#!/usr/bin/env bash # Train an ACT base policy (CVAE, chunk25). Pick the GPU via CUDA_VISIBLE_DEVICES. # Usage: # CUDA_VISIBLE_DEVICES=0 bash launch_act_train.sh sim_open_laptop_100demos open_laptop_100demos_chunk25 # Overrides: EPOCHS (10000), CHUNK (25), BS (8), PY. set -e TASK="${1:?task_name (a key in act/constants.py SIM_TASK_CONFIGS)}" RUN="${2:?ckpt subdir name under /data0/hongyi/act_runs}" EPOCHS="${EPOCHS:-10000}"; CHUNK="${CHUNK:-25}"; BS="${BS:-8}" PY="${PY:-/home/hongyi/miniconda3/envs/pwm/bin/python}" CKPT="/data0/hongyi/act_runs/${RUN}" LOG="/home/hongyi/PWM/policy_rl/training_logs/act_${RUN}.log" mkdir -p "$CKPT" "$(dirname "$LOG")" cd /home/hongyi/PWM/policy_rl/act # act/ for constants/utils/policy/detr; act/detr for DETR's internal `util`/`models` imports export PYTHONPATH="/home/hongyi/PWM/policy_rl/act:/home/hongyi/PWM/policy_rl/act/detr:${PYTHONPATH}" echo "=== ACT train: task=$TASK ckpt=$CKPT GPU=$CUDA_VISIBLE_DEVICES epochs=$EPOCHS chunk=$CHUNK bs=$BS" nohup $PY imitate_episodes.py \ --task_name "$TASK" --ckpt_dir "$CKPT" --policy_class ACT \ --batch_size $BS --seed 0 --num_epochs $EPOCHS --lr 1e-5 \ --kl_weight 10 --chunk_size $CHUNK --hidden_dim 512 --dim_feedforward 3200 \ > "$LOG" 2>&1 & echo "trainer pid=$! log=$LOG"