training/Policy/run_v5_ablations.sh

#!/bin/bash
# ──────────────────────────────────────────────────────────────────────────────
# V5 Hierarchical PolicyHead — Ablation Suite
#
# Runs 4 experiments to validate the hierarchical decomposition:
#   1) v5_base:     Hierarchical head, default hyperparams
#   2) v5_focal:    Higher focal emphasis (α=0.85, γ=3.0)
#   3) v5_mono:     + VideoGroupedSampler + monotonic constraint (λ=0.1)
#   4) v5_smooth:   + label smoothing (ε=0.05)
#
# Each experiment: ~12-15min  (early stop ~epoch 3, belief cache mode)
# Total:           ~50-60min
#
# Results go to checkpoints/Policy/policy_warmstart_v5_*/best
#
# Usage:
#   bash training/Policy/run_v5_ablations.sh
# ──────────────────────────────────────────────────────────────────────────────

set -euo pipefail

cd "$(dirname "$0")/../.."

SFT_CKPT="checkpoints/SFT/sft_v2/best"
LABEL_DIR="data/policy_labels"
CACHE_DIR="data/belief_cache"
OUTPUT_DIR="checkpoints/Policy"

echo "╔══════════════════════════════════════════════════════════╗"
echo "║  V5 Hierarchical PolicyHead — Ablation Suite            ║"
echo "║  4 experiments × ~15 min ≈ 50-60 min total              ║"
echo "╚══════════════════════════════════════════════════════════╝"

# ── 1) v5_base: Default hierarchical head ────────────────────────────────────
echo ""
echo "━━━ [1/4] v5_base: default hierarchical head ━━━"
TAG=base \
FOCAL_ALPHA=0.75 FOCAL_GAMMA=2.0 \
ALERT_LOSS_W=1.0 DANGER_LOSS_W=0.5 \
bash training/Policy/train_policy_v5.sh --balanced

# ── 2) v5_focal: Stronger focal emphasis ─────────────────────────────────────
echo ""
echo "━━━ [2/4] v5_focal: α=0.85 γ=3.0 ━━━"
TAG=focal \
FOCAL_ALPHA=0.85 FOCAL_GAMMA=3.0 \
ALERT_LOSS_W=1.5 DANGER_LOSS_W=0.5 \
bash training/Policy/train_policy_v5.sh --balanced

# ── 3) v5_mono: + video-grouped sampler + monotonic ─────────────────────────
echo ""
echo "━━━ [3/4] v5_mono: hierarchical + VideoGroupedSampler + mono_λ=0.1 ━━━"
TAG=mono \
FOCAL_ALPHA=0.75 FOCAL_GAMMA=2.0 \
ALERT_LOSS_W=1.0 DANGER_LOSS_W=0.5 \
MONO_LAMBDA=0.1 \
bash training/Policy/train_policy_v5.sh --video-sampler

# ── 4) v5_smooth: + label smoothing ─────────────────────────────────────────
echo ""
echo "━━━ [4/4] v5_smooth: hierarchical + label_smoothing=0.05 ━━━"
TAG=smooth \
FOCAL_ALPHA=0.75 FOCAL_GAMMA=2.0 \
ALERT_LOSS_W=1.0 DANGER_LOSS_W=0.5 \
LABEL_SMOOTHING=0.05 \
bash training/Policy/train_policy_v5.sh --balanced

echo ""
echo "╔══════════════════════════════════════════════════════════╗"
echo "║  All 4 ablations complete!                              ║"
echo "║                                                         ║"
echo "║  Checkpoints:                                           ║"
echo "║    ${OUTPUT_DIR}/policy_warmstart_v5_base/best"
echo "║    ${OUTPUT_DIR}/policy_warmstart_v5_focal/best"
echo "║    ${OUTPUT_DIR}/policy_warmstart_v5_mono/best"
echo "║    ${OUTPUT_DIR}/policy_warmstart_v5_smooth/best"
echo "║                                                         ║"
echo "║  Compare results in each best/policy_meta.json          ║"
echo "╚══════════════════════════════════════════════════════════╝"