code/scripts/run_grpo_forensics_ab_noAug.sh

#!/bin/bash
# Ablation AB2: v10_r2 w/o Spatio-Temporal consistency augmentation block.
# Disables both SPI (splice-position invariance) and FBR (forward-backward
# reversal). Everything else inherits v10_r2.

export WANDB_PROJECT=Forensics-GRPO
export WANDB_NAME="${WANDB_NAME:-AF_ab_noAug}"

export PYTHONPATH=".:$PYTHONPATH"
export DEBUG_MODE="true"
export LOG_PATH="./qwen2.5_7b_vl_forensics_ab_noAug.txt"

export PATH="/mnt/local-fast/zhangt/torch_env/bin:$PATH"
export LD_LIBRARY_PATH="/opt/conda/lib:${LD_LIBRARY_PATH}"

# === v10_r2 baseline (inherited) ============================================
export FORENSICS_COT=false
export FORENSICS_FORGERY_AWARE=false

export FORENSICS_HUNGARIAN_DECOMP=true
export FORENSICS_HUNGARIAN_MAXK=8
export FORENSICS_BOUNDARY_WEIGHT=0.3
export FORENSICS_BOUNDARY_TAU=2.0

# --- AB2 change: Spatio-Temporal augmentation block OFF --------------------
export FORENSICS_SPI_AUG=false
export FORENSICS_FBR_AUG=false
# ----------------------------------------------------------------------------

export FORENSICS_STRICT_TAUS="0.3,0.5,0.7"
export FORENSICS_STRICT_EDGE_SIGMA=0.5

unset FORENSICS_COT_CONSIS_WEIGHT
unset FORENSICS_COT_CF_WEIGHT
unset FORENSICS_COT_ALIGN_WEIGHT
unset FORENSICS_K_REQUIRE_TIMESTEP
unset FORENSICS_IOU_REQUIRE_THINK
# ============================================================================

OUTDIR="${OUTDIR:-outputs_forensics/ab_noAug}"
MODEL_PATH="${MODEL_PATH:-/mnt/local-fast/zhangt/Qwen2.5-VL-7B-Instruct}"
PREPROCESSED="${PREPROCESSED:-/mnt/local-fast/zhangt/forensics_grpo_cache_uniform3584_fps2.0}"

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

torchrun --nproc_per_node="8" \
    --nnodes="1" \
    --node_rank="0" \
    --master_addr="127.0.0.1" \
    --master_port="12372" \
    src/open_r1/grpo_forensics.py \
    --deepspeed scripts/zero3_offload.json \
    --output_dir "$OUTDIR" \
    --model_name_or_path "$MODEL_PATH" \
    --annot_dir /mnt/local-fast/zhangt/annot/annot \
    --video_root /mnt/local-fast/zhangt/video \
    --preprocessed_data_path "$PREPROCESSED" \
    --reward_funcs iou strict_boundary strict_edge \
    --dataset_name activityforensics \
    --max_prompt_length 8192 \
    --max_completion_length 512 \
    --num_generations 4 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 2 \
    --logging_steps 1 \
    --bf16 \
    --torch_dtype bfloat16 \
    --data_seed 42 \
    --gradient_checkpointing true \
    --attn_implementation flash_attention_2 \
    --num_train_epochs 4 \
    --run_name "$WANDB_NAME" \
    --report_to none \
    --save_steps 240 \
    --learning_rate 1e-6 \
    --save_only_model true