eval_scripts/video-SALMONN2-testset/evaluation.py

import time
import copy
import os
import json
import concurrent.futures
import threading
from tqdm import tqdm
import ast
import random
import re
import sys
from pathlib import Path

try:
    with open("apikey.txt", "r") as f:
        api_key = f.read()
except:
    api_key = ''

def call_gpt41(msg):
    while True:
        try:
            response = openai.ChatCompletion.create(
                model="gpt-4.1", 
                messages=msg, 
                api_key=api_key, 
                request_timeout=5)
            break
        except:
            print("Timeout, retrying...")
            time.sleep(5)

    output_text = response['choices'][0]['message']['content']
    return output_text


script_dir = Path(__file__).resolve().parent
events_file = script_dir / "video_salmonn2_test.json"
LOCK = threading.Lock()

summarize_prompt_system = '''
### Task:
A good video description is one that describes the various details in the video. You task is to judge whether a video description is good or not. You will be provided all the base events in the video, and also a video description to be evaluated. You need to determine which video events are described in the given video description.

### Instruction:
Given the video desciption, besides the events described correctly, there might be events that are missed, described incorrectly and hallucination. You need to determine the number of missed events, incorrect events and hallucination events.
- Missed Event: An event from the given base events set whose central action, participants, and context are entirely absent in the caption to be evaluated, meaning the main occurrence is not conveyed in any form, even if paraphrased or summarized.
- Incorrect Event: An event that is in the base events set and is mentioned in the caption, but described with significant factual inaccuracies.
- Hallucination Event: An event mentioned in the caption that is not in the base events set and is not a plausible variation or inference from them; it is entirely fabricated or clearly inappropriate given the context.
You are also required to list these events out. Note that there should be no overlap between Incorrect Events and Hallucination Events.

### Input Format:
- There are totally {event_num} base events in the video. All the base events in the video will be provided in List format, i.e. ["xxx", "xxx", ...]
- The video description to be evaluated will be provided as well.

### Output Format:
Your output should be strict in the following Python dictionary format without anything else:
{{"Missed": x, "Incorrect": x, "Hallucination": x, "Missed Event": [...], "Incorrect Event": [...], "Hallucination Event": [...] }}
'''

summarize_prompt_system_2 = '''
### Task:
A good video description is one that describes the various details in the video. You task is to judge whether a video description is good or not. You will be provided all the base events in the video, and also a video description to be evaluated. You need to determine which video events are described in the given video description.

### Instruction:
Given the video desciption, you need to determine the number of missed events, correct events, incorrect events and hallucination events. Make sure that: "missed" + "correct" + "incorrect" + "hallucination" = {event_num}
- Missed Event: An event from the given base events set whose central action, participants, and context are entirely absent in the caption to be evaluated, meaning the main occurrence is not conveyed in any form, even if paraphrased or summarized.
- Incorrect Event: An event that is in the base events set and is mentioned in the caption, but described with significant factual inaccuracies.
- Hallucination Event: An event mentioned in the caption that is not in the base events set and is not a plausible variation or inference from them; it is entirely fabricated or clearly inappropriate given the context.
Note that there should be no overlap between Incorrect Events and Hallucination Events.

### Input Format:
- There are totally {event_num} base events in the video. All the base events in the video will be provided in List format, i.e. ["xxx", "xxx", ...]
- The video description to be evaluated will be provided as well.

### Output Format:
Your output should be strict in the following Python dictionary format without anything else:
{{"Missed": x, "Correct": x, "Incorrect": x, "Hallucination": x}}
''' 

summarize_prompt = '''
#### Events In The Video
{events_in_video}

#### Video Description To Be Rated
{cap_to_be_rated}

Given base events in the video and the video description, please count the missed, incorrect and hallucination events and list them out. 
'''

summarize_prompt_2 = '''
#### Events In The Video
{events_in_video}

#### Video Description To Be Rated
{cap_to_be_rated}

Given base events in the video and the video description, please count the missed, correct, incorrect and hallucination events. 
'''

seed = 2024

def gpt_caption(ref, pred, summarize_prompt_system=None, summarize_prompt=None):
    temp_query_system = summarize_prompt_system.format(event_num=len(ref))
    temp_summarize_prompt = summarize_prompt.format(events_in_video=ref, cap_to_be_rated=pred)

