#!/bin/bash
# Fallback if SmolVLA does not learn the baseline: train ACT from scratch (ResNet-18
# image backbone, pretrained on ImageNet) on the same data, overhead and wrist cameras,
# and push it to the Hugging Face Hub. ACT reads the camera keys as they are, so no
# rename map. Run from the project root on the GPU pod:
#   ROOT=data/so101_chess_baseline REPO=Machanize/chess_phase_act_baseline bash sim/train_act.sh
set -euo pipefail
cd "$(dirname "$0")/.."
ROOT=${ROOT:-data/so101_chess_baseline}
REPO=${REPO:-Machanize/chess_phase_act_baseline}
JOB=${JOB:-chess_phase_act_baseline}
.venv/bin/lerobot-train \
  --policy.type=act \
  --dataset.repo_id="${DATASET_ID:-local/so101_chess_baseline}" \
  --dataset.root="$ROOT" \
  --batch_size="${BATCH:-32}" \
  --steps="${STEPS:-40000}" \
  --save_freq="${SAVE_FREQ:-10000}" \
  --log_freq=100 \
  --num_workers="${WORKERS:-16}" \
  --output_dir="${OUT:-outputs/$JOB}" \
  --job_name="$JOB" \
  --policy.device=cuda \
  --wandb.enable=false \
  --policy.push_to_hub="${PUSH:-true}" \
  --policy.repo_id="$REPO" \
  --policy.private=false