eval_only.py

from vlmeval.dataset import build_dataset
from vlmeval.smp import *


load_env()
# dataset_name = "MMMU_DEV_VAL"
# dataset_name = "MathVista_MINI"
dataset_name = "DynaMath"
dataset = build_dataset(dataset_name)
judge_kwargs = {
    'nproc': 16,
    'verbose': True,
    'retry': 10,
}
if dataset.TYPE in ['MCQ', 'Y/N', 'MCQ_MMMU_Pro'] or listinstr(['moviechat1k'], dataset_name.lower()):
    if listinstr(['WeMath'], dataset_name):
        judge_kwargs['model'] = 'gpt-4o-mini'
    else:
        judge_kwargs['model'] = 'chatgpt-0125'
elif listinstr(['MMVet', 'LLaVABench', 'MMBench_Video'], dataset_name):
    judge_kwargs['model'] = 'gpt-4-turbo'
elif listinstr(['MathVista', 'MathVerse', 'MathVision', 'DynaMath', 'VL-RewardBench', 'LogicVista', 'MOAT'], dataset_name):  # noqa: E501
    judge_kwargs['model'] = 'gpt-4o-mini'
elif listinstr(['MMLongBench', 'MMDU', 'DUDE', 'SLIDEVQA', 'MIA-Bench', 'WildVision', 'MMAlignBench'], dataset_name):  # noqa: E501
    judge_kwargs['model'] = 'gpt-4o'

fs = [
    "/user/konglingyu/VLMEvalKit/public_eval/grpo_v7_exp0_qwen25vl_scalable_rl_opensource_math_grpo_bs96_wofilter_scoreB_std_filter_0523_global_step_200/DynaMath_train_prompt_greedy/20250524/grpo_v7_exp0_qwen25vl_scalable_rl_opensource_math_grpo_bs96_wofilter_scoreB_std_filter_0523_global_step_200/T20250524_G/grpo_v7_exp0_qwen25vl_scalable_rl_opensource_math_grpo_bs96_wofilter_scoreB_std_filter_0523_global_step_200_DynaMath.xlsx"
    # "/user/konglingyu/VLMEvalKit/outputs/Qwen2.5-VL-7B-Instruct-original/T20250412_G/Qwen2.5-VL-7B-Instruct-original_DynaMath.xlsx",
    # "/user/konglingyu/VLMEvalKit/outputs/Qwen2.5-VL-7B-RL-greedy/T20250414_G/Qwen2.5-VL-7B-RL-greedy_DynaMath.xlsx",
    # "/user/konglingyu/VLMEvalKit/public_eval/bbox_step_300/DynaMath/20250418/bbox_step_300/T20250418_G/bbox_step_300_DynaMath.xlsx",
    # "/user/konglingyu/VLMEvalKit/public_eval/clip_high_step_600/DynaMath/20250419/clip_high_step_600/T20250419_G/clip_high_step_600_DynaMath.xlsx",
    # "/user/konglingyu/VLMEvalKit/public_eval/dr_grpo_step_600/DynaMath/20250419/dr_grpo_step_600/T20250419_G/dr_grpo_step_600_DynaMath.xlsx",
    # "/user/konglingyu/VLMEvalKit/public_eval/dr_grpo_step_800/DynaMath/20250418/dr_grpo_step_800/T20250418_G/dr_grpo_step_800_DynaMath.xlsx",
    # "/user/konglingyu/VLMEvalKit/public_eval/grpo_v7_exp9_qwen25vl_grpo_opensource_math_doc_dr_grpo_500/DynaMath/20250417/grpo_v7_exp9_qwen25vl_grpo_opensource_math_doc_dr_grpo/T20250417_G/grpo_v7_exp9_qwen25vl_grpo_opensource_math_doc_dr_grpo_DynaMath.xlsx",
    # "/user/konglingyu/VLMEvalKit/public_eval/naive_grpo_step_400/DynaMath/20250418/naive_grpo_step_400/T20250418_G/naive_grpo_step_400_DynaMath.xlsx",
]
# file = "/user/konglingyu/VLMEvalKit/public_eval/bbox_step_300/DynaMath/20250418/bbox_step_300/T20250418_G/bbox_step_300_DynaMath.xlsx"
for file in fs:
    try:
        os.remove(file.replace(".xlsx", "_gpt-4o-mini_score.csv"))
        os.remove(file.replace(".xlsx", "_gpt-4o-mini.pkl"))
        os.remove(file.replace(".xlsx", "_gpt-4o-mini.xlsx"))
        print("Removed old files")
    except:
        pass
    dataset.evaluate(file, **judge_kwargs)
    with open(file.replace(".xlsx", "_gpt-4o-mini_score.csv")) as f:
        lines = f.readlines()
        print(f"File: {file.split('/')[-1]}")
        for line in lines:
            print(line.strip())