code/scripts/preprocess_forensics.sh

#!/bin/bash
# Pre-encode AF videos with the Qwen2.5-VL processor.
# Output cache layout: <OUTPUT_DIR>/<split>/<generator>/<sample_id>/{video_inputs.pt,video_kwargs.json}

FPS="${FPS:-2.0}"
OUTPUT_DIR="${OUTPUT_DIR:-/ces/zt/forensics_grpo_cache_uniform3584_fps${FPS}}"

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

python preprocess_forensics.py \
    --annot_dir /ces/zt/activityforensics/annot \
    --video_root /ces/zt \
    --output_dir "$OUTPUT_DIR" \
    --max_pix_size 3584 \
    --min_pix_size 16 \
    --fps "$FPS" \
    --num_workers 8 \
    --splits train test \
    --skip_existing