#!/usr/bin/env bash
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; UPY=$CONDA/envs/ur5dp/bin/python
TASK=${TASK:-battery}; NAME=${TASK}_v1
log() { echo "== $(date '+%F %T') [$TASK] $*"; }
FINAL=$RUN/data/${NAME}_1028

source $CONDA/etc/profile.d/conda.sh
kept=$($UPY -c "import json; d=json.load(open('$FINAL/collect_meta.json')); print(d['n_kept'])" 2>/dev/null || echo 0)
log "Training DP ($kept demos) with AMP + batch_size=256 + num_workers=16"

cd $V5CODE/ur5_lab_code && export PYTHONPATH=src:.
rm -f $V5CODE/train_runs/${NAME}_1028_aug_s0/logs.json.txt  # clear old logs to allow fresh start

$UPY train.py --config-name=train_baseline_diffusion_unet task=sim_baseline \
    task_name=${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_aug_s0" \
    2>&1 | tee $RUN/logs/${NAME}_1028_train.log
rc=$?; log "Train exit $rc"
[ $rc -eq 0 ] && log "DONE: $V5CODE/train_runs/${NAME}_1028_aug_s0/checkpoints/" || { log "TRAIN FAILED"; exit 1; }