flashvtg-experiment-backup / FlashVTG /scripts /run_lab_preprocess.sh
zhaoshiwen's picture
Add files using upload-large-folder tool
57259ac verified
Raw
History Blame Contribute Delete
1.04 kB
#!/bin/bash
# 一键完成 Lab 数据集预处理:生成 QV 风格 jsonl + 提取 CLIP 视频特征 + 提取 LLaMA 文本特征
# 用法: bash scripts/run_lab_preprocess.sh [gpu_id]
# 默认 GPU 0;无 GPU 时自动用 CPU(CLIP/LLaMA 会较慢)
set -e
cd "$(dirname "$0")/.."
LAB_DIR=${LAB_DIR:-features/lab}
OUT_JSONL=${OUT_JSONL:-data/lab_val.jsonl}
GPU_ID=${1:-0}
echo "Lab dir: $LAB_DIR"
echo "Output jsonl: $OUT_JSONL"
echo "GPU: $GPU_ID"
# 1) 生成 QV 风格标注(不依赖 GPU)
python scripts/prepare_lab_dataset.py \
--lab_dir "$LAB_DIR" \
--out_jsonl "$OUT_JSONL"
# 2) 提取 CLIP + LLaMA(需要 GPU 和 open_clip / transformers)
export CUDA_VISIBLE_DEVICES=$GPU_ID
python scripts/prepare_lab_dataset.py \
--lab_dir "$LAB_DIR" \
--out_jsonl "$OUT_JSONL" \
--extract_clip \
--extract_llama \
--device "${DEVICE:-cuda}"
echo "Preprocess done. Next: run inference with --eval_path $OUT_JSONL and v_feat_dirs/t_feat_dir pointing to features/lab/clip_features and features/lab/llama_text_feature"