eval/run_eval_rl.sh

#!/bin/bash
# RL v5e evaluation: rep=1.0 then rep=1.1, GPU 4-7
set -e
source /root/miniconda3/etc/profile.d/conda.sh
conda activate Video-R1
export DECORD_EOF_RETRY_MAX=20480
export CUDA_VISIBLE_DEVICES=4,5,6,7

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
MODEL="/root/zwj/DM-GRPO/Model/DeltaMask-GRPO-v5e_SFT-v5e-32f_ts-reward"

# rep=1.0
python "$SCRIPT_DIR/eval.py" \
    --model_path "$MODEL" \
    --file_name "DM-GRPO-v5e_rep1.0" \
    --benchmarks tempcompass \
    --bsz 64 --max_tokens 4096 --max_model_len 16384 \
    --gpu_memory_utilization 0.8 --repetition_penalty 1.0 --nframes 32

# rep=1.1
python "$SCRIPT_DIR/eval.py" \
    --model_path "$MODEL" \
    --file_name "DM-GRPO-v5e_rep1.1" \
    --benchmarks tempcompass \
    --bsz 64 --max_tokens 4096 --max_model_len 16384 \
    --gpu_memory_utilization 0.8 --repetition_penalty 1.1 --nframes 32

echo "RL v5e evaluation complete."