import re, time, os from tqdm import tqdm import json from datetime import datetime import argparse import Levenshtein from base_agent import BaseAgent from system_prompts import sys_prompts from tools import ToolCalling from process import * import pandas as pd def parse_args(): parser = argparse.ArgumentParser(description='Eval-Agent-VBench', formatter_class=argparse.RawTextHelpFormatter) parser.add_argument( "--user_query", type=str, required=True, help="user query", ) parser.add_argument( "--model", type=str, default="latte1", help="target model", ) args = parser.parse_args() return args def most_similar_string(prompt, string_list): similarities = [Levenshtein.distance(prompt, item["Prompt"]) for item in string_list] most_similar_idx = similarities.index(min(similarities)) return string_list[most_similar_idx] def check_and_fix_prompt(chosed_prompts, prompt_list): results_dict={} for key, item in chosed_prompts.items(): thought = item["Thought"] sim_item = most_similar_string(item["Prompt"], prompt_list) sim_item["Thought"] = thought results_dict[key] = sim_item return results_dict def format_dimension_as_string(df, dimension_name): row = df.loc[df['Dimension'] == dimension_name] if row.empty: return f"No data found for dimension: {dimension_name}" formatted_string = ( f"{row['Dimension'].values[0]}: " f"Very High -> {row['Very High'].values[0]}, " f"High -> {row['High'].values[0]}, " f"Moderate -> {row['Moderate'].values[0]}, " f"Low -> {row['Low'].values[0]}, " f"Very Low -> {row['Very Low'].values[0]}" ) return formatted_string class EvalAgent: def __init__(self, sample_model="latte1", save_mode="video", refer_file="vbench_dimension_scores.tsv"): # self.tools = ToolCalling(sample_model=sample_model, save_mode=save_mode) self.sample_model = sample_model self.user_query = "" self.tsv_file_path = refer_file def init_agent(self): self.prompt_agent = BaseAgent(system_prompt=sys_prompts["vbench-prompt-sys"], use_history=False, temp=0.7) self.plan_agent = BaseAgent(system_prompt=sys_prompts["vbench-plan-sys"], use_history=True, temp=0.7) def search_auxiliary(self, designed_prompts, prompt): for _, value in designed_prompts.items(): if value['Prompt'] == prompt: return value["auxiliary_info"] raise "Didn't find auxiliary info, please check your json." def sample_and_eval(self, designed_prompts, save_path, tool_name): prompts = [item["Prompt"] for _, item in designed_prompts.items()] video_pairs = self.tools.sample(prompts, save_path) if 'auxiliary_info' in designed_prompts["Step 1"]: for item in video_pairs: item["auxiliary_info"] = self.search_auxiliary(designed_prompts, item["prompt"]) eval_results = self.tools.eval(tool_name, video_pairs) return eval_results def reference_prompt(self, search_dim): file_path = "./eval_tools/vbench/VBench_full_info.json" data = json.load(open(file_path, "r")) results = [] for item in data: if search_dim in item["dimension"]: item.pop("dimension") item["Prompt"] = item.pop("prompt_en") if 'auxiliary_info' in item and search_dim in item['auxiliary_info']: item["auxiliary_info"] = list(item["auxiliary_info"][search_dim].values())[0] results.append(item) return results def format_eval_result(self, results, reference_table): question = results["Sub-aspect"] tool_name = results["Tool"] average_score = results["eval_results"]["score"][0] video_results = results["eval_results"]["score"][1] output = f"Sub-aspect: {question}\n" output += f"The score categorization table for the numerical results evaluated by the '{tool_name}' is as follows:\n{reference_table}\n\n" output += f"Observation: The evaluation results using '{tool_name}' are summarized below.\n" output += f"Average Score: {average_score:.4f}\n" output += "Detailed Results:\n" for i, video in enumerate(video_results, 1): prompt = video["prompt"] score = video["video_results"] output += f"\t{i}. Prompt: {prompt}\n" output += f"\tScore: {score:.4f}\n" return output def update_info(self): folder_name = datetime.now().strftime('%Y-%m-%d-%H:%M:%S') + "-" + self.user_query.replace(" ", "_") self.save_path = f"./eval_vbench_results/{self.sample_model}/{folder_name}" os.makedirs(self.save_path, exist_ok=True) self.video_folder = os.path.join(self.save_path, "videos") self.file_name = os.path.join(self.save_path, f"eval_results.json") def explore(self, query, all_chat=[]): self.user_query = query self.update_info() self.init_agent() df = pd.read_csv(self.tsv_file_path, sep='\t') plan_query = query all_chat.append(plan_query) n = 0 while True: breakpoint() plans = self.plan_agent(plan_query, parse=True) if plans.get("Analysis"): all_chat.append(plans) print("Finish!") break tool_name = plans["Tool"].lower().strip().replace(" ", "_") reference_table = format_dimension_as_string(df, plans["Tool"]) prompt_query = json.dumps(plans) prompt_list = self.reference_prompt(tool_name) prompt_query = f"Context:\n{prompt_query}\n\nPrompt list:\n{json.dumps(prompt_list)}" designed_prompts = self.prompt_agent(prompt_query, parse=True) designed_prompts = check_and_fix_prompt(designed_prompts, prompt_list) plans["eval_results"] = self.sample_and_eval(designed_prompts, self.video_folder, tool_name) plan_query = self.format_eval_result(plans, reference_table=reference_table) all_chat.append(plans) if n > 9: break n += 1 all_chat.append(self.plan_agent.messages) save_json(all_chat, self.file_name) def main(): args = parse_args() user_query = args.user_query eval_agent = EvalAgent(sample_model=args.model, save_mode="video") eval_agent.explore(user_query) if __name__ == "__main__": main()