setup once if helper exists and no private macros file is present.
if [ -f "$ROBOSUITE_ROOT/robosuite/scripts/setup_macros.py" ] && [ ! -f "$ROBOSUITE_ROOT/robosuite/macros_private.py" ]; then
  uv run python "$ROBOSUITE_ROOT/robosuite/scripts/setup_macros.py" >/dev/null 2>&1 || true
fi

run_one_shard() {
  local cuda_device="$1"
  local start_ep="$2"
  local end_ep="$3"
  local target_repo_id="$4"
  local log_file="${5:-}"

  export CUDA_VISIBLE_DEVICES="$cuda_device"
  export LIBERO_PATH="$LIBERO_CODE_ROOT"
  export PYTHONPATH="$OPENPI_ROOT/third_party:$LIBERO_CODE_ROOT:$EXTRA_LIBERO_PATH:$ROBOSUITE_ROOT${PYTHONPATH:+:$PYTHONPATH}"

  local -a cmd=(
    uv run python scripts/lerobot_preprocess_cotracker.py
    --source-repo-id "$SOURCE_REPO_ID"
    --target-repo-id "$target_repo_id"
    --cotracker-checkpoint "$COTRACKER_CHECKPOINT"
    --horizon "$HORIZON"
    --mesh-source "$MESH_SOURCE"
    --mesh-repo-id "$MESH_REPO_ID"
    --overwrite
    --start-episode "$start_ep"
    --end-episode "$end_ep"
    --libero-root "$LIBERO_CODE_ROOT"
    --libero-data "$LIBERO_CODE_ROOT"
    --extra-libero-path "$EXTRA_LIBERO_PATH"
  )

  echo "[run_dense_cotracker] CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES START_EP=$start_ep END_EP=$end_ep TARGET_REPO_ID=$target_repo_id"
  if [ -n "$log_file" ]; then
    "${cmd[@]}" >"$log_file" 2>&1
  else
    exec "${cmd[@]}"
  fi
}

if [ "${1:-}" != "--all" ]; then
  CUDA_DEVICE=${1:-0}
  START_EP=${2:-212}
  END_EP=${3:-213}
  run_one_shard "$CUDA_DEVICE" "$START_EP" "$END_EP" "$TARGET_REPO_ID"
  exit 0
fi

# --all mode
mkdir -p "$SHARD_LOG_DIR"
IFS=',' read -r -a GPUS_RAW <<< "$GPU_IDS"
GPUS=()
for g in "${GPUS_RAW[@]}"; do
  g_trimmed="$(echo "$g" | xargs)"
  if [ -n "$g_trimmed" ]; then
    GPUS+=("$g_trimmed")
  fi
done
NUM_GPUS=${#GPUS[@]}
if [ "$NUM_GPUS" -lt 1 ]; then
  echo "[run_dense_cotracker] ERROR: GPU_IDS produced zero workers."
  exit 1
fi

TOTAL_EPS=$(uv run python - <<PY
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds=LeRobotDataset("$SOURCE_REPO_ID")
print(len(ds.meta.episodes))
PY
)

if ! [[ "$TOTAL_EPS" =~ ^[0-9]+$ ]]; then
  echo "[run_dense_cotracker] ERROR: failed to read episode count from SOURCE_REPO_ID=$SOURCE_REPO_ID"
  exit 1
fi

echo "[run_dense_cotracker] --all mode"
echo "[run_dense_cotracker] SOURCE_REPO_ID=$SOURCE_REPO_ID TOTAL_EPS=$TOTAL_EPS"
echo "[run_dense_cotracker] GPU_IDS=$GPU_IDS NUM_GPUS=$NUM_GPUS"
echo "[run_dense_cotracker] MESH_SOURCE=$MESH_SOURCE"
echo "[run_dense_cotracker] MESH_REPO_ID=$MESH_REPO_ID"
echo "[run_dense_cotracker] SHARD_TARGET_PREFIX=$SHARD_TARGET_PREFIX"
echo "[run_dense_cotracker] SHARD_LOG_DIR=$SHARD_LOG_DIR"

PIDS=()
NAMES=()
for idx in "${!GPUS[@]}"; do
  gpu="${GPUS[$idx]}"
  start=$(( idx * TOTAL_EPS / NUM_GPUS ))
  end=$(( (idx + 1) * TOTAL_EPS / NUM_GPUS ))
  shard_target="${SHARD_TARGET_PREFIX}_gpu${gpu}_${start}_${end}"
  shard_log="${SHARD_LOG_DIR}/gpu${gpu}_${start}_${end}.log"

  # Skip empty range guard.
  if [ "$end" -le "$start" ]; then
    echo "[run_dense_cotracker] skip gpu=$gpu empty range [$start,$end)"
    continue
  fi

  (
    run_one_shard "$gpu" "$start" "$end" "$shard_target" "$shard_log"
  ) &
  pid=$!
  PIDS+=("$pid")
  NAMES+=("gpu=${gpu} range=[${start},${end}) target=${shard_target}")
  echo "[run_dense_cotracker] launched pid=$pid ${NAMES[-1]} log=$shard_log"
done

FAIL=0
for i in "${!PIDS[@]}"; do
  pid="${PIDS[$i]}"
  name="${NAMES[$i]}"
  if wait "$pid"; then
    echo "[run_dense_cotracker] done: $name"
  else
    echo "[run_dense_cotracker] FAILED: $name"
    FAIL=1
  fi
done

if [ "$FAIL" -ne 0 ]; then
  echo "[run_dense_cotracker] one or more shards failed. Check logs in $SHARD_LOG_DIR"
  exit 1
fi

echo "[run_dense_cotracker] all shards finished successfully."
echo "[run_dense_cotracker] produced shard datasets with prefix: $SHARD_TARGET_PREFIX"