utils/count.py

import jsonlines
import re
import os
from matplotlib import pyplot as plt

def count_bbox(answer: str):
    PATTERN = re.compile(r'\((.*?)\),\((.*?)\)')
    bbox_num = len(re.findall(PATTERN, answer))
    return bbox_num

save_ckpt_lst = [350]

mmmu_box_ratio, mmmu_avg_box_num = [], []
mathvista_box_ratio, mathvista_avg_box_num = [], []
wtq_box_ratio, wtq_avg_box_num = [], []
mmdocbench_box_ratio, mmdocbench_avg_box_num = [], []
tablefact_box_ratio, tablefact_avg_box_num = [], []

for results in os.listdir("/user/xuqixin/checkpoints/qwen25_vl-7b/grpo_v7_exp10_qwen25_vl_sft_bbox_grpo_opensource_doc/results"):
    for ckpt in save_ckpt_lst:
        if not results.endswith(f"_{ckpt}.jsonl"):
            continue
        with jsonlines.open(f"/user/xuqixin/checkpoints/qwen25_vl-7b/grpo_v7_exp10_qwen25_vl_sft_bbox_grpo_opensource_doc/results/{results}") as reader:
            total_num = 0
            total_bbox_num = 0
            bbox_sample_num = 0
            already_print = 0
            print("##############################################################################")
            print(f"results: {results}")
            for obj in reader:
                total_num += 1
                answer = obj["output"]
                bbox_num = count_bbox(answer)
                total_bbox_num += bbox_num
                if bbox_num > 0:
                    bbox_sample_num += 1
                    if already_print < 5:
                        print(obj)
                        print("\n")
                        already_print += 1
            # print(f"results: {results}")
            # print(f"total_num: {total_num}, total_bbox_num: {total_bbox_num}, bbox_sample_num: {bbox_sample_num}")
            # print(f"average box num: {total_bbox_num / bbox_sample_num}, bbox num ratio: {bbox_sample_num / total_num}")

            if "mmmu" in results:
                # maintain only 2 decimal places
                mmmu_box_ratio.append(round(bbox_sample_num / total_num, 3))
                mmmu_avg_box_num.append(round(total_bbox_num / bbox_sample_num, 3))
            elif "mathvista" in results:
                mathvista_box_ratio.append(round(bbox_sample_num / total_num, 3))
                mathvista_avg_box_num.append(round(total_bbox_num / bbox_sample_num, 3))
            elif "wtq" in results:
                wtq_box_ratio.append(round(bbox_sample_num / total_num, 3))
                wtq_avg_box_num.append(round(total_bbox_num / bbox_sample_num, 3))
            elif "mmdocbench" in results:
                mmdocbench_box_ratio.append(round(bbox_sample_num / total_num, 3))
                mmdocbench_avg_box_num.append(round(total_bbox_num / bbox_sample_num, 3))
            elif "tablefact" in results:
                tablefact_box_ratio.append(round(bbox_sample_num / total_num, 3))
                tablefact_avg_box_num.append(round(total_bbox_num / bbox_sample_num, 3))


# plot the figure
plt.figure(figsize=(10, 5))
plt.subplot(1, 2, 1)
plt.plot(save_ckpt_lst, mmmu_box_ratio, label='mmmu')
plt.plot(save_ckpt_lst, mathvista_box_ratio, label='mathvista')
plt.plot(save_ckpt_lst, wtq_box_ratio, label='wtq')
plt.plot(save_ckpt_lst, mmdocbench_box_ratio, label='mmdocbench')
plt.plot(save_ckpt_lst, tablefact_box_ratio, label='tablefact')
plt.xlabel('Checkpoint')
plt.ylabel('Box Ratio')
plt.title('Box Ratio vs Checkpoint')
plt.legend()
plt.subplot(1, 2, 2)
plt.plot(save_ckpt_lst, mmmu_avg_box_num, label='mmmu')
plt.plot(save_ckpt_lst, mathvista_avg_box_num, label='mathvista')
plt.plot(save_ckpt_lst, wtq_avg_box_num, label='wtq')
plt.plot(save_ckpt_lst, mmdocbench_avg_box_num, label='mmdocbench')
plt.plot(save_ckpt_lst, tablefact_avg_box_num, label='tablefact')
plt.xlabel('Checkpoint')
plt.ylabel('Average Box Number')
plt.title('Average Box Number vs Checkpoint')
plt.legend()
plt.tight_layout()
plt.savefig('/user/xuqixin/checkpoints/qwen25_vl-7b/grpo_v7_exp10_qwen25_vl_sft_bbox_grpo_opensource_doc/box_ratio.png')
plt.show()

print("##############################################################################")
print("mmmu box ratio:", mmmu_box_ratio)
print("mathvista box ratio:", mathvista_box_ratio)
print("wtq box ratio:", wtq_box_ratio)
print("mmdocbench box ratio:", mmdocbench_box_ratio)
print("tablefact box ratio:", tablefact_box_ratio)

print("mmmu avg box num:", mmmu_avg_box_num)
print("mathvista avg box num:", mathvista_avg_box_num)
print("wtq avg box num:", wtq_avg_box_num)
print("mmdocbench avg box num:", mmdocbench_avg_box_num)
print("tablefact avg box num:", tablefact_avg_box_num)
            
# results: mmmu_350.jsonl
# total_num: 894, total_bbox_num: 52, bbox_sample_num: 31
# average box num: 1.6774193548387097, bbox num ratio: 0.03467561521252797
# results: mathvista_350.jsonl
# total_num: 998, total_bbox_num: 157, bbox_sample_num: 134
# average box num: 1.171641791044776, bbox num ratio: 0.1342685370741483
# results: wtq_350.jsonl
# total_num: 868, total_bbox_num: 1015, bbox_sample_num: 738
# average box num: 1.3753387533875339, bbox num ratio: 0.8502304147465438
# results: mmdocbench_350.jsonl
# total_num: 1000, total_bbox_num: 517, bbox_sample_num: 409
# average box num: 1.2640586797066016, bbox num ratio: 0.409
# results: tablefact_350.jsonl
# total_num: 868, total_bbox_num: 816, bbox_sample_num: 597
# average box num: 1.3668341708542713, bbox num ratio: 0.6877880184331797