#!/usr/bin/env bash
# ringpeg.sh (1028_run) — 4x parallel collection + AMP training
set -uo pipefail
ROOT=/lustre/fsw/portfolios/nvr/users/jtremblay/yu
RUN=$ROOT/1028_run; TASKS=$ROOT/1027_sim_tasks
V5CODE=$TASKS/share_v5_sim2real_20260929/code
CONDA=$ROOT/miniconda3; SPY=$CONDA/envs/simfoundry/bin/python; UPY=$CONDA/envs/ur5dp/bin/python
HOME_FAKE=$RUN/home_fake; NAME=ringpeg_dart_v1; RDIR=$TASKS/ringpeg
YUASSET=$TASKS/assets/yuasset_v1_docs_only
log() { echo "== $(date '+%F %T') $*"; }

mkdir -p $HOME_FAKE/ur5_sim2real_v5 $RUN/logs $V5CODE/sim_data $V5CODE/train_runs
ln -sfn $CONDA $HOME_FAKE/miniconda3
ln -sfn $TASKS/share_v5_sim2real_20260929 $HOME_FAKE/ur5_sim2real_v5/share_v5_sim2real_20260929
ln -sfn $RDIR $HOME_FAKE/ur5_sim2real_v5/ringpeg
ln -sfn $YUASSET $HOME_FAKE/ur5_sim2real_v5/yuasset
export HOME=$HOME_FAKE
source $CONDA/etc/profile.d/conda.sh
export TORCHINDUCTOR_COMPILE_THREADS=1 OMNIGIBSON_HEADLESS=1 OMNI_KIT_ACCEPT_EULA=YES PYTHONUNBUFFERED=1
export OMNIGIBSON_RENDER_VIEWER_CAMERA=0 TORCH_COMPILE_DISABLE=1 TORCHDYNAMO_DISABLE=1
export OMNIGIBSON_DATA_PATH=/lustre/fsw/portfolios/nvr/users/jtremblay/yu/SimFoundry/deps/BEHAVIOR-1K/datasets

DATA_A=$RUN/data/${NAME}_1028a; DATA_B=$RUN/data/${NAME}_1028b
DATA_C=$RUN/data/${NAME}_1028c; DATA_D=$RUN/data/${NAME}_1028d
mkdir -p $DATA_A $DATA_B $DATA_C $DATA_D

log "Collecting ringpeg: 4x parallel (GPU 0/1/2/3)"

OMNIGIBSON_GPU_ID=0 \
$SPY -u $RDIR/ringpeg_collect.py \
    --out $DATA_A --seeds 300-399 --max-kept 15 \
    --layouts 0 --jitter 0.03 --ring-yaw-deg 30 --peg-yaw-deg 15 --dart --start-offset \
    2>&1 | tee $RUN/logs/${NAME}_1028a.log &
PID_A=$!
sleep 55

OMNIGIBSON_GPU_ID=1 \
$SPY -u $RDIR/ringpeg_collect.py \
    --out $DATA_B --seeds 400-499 --max-kept 15 \
    --layouts 0 --jitter 0.03 --ring-yaw-deg 30 --peg-yaw-deg 15 --dart --start-offset \
    2>&1 | tee $RUN/logs/${NAME}_1028b.log &
PID_B=$!
sleep 55

OMNIGIBSON_GPU_ID=2 \
$SPY -u $RDIR/ringpeg_collect.py \
    --out $DATA_C --seeds 500-599 --max-kept 15 \
    --layouts 0 --jitter 0.03 --ring-yaw-deg 30 --peg-yaw-deg 15 --dart --start-offset \
    2>&1 | tee $RUN/logs/${NAME}_1028c.log &
PID_C=$!
sleep 55

OMNIGIBSON_GPU_ID=3 \
$SPY -u $RDIR/ringpeg_collect.py \
    --out $DATA_D --seeds 600-699 --max-kept 20 \
    --layouts 0 --jitter 0.03 --ring-yaw-deg 30 --peg-yaw-deg 15 --dart --start-offset \
    2>&1 | tee $RUN/logs/${NAME}_1028d.log &
PID_D=$!

wait $PID_A; log "GPU0 done"
wait $PID_B; log "GPU1 done"
wait $PID_C; log "GPU2 done"
wait $PID_D; log "GPU3 done"

FINAL=$RUN/data/${NAME}_1028
python3 $RUN/merge_shards.py $DATA_A $DATA_B $DATA_C $DATA_D $FINAL 50
kept=$($UPY -c "import json; d=json.load(open('$FINAL/collect_meta.json')); print(d['n_kept'])" 2>/dev/null || echo 0)
log "Kept: $kept"; [ "$kept" -lt 40 ] && { log "ERROR: only $kept"; exit 1; }

ln -sfn $FINAL $V5CODE/sim_data/${NAME}_1028
log "Packing zarr"
cd $V5CODE && $UPY ur5_import/scripts/pack_zarr.py --dir sim_data/${NAME}_1028 2>&1 | tee $RUN/logs/${NAME}_1028_train.log
log "Training DP ($kept demos)"
cd $V5CODE/ur5_lab_code && export PYTHONPATH=src:.
$UPY train.py --config-name=train_baseline_diffusion_unet task=sim_baseline \
    task_name=${NAME}_1028 dataset_path=$FINAL n_demo=$kept \
    training.device=cuda:0 training.seed=0 training.num_epochs=130 training.checkpoint_every=129 \
    checkpoint.topk.k=0 training.resume=True checkpoint.save_last_ckpt=True \
    +training.use_amp=True dataloader.batch_size=256 dataloader.num_workers=16 logging.mode=offline \
    "policy._target_=algorithms.policy.diffusion_unet_baseline_aug_policy.DiffusionUnetBaselineAugPolicy" \
    +policy.cam1_dropout=0.1 +policy.cam1_aug_p=0.5 +policy.cam1_erase_p=0.3 +policy.cam0_aug_p=0.3 \
    "hydra.run.dir=$V5CODE/train_runs/${NAME}_1028_s0" \
    2>&1 | tee -a $RUN/logs/${NAME}_1028_train.log
rc=$?; log "Train exit $rc"; [ $rc -eq 0 ] || { log "TRAIN FAILED"; exit 1; }
log "DONE: $V5CODE/train_runs/${NAME}_1028_s0/checkpoints/"