from pathlib import Path import glob import json import csv import tqdm import hashlib from convert_utils.base_convertor import BaseConvertor from time_utils.time_utils import format_time, time_to_seconds, seconds_to_time prompt = '''You are an advanced video question-answering AI assistant. You have been provided with some frames from the video and a multiple-choice question related to the video. Your task is to carefully analyze the video and provide the best answer to question, choosing from the four options provided. Respond with only the letter (A, B, C, or D) of the correct option. Question: {} Is it the right time to output {}? You can only answer yes or no. The answer is: ''' class PrepareBenchmarkTemplate(BaseConvertor): VIDEO_ROOT = "/home/dyvm6xra/dyvm6xrauser36/dyvm6xrauser36/huankguan2/Project/OmniEval/benchmarks/streamingbench/StreamingBench/videos" def read_data(self, data_file: str) -> list: with open(data_file, mode="r", encoding="utf-8") as f: reader = csv.DictReader(f) data = list(reader) return data def get_video_uuid(self, item) -> str: video_uuid = hashlib.sha256( self.get_video_path(item).encode("utf-8") ).hexdigest()[:8] return video_uuid def get_video_path(self, item) -> str: task_name, sample_name, sample_id, _ = item["question_id"].split("_") video_path = ( f"{self.VIDEO_ROOT}/{task_name}/{sample_name}_{sample_id}/video.mp4" ) return video_path def get_stream_addr(self, item) -> str: return f"/home/dyvm6xra/dyvm6xrauser36/dyvm6xrauser36/huankguan2/Project/OmniEval/benchmarks/streamingbench/stream_chunks/{self.get_video_uuid(item)}" def build_sqa(self, item) -> list: """event_id: int, timestamp: str""" question_t = format_time(item["time_stamp"]) response_t = format_time(item.get("ground_truth_time_stamp", question_t)) # question_id = item["question_id"] event_dict = { "question_id": item["question_id"], "task_type": item["task_type"], "question": prompt.format(item["question"], item["ground_truth_output"]), "response": None } qt = time_to_seconds(question_t) rt = time_to_seconds(response_t) round_t = [ max(qt, rt+i) for i in range(-4, 5)] event_ts = [question_t] + [seconds_to_time(t) for t in round_t] sqa = [] for i, ts in enumerate(self.deduplicate_and_sorting(event_ts)): event = { "event_id": i + 1, "timestamp": ts, } if ts >= response_t: event_dict["response"] = "yes" else: event_dict["response"] = "no" sqa.append(event | event_dict) return sqa inputs = [ ( "proactive_output_50.jsonl", ["StreamingBench/annotations/Proactive_Output_50.csv"], ), ( "proactive_output.jsonl", ["StreamingBench/annotations/Proactive_Output.csv"], ), ] for outfile, samples in inputs: converter = PrepareBenchmarkTemplate(outfile=outfile, start_id=1) converter.clear() for x in tqdm.tqdm(samples): converter.convert(x) print("done:", outfile)