    msg = [
        {"role": "system", "content": temp_query_system},
        {"role": "user", "content": temp_summarize_prompt}
    ]
    completion = call_gpt41(msg)

    return completion


def reduce_repeated_words(text):
    pattern = "."
    for i in range(1, 50):
        p = pattern * i
        text = re.sub(f'({p})' + r'\1{4,200}', r'\1', text)
    for i in range(50, 100):
        p = pattern * i
        text = re.sub(f'({p})' + r'\1{3,200}', r'\1', text)
    return text

def gpt_extract(item):
    try:
        if isinstance(item['pred'], list):
            item['pred'] = item['pred'][0]
        if "<|im_end|>" not in item['pred']:
            text = reduce_repeated_words(item["pred"])
        else:
            text = item['pred'].replace("<|im_end|>", "")
        res = gpt_caption(item['events'], text, summarize_prompt_system=summarize_prompt_system, summarize_prompt=summarize_prompt)

        miss = int(res.split('"Missed":')[1].split('"Incorrect"')[0].strip().replace(",", ""))
        incor = int(res.split('"Incorrect":')[1].split('"Hallucination"')[0].strip().replace(",", ""))
        hall = int(res.split('"Hallucination":')[1].split('"Missed Event"')[0].strip().replace(",", ""))

        try:
            miss_event = json.loads(res.split('"Missed Event":')[1].split('"Incorrect Event"')[0].strip()[:-1])
        except Exception as e:
            miss_event = eval(res.split('"Missed Event":')[1].split('"Incorrect Event"')[0].strip()[:-1])
        
        try:
            incor_event = json.loads(res.split('"Incorrect Event":')[1].split('"Hallucination Event"')[0].strip()[:-1])
        except Exception as e:
            incor_event = eval(res.split('"Incorrect Event":')[1].split('"Hallucination Event"')[0].strip()[:-1])
        
        try:
            hall_event = json.loads(res.split('"Hallucination Event":')[1].split('}')[0].strip())
        except Exception as e:
            hall_event = eval(res.split('"Hallucination Event":')[1].split('}')[0].strip())

        item["cutted_pred"] = text
        item["Missed"] = miss
        item["Incorrect"] = incor
        item["Hallucination"] = hall
        item["Missed Event"] = miss_event
        item["Incorrect Event"] = incor_event
        item["Hallucination Event"] = hall_event

        return item

    except Exception as e:
        return item

def gpt_extract_2(item):
    try:
        if "<|im_end|>" not in item['pred']:
            text = reduce_repeated_words(item["pred"])
        else:
            text = item['pred'].replace("<|im_end|>", "")
        res = gpt_caption(item['events'], text, summarize_prompt_system=summarize_prompt_system_2, summarize_prompt=summarize_prompt_2)

        miss = int(res.split('"Missed":')[1].split('"Correct"')[0].strip().replace(",", ""))
        cor = int(res.split('"Correct":')[1].split('"Incorrect"')[0].strip().replace(",", ""))
        incor = int(res.split('"Incorrect":')[1].split('"Hallucination"')[0].strip().replace(",", ""))
        hall = int(res.split('"Hallucination":')[1].split('}')[0].strip().replace(",", ""))

        assert miss + cor + incor == len(item["events"])

        item["cutted_pred"] = text
        item["Missed"] = miss
        item["Incorrect"] = incor
        item["Hallucination"] = hall
        item["Correct"] = cor

        return item

    except Exception as e:
        return item

if __name__ == "__main__":
    if len(sys.argv) < 2:
        print("require for res_file_path")
        sys.exit(1)
    elif len(sys.argv) < 3:
        res_file = sys.argv[1]
        output_json = res_file.replace(".json", f"_eval_results_gpt41.json")
        log_path = res_file.replace(".json", f"_eval_results_gpt41.log")
    else:
        raise ValueError()

    with open(res_file, 'r', encoding='utf-8') as fp:
        res_data = json.load(fp)

    with open(events_file, 'r', encoding='utf-8') as fp:
        events_data = json.load(fp)

    map_dic = {}
    for item in events_data:
        map_dic[item["video"]] = item
        
    for item in res_data:
        video_path = item['id'][0].split("/")[-2] + '/' +  item['id'][0].split("/")[-1]
        if video_path in map_dic:
            events = map_dic[video_path]["events"]
            map_dic[video_path] = item
            map_dic[video_path]["events"] = events

    res_data = list(map_dic.values())
    print(len(res_data))

    total_result = []

    for i in range(7):
        with concurrent.futures.ThreadPoolExecutor(max_workers=100) as executor:
            responses = list(tqdm(executor.map(gpt_extract, res_data)))

        seed += 1
        result = [r for r in responses if "Hallucination" in r]
        ignore = [r for r in responses if "Hallucination" not in r]
        try:
            with open(output_json, 'w', encoding='utf-8') as fp:
                json.dump(result, fp, indent=4, ensure_ascii=False)
        except Exception as e:
            print(e)
        print(output_json)
        print(len(result), len(ignore))

        k = 0

        while len(ignore) != 0 and k < 8:
            with concurrent.futures.ThreadPoolExecutor(max_workers=min(100, len(ignore))) as executor:
                responses = list(tqdm(executor.map(gpt_extract, ignore)))

            seed += 1
            result += [r for r in responses if "Hallucination" in r]
            ignore = [r for r in responses if "Hallucination" not in r]
            try:
                with open(output_json, 'w', encoding='utf-8') as fp:
                    json.dump(result, fp, indent=4, ensure_ascii=False)
            except Exception as e:
                print(e)
            print(output_json)
            print(len(result), len(ignore))

            k += 1

        if len(ignore) > 0:
            k = 0
            print("Version 2 GPT request")
            while len(ignore) != 0 and k < 3:
                with concurrent.futures.ThreadPoolExecutor(max_workers=10) as executor:
                    responses = list(tqdm(executor.map(gpt_extract_2, ignore)))
                
                seed += 1
                result += [r for r in responses if "Hallucination" in r]
                ignore = [r for r in responses if "Hallucination" not in r]
                try:
                    with open(output_json, 'w', encoding='utf-8') as fp:
                        json.dump(result, fp, indent=4, ensure_ascii=False)
                except Exception as e:
                    print(e)
                print(output_json)
                print(len(result), len(ignore))
                k += 1

        eer = sum([it['Missed'] + it['Incorrect'] + it['Hallucination'] for it in result]) / sum([len(it["events"]) for it in result])
        left_r = sum([it['Missed'] for it in result]) / sum([len(it["events"]) for it in result])
        inc_r = sum([it['Incorrect'] for it in result]) / sum([len(it["events"]) for it in result])
        fan_r = sum([it['Hallucination'] for it in result]) / sum([len(it["events"]) for it in result])

        print(f"EER: {eer}, MISS: {left_r}, INCORRECT: {inc_r}, HALLUCINATION: {fan_r}, Success_samples: {len(result)}")
        print(f"{eer:.3f}, {left_r:.3f}, {inc_r:.3f}, {fan_r:.3f}, {len(result)}")

        total_result.append([eer, left_r, inc_r, fan_r, len(result)])


    sorted_result = sorted(total_result, key=lambda x: x[0])
    print(f"EER: {sorted_result[3][0]:.3f}, MISS: {sorted_result[3][1]:.3f}, INCORRECT: {sorted_result[3][2]:.3f}, HALLUCINATION: {sorted_result[3][3]:.3f}")
    with open(log_path, 'w', encoding='utf-8') as f_log:
        f_log.write("=" * 80 + "\n")
        f_log.write("                    PERFORMANCE SUMMARY\n")
        f_log.write("=" * 80 + "\n")
        summary_line = (
            f"  EER         : {highlight_result[0]:.4f}\n"
            f"  MISS        : {highlight_result[1]:.4f}\n"
            f"  INCORRECT   : {highlight_result[2]:.4f}\n"
            f"  HALLUCINATION: {highlight_result[3]:.4f}\n"
        )
        f_log.write(summary_line)
        f_log.write("=" * 80 + "\n\n")

        f_log.write("Full Results Table:\n")
        
        header = f"{'EER':<18}{'MISS':<18}{'INCORRECT':<18}{'HALLUCINATION':<18}{'Success Samples':<18}"
        separator = "-" * len(header)
        
        f_log.write(header + "\n")
        f_log.write(separator + "\n")
        
        # Write the data rows
        for sublist in sorted_result:
            row_str = (
                f"{sublist[0]:<18.4f}"
                f"{sublist[1]:<18.4f}"
                f"{sublist[2]:<18.4f}"
                f"{sublist[3]:<18.4f}"
                f"{sublist[4]:<18}"
            )
            f_log.write(row_str + '\n')
        
        f_log.flush